Yixin Zhu | PKU

CivRealm: 基于文明演化的多智能体学习与推理决策仿真

Thu, 20 Aug 2026 00:00:00 +0000

项目背景与目标

在人类社会的发展进程中，决策常常需要“基于历史经验的学习”与“复杂环境下的推理”协同进行。当前主流人工智能决策体在绝大多数交互环境中，更重视单纯的学习，而忽视了推理与协商等真实决策关键环节。

本项目旨在基于文明系列策略游戏规则和真实社会演变规律，开发多智能体交互仿真平台 CivRealm，推动智能体在“开放型、不完全信息、复杂社会博弈环境”下同时实现高效学习与创新性推理。系统支持多类型AI智能体（如RL-agent与LLM-agent），能够模拟发展、战争、外交等多样历史与社会情境，挑战智能体在资源管理、社会协商、长期规划中的泛化能力。项目的最终目标是构建一个既能支持强化学习又能支持大模型推理，并涵盖多样子任务评测的文明决策环境，促进智能体泛化研究和社会决策应用进步。

核心任务设计

任务一：高复杂度多智能体文明环境构建（20分）

CivRealm是一个类似于文明6的仿真游戏平台，支持多智能体，战斗，交易等行为，采用模块化仿真引擎，支持自定义地图、任务与事件脚本。基于此,本试验从探索一些关键的场景设计开始。注意，以下的场景设计原有仿真器都支持，所以你不需要从头开发，但你需要设计简单的脚本验证这些设计：

基于“文明”类游戏规则，探索仿真城市发展、土地扩张、单位管理、科技进步等元素，探索动态玩家数量和不完全信息场景。
探索社会历史事件、外交互动、环境变化等机制，探索智能体需应对资源的不确定性、动态联盟、突发事件（如战争和自然灾害）。
探索多智能体通信与交易系统，支持自主协商、联盟、竞争等社会行为。需要交付一个简单的脚本或是仿真的录屏，验证仿真器的这些行为即可。

任务二：多智能体学习与推理（40分）

CivRealm提供张量型（Tensor-based）API，支持RL智能体获取观察、做出动作、获得奖励，支持A2C/DQN等主流强化学习算法，并提供自然语言（Language-based）API，为LLM智能体提供环境播报、战略决策、外交协商等语境输入输出，支持prompt engineering与多轮会话。你需要在下列任务中二选一：

基线 RL 智能体训练与对比 实现并训练一个强化学习基线智能体（如 BaseRL），并可选实现分层/协同智能体（如 Mastaba）。

使用张量（Tensor）API获取环境观察，输出动作并获得奖励。
在文明仿真任务（如城市扩张、资源管理或战斗任务）下，训练并评估算法性能。
对比不同决策范式（如传统 RL vs. 分层协同）的表现，分析各自优劣。

基于大语言模型的 Agent Pipeline 搭建一个基于大语言模型（如 GPT-4）驱动的智能体管线，通过自然语言 API 与环境互动。

实现一个基于大语言模型的agent pipeline。Agent需要观察、推理和给出命令。你可以参考论文中的实现，使用AutoGPT框架复现论文中实现的智能体。对于论文中提到的任务，你可以任选其一。
Agent 能够接收环境播报，进行推理，输出战略命令，并参与外交协商/社会事件。
展示 Agent 在关键任务上的表现和推理过程，并与 RL 智能体进行简要对比分析（如果可以）。

你可以参考论文中的baseline实现。复现部分实验结果并分析即可。

任务三：开放问题（40分）

用任何方法，设计强化学习或基于大语言模型的baseline，在某些方面超过原有的baseline结果（推荐以每场游戏结束时的平均分数为基准）。你可以考虑model-based rl或者llm的sft或者dpo等优化方法（但是单纯替换基座模型带来的效果提升不算）。你可以尝试的方向有：

模型规划的战略性和长期性
对于其他玩家（或智能体）的偏好/规划预测
基于self-play的左右互搏
…（任何你感兴趣的）

如果你感兴趣并且性能提升明显，可以考虑投稿相关论文。

数据集与技术栈

核心数据集：CivRealm

推荐技术栈：

仿真引擎：Python3.9+，NumPy/Pandas为基础，支持自定义状态空间与事件脚本
强化学习：PyTorch/TensorFlow，StableBaselines
大模型推理：OpenAI API、LangChain、transformers（LLM agent调用与prompt设计）
前端与系统：Streamlit/FastAPI/WebGL（可视化界面与API集成）

评估标准与预期成果

评分分配（100分）：

任务一：20分——成功运行场景，验证场景功能设计
任务二：40分——智能体实现、实验复现、对比分析的清晰度与深度
任务三：40分——优化方法的新颖性、效果提升与分析说明

预期交付物：完整代码（复现脚本+readme file）、详细实验报告（含背景、方法、结果与分析），录屏/截图示例，创新想法及不足反思

参考文献

Qi, S., Chen, S., Li, Y., Kong, X., Wang, J., Yang, B., … & Zhu, S. C. CivRealm: A Learning and Reasoning Odyssey in Civilization for Decision-Making Agents. In The Twelfth International Conference on Learning Representations.

CognitiveMap: 从稀疏、有噪的成对比较中恢复关系结构

Thu, 20 Aug 2026 00:00:00 +0000

小组作业：从稀疏、有噪的成对比较中恢复关系结构

形式：4–5 人小组｜周期：4 周｜ 建议投入：每人每周 6–8 小时

提交物（一份 zip + 一个仓库链接）

技术报告（英文 LaTeX，正文 ≥ 8 页，不含参考文献）
代码仓库（含 README 与一键复现脚本，他人可完整重跑全部图表）
结果数据（严格按第 8 节的格式）
15 分钟英文口头汇报（录像）

分工要求：报告首页给出 CRediT 格式贡献声明，每位成员至少主责一个模块。全组同分，贡献声明与答辩表现不一致者单独调分。

**⚠️ 报告须含一节「模块间一致性检查」（占报告分的 1/3）。**本作业的六个模块不是六份独立作业，它们相互咬合：

检查	涉及模块
A.4 的解析式能否预测 D.1 在 ε = 0 时实测的可解对比例	A → D
训练出的网络放到相图上，落在 flat 与闭包的哪一侧	B → D
C 给出的预算分配建议，在 D 的相图上是否真的更优	C → D
E 的真实概念空间，落在 D 相图的哪个位置	E → D

四条至少答三条。若某条检查的结果互相矛盾，如实写出矛盾并分析原因——发现并说清一处矛盾，比四条都"一致"更能拿分。这一节也是答辩必问的部分：如果四个人各做各的、从不对话，这一节写不出来。

0. 这份作业问的是一个问题

只学过相邻关系 A>B、B>C、C>D，人和动物却能判断从未学过的 A>D[1]。一个直觉的解释是联想强度（flat associative）：反复配对让 A 带上高分、D 带上低分，比较分数即可。另一类解释认为大脑构建了一张关系图，推理是在图上找路径[2,3]；这一视角把空间编码视为人类思维的通用底层[10]，并已有形式化的实现——把关系表示成「从此处出发会到达何处」的预测性结构[9]。海马–内嗅系统不仅编码物理空间[4]，也以同样方式组织抽象关系与社会关系[5,6,7,8]。

这两类模型在观测充分时都能答对，因此区分它们的地方在观测不充分处。本作业只问一个问题：

观测要多稀疏、多脏，一张关系图才推不动了？一维序和二维支配的门槛，是同一条吗？

这个问题看起来是工程问题，其实是理论问题：如果两种推理的可行域边界不同，那它们就不可能由同一个瓶颈决定；如果边界重合，则"二维更难"只是"二维平均距离更短"造成的假象。目前没有人系统地量过这条边界。

四个人的建议分工（4 人组可合并 C、F）：

模块	内容	分值
A	基线模型与三条行为签名	15
B	神经网络的可学习性边界	20
C	固定观测预算下的分配问题	15
D	相图：一维与二维的可行域	25
E	真实概念空间上的复核	15
F	产出可证伪的人类预测	10

**模块 D 是本作业的主线，其余模块都是为它服务的工具或对照。**如果时间不够，优先保证 A + D 的完整性。

1. 通用协议（所有模块共用，不得更改）

真值结构：N 个个体，随机打乱得到一个真实全序；二维情形为两个正交的序（即每个个体有一对秩 (x, y)）。真值只用于评分，不得喂给模型。
观测（前提）：从全部 C(N,2) 个配对中以概率 p 独立采样。被采样到的配对给出一条有向边"i > j"。
噪声：每条被观测的边以概率 ε 方向反转。ε 是边错误率，不是试次错误率。
提问：在未被观测的配对（held-out）上提问。二维情形只在可比配对（一项在两维上都不低于另一项且至少一维更高）上统计支配判断准确率。
重复与种子——这是两件不同的事，两件都要做：
- 重复（repetition）：在一个种子内，每个数据点 ≥ 200 次重新抽真值 + 重新抽观测；
- 种子（seed）：整条流水线用 ≥ 3 个独立随机种子各跑一遍（建议 42 / 43 / 44）。最终报告的形式是 mean ± s.d. over seeds，其中每个 seed 的值本身已是 200 次重复的均值。
禁止调用现成的传递闭包库，闭包需自己实现（三重循环或 Floyd–Warshall 皆可）。

⚠️ 二维的组合方式必须写清楚：你是先对每一维各取闭包再组合，还是先组合再取闭包？两者不等价，结论也不同。报告中须明确说明并给出理由。

1.1 种子、报告与复现（全局硬性要求，适用于每一个模块）

这三条是本作业的红线，任何一条不满足，相关模块直接计 0 分。

(a) 每一个数字都必须来自 ≥ 3 个独立种子。

没有例外。不要因为"这个结果是确定性的"就只跑一次——确定性的算法不等于确定性的结果：真值次序是随机抽的，观测掩码是随机抽的，噪声是随机加的，网络初始化是随机的。只要流水线里还有一处随机，就必须多种子。
唯一豁免的是解析结果（模块 A.4 的闭式表达式）。豁免的部分请在报告中明确标注为 analytic。
种子必须显式设置并记录：random、numpy、torch（含 torch.manual_seed）、以及环境变量 PYTHONHASHSEED。种子作为命令行参数暴露出来，使任何一次运行都能被单独重放。

(b) 报告所有结果，不是最好的结果。

每张图、每个表格单元都写成 mean ± s.d.（n = 种子数），并在附录中逐种子列出原始数值。
禁止只报告"表现最好的那个种子/维度/超参"。若某个种子的行为与其他种子明显不同，把它写出来并讨论——这通常是最有价值的观察，不是需要藏起来的瑕疵。
若某项实验跑失败了、或没跑完，在报告中明确写出来，并说明缺口在哪。留白不扣分，掩盖要扣分。

(c) 每一个数字都必须可追溯到代码。

仓库中须有 README，说明从零开始如何复现每一张图、每一个表——具体到"运行哪个脚本、带什么参数、产出哪个文件"。
每个结果文件（.json / .csv）中须记录产生它的种子、参数、脚本名与 git commit。
报告中每张图的题注须注明它由哪个脚本、哪个数据文件生成。
**助教会随机抽取至少两个结果重跑。**重跑不出来（或与提交值不一致，超出浮点末位差异）的，该模块计 0 分。

这三条看起来繁琐，但它们正是把"我跑出来了"变成"这个结果成立"的全部区别。本作业的很多结论是否定性的（某个方法不行、某条边界不存在），而否定性结论对上述要求最敏感——一个没记种子、没法重跑的否定结果，等于没有结果。

1.2 数据从哪来、需要装什么

模块 A、B、C、D、F 不使用任何外部数据集。全部数据由你按 §1 的协议自行生成——每一次"实验"就是：抽一个真值次序 → 按 p 抽观测 → 按 ε 加噪 → 在 held-out 上提问。不要去找现成的传递推理数据集，本作业不需要，也没有公认可用的。

这样设计不是图省事，有三条实在的理由：

**只有真值已知，准确率才有定义。**在真实数据上你根本不知道"正确答案"是什么，就无从判断闭包补出的关系对不对。
**只有能自由拧 N、p、ε 三个旋钮，模块 D 的相图才画得出来。**相图的横纵轴就是这些旋钮，外部数据集把它们钉死了。
**可复现。**数据由种子完全决定，助教重跑你的脚本能得到逐位相同的结果——这正是 §1.1(c) 要求的。

唯一用到外部数据的是模块 E（预训练词向量），下载与规格见 E.1。

环境：Python 3.9+、NumPy、Matplotlib、PyTorch（仅模块 B 需要，CPU 版即可）。不需要 GPU——所有实验在笔记本上可完成，模块 D 的机时开销来自格点数量而非单次计算（见第 10 节的估算与并行化建议）。仓库的 README 中请写明你实际使用的版本号。

2. 模块 A：基线模型与三条行为签名（15 分）

A.1（3 分）两个 agent

def flat_agent(observed_pairs, query):
 """查表：只对观测过的配对作答，其余随机猜。"""

def map_agent(observed_pairs, query):
 """关系图：先对观测边取传递闭包，再回答查询；仍不可推者随机猜。"""

A.2（4 分）符号距离效应

N = 7，p ∈ {0.2, 0.35, 0.5}，按秩距离 1…6 分组统计 held-out 准确率，两个 agent 各一条线，含误差棒。

报告要求：关系图 agent 是否出现准确率随距离上升？查表 agent 呢？给出机制解释，并说明这一效应为什么不能由联想强度产生。

A.3（4 分）端点锚定效应（end-anchor）

统计含端点（最高或最低项）的配对 vs 纯中间配对的准确率，画"准确率 vs 序位"曲线，检验是否呈两端高、中间低的弓形。

报告要求：解释弓形如何从闭包中自然出现（提示：端点参与了多少条可推链？）。并回答——若把结构从全序换成环（首尾相接），弓形还在吗？先给预测，再跑。

A.4（4 分）解析部分

对一维、无噪（ε = 0）情形，推导 held-out 配对（即该对本身未被观测）被闭包解出的概率 P_resolve(d; N, p)，其中 d 为两者的秩距离。

这一题有精确解，不要一上来就写模拟。

按以下顺序做，每一步都比上一步难一点：

先想清楚谁能当中间站。闭包里的路径必须沿真值次序单调，所以能把 i 和 j 接起来的点，秩必须严格介于两者之间。请先回答：这样的点有几个？
**d = 1 的精确值。**由上一步直接得出，写出来并说明理由。
**一般 d 的精确表达式。**顺着"可达集逐个吸收中间点"的思路，写出一个可计算的递推（提示：新中间点能否被接上，只取决于当前已可达的点数）。这不是近似，是精确解。
**一个可以被证伪的推论。**你的表达式里出现 N 了吗？据此预测 P_resolve(d; N, p) 与 N 的关系，然后用 N = 7、12、20 的模拟去检验这条预测。先写预测，再跑。
把理论曲线与模拟画在同一张图上，报告最大偏差，并说明该偏差是否在蒙特卡洛误差范围内。

若第 3 步实在做不出，可以给上下界并说明界在何处松，但必须完成第 4 步的预测–检验。这一小题是模块 D 的理论锚点，D.3 会直接用到它的结论。

3. 模块 B：神经网络的可学习性边界（20 分）

前两个 agent 是符号模型。一个自然的问题是：一个从数据里学的连续模型，会不会自己长出同样的能力？

用 nn.Embedding（每个个体一个可学习向量）+ 两层 MLP（输入为两个个体向量的拼接），在观测到的边上以交叉熵训练，在 held-out 配对上测试。

B.1（5 分）一维

N = 7，相邻采样（只给 N−1 条相邻边）。画准确率 vs 秩距离。

网络是否自发出现符号距离效应？与 A.2 的闭包曲线对比。
把学到的嵌入向量降到 1 维可视化：它排成一条线吗？该一维投影与真实秩的相关有多高？
≥ 20 个随机结构 × ≥ 3 个种子，报告 mean ± s.d.（跨种子），并在附录逐种子列出。

B.2（8 分）二维与"不可比"

4×4 网格，三分类：a 支配 b / b 支配 a / 不可比。

关键设定：训练集只含单轴相邻对——因而训练集里一条"不可比"样本都没有。测试集包含全部 held-out 配对，其中相当一部分是不可比的。

嵌入维度取 1、2、4、8、32，每个维度 ≥ 6 个随机种子（此处种子数高于全局下限 3，因为网络初始化的方差更大）；
分别报告可比配对与不可比配对的判别准确率（不要只报总体准确率，它会掩盖问题）；
画出准确率 vs 嵌入维度。

B.3（7 分）诊断

若网络在"不可比"上失败，请判断失败来源属于以下哪一类，并给出能相互分离的对照实验：

	解释	你的对照应当预测什么
(a)	训练集不含该标签
(b)	损失函数
(c)	网络容量/结构
(d)	“给所有个体打一个分数再排序"这件事本身

至少必须做的对照：把"不可比"样本按比例 r 加入训练集，r 从 0 扫到 0.5，看准确率怎么动。

报告要求：只写"可能是因为……“不得分。必须给出每种解释各自预测什么、你的数据支持哪一个、又排除了哪一个。若你认为 (d) 成立，请把它写成一个关于表示的命题并给出论证——注意任意两个实数必有大小。

B.4（与模块 D 的接口，不单独计分，但缺失则 D 扣 5 分）

把训练好的网络当作第三个 agent，放到模块 D 相图的至少 6 个格点上评测（建议沿 p 与 ε 各取一条线），与 flat / 关系图两个 agent 画在同一张图里。

必答：网络的可行域边界落在另外两个 agent 的哪一侧？ 它更像查表，更像闭包，还是两者都不像？在噪声大的格点上，它是退化成"猜”，还是自信地答错？——这两种失败方式的后果完全不同。

4. 模块 C：固定观测预算下的分配问题（15 分）

现实中每一次比较都有成本（一场比赛、一次试验、一次标注）。给定总预算 B 次比较，你可以：

覆盖 B 个不同配对，每对只看 1 次（密但脏）；
或覆盖 B/k 个配对，每对看 k 次取多数（稀但干净）。

C.1（8 分）最优 k

固定 B，扫 k ∈ {1, 2, 3, 5, 7, 9}，在 ε ∈ {0.05, 0.1, 0.2, 0.3} 下分别找出使 held-out 准确率最高的 k*。

注意约束：覆盖的配对数 ⌊B/k⌋ 不能超过 C(N,2)。请选一个使 k = 1 时不会覆盖满全部配对的 B（否则 k = 1 那一档没有意义），并在报告中写明你取的 B 与对应的覆盖率。

画 “准确率 vs k”，每个 ε 一条线；
给出 k*(ε) 的关系，并解释它的形状。

C.2（4 分）它依赖什么

k* 随 N 变化吗？随 B/C(N,2) 的比值变化吗？至少扫两个 N。

C.3（3 分）一维 vs 二维

同一预算下，一维与二维的 k* 相同吗？ 若不同，方向是什么，为什么？

报告要求：给出一条可直接用于实验设计的建议——“在边错误率约为 ε 时，预算应当这样花”。这条建议必须在模块 D 的相图上被检验：按你的建议分配预算，落点应当在可行域内；按等预算的 k = 1 分配，落点在哪？把两个落点标在 D.1 的图上。建议与检验不一致时，如实报告并分析原因——这比给一条没验过的建议有价值得多。

5. 模块 D：相图 —— 一维与二维的可行域（25 分）

本模块是主线。

D.1（8 分）扫描

对一维和二维分别扫描：

个体数 N：一维取 {9, 12, 16, 25, 36, 49}；二维取对应的矩形网格 3×3、3×4、4×4、5×5、6×6、7×7，使两种结构在同一 N 下逐点可比（这是 D.2 能做比较的前提，不要各扫各的）
观测密度 p ∈ {0.2, 0.3, 0.4, 0.6, 0.8}
边错误率 ε ∈ {0, 0.1, 0.2, 0.3}

每个格点报告两个 agent 在 held-out 上的准确率：≥ 3 个种子 × 每种子 ≥ 200 次重复，报告跨种子的 mean ± s.d.

输出：每个 ε 一张 (N, p) 热力图，画 map 减 flat 的优势；在图上标出优势 = 5 个百分点的等值线（这条线就是"可行域边界”）。

必答：在固定 (p, ε) 下，优势随 N 单调吗？若不单调，峰值在哪、之后为什么会掉？把"更多个体"与"更多错边"这两个相反的作用分别量化，说明是哪一个在何处占上风。

D.2（6 分）两条边界是同一条吗

把一维与二维的等值线画在同一张图上。

在同一 (p, ε) 下，二维需要多少个体才能达到一维在 N = 12 时的优势水平？给出比值，并给出它随 ε 的变化。
若二维的边界系统性地更靠右（需要更多个体），请给出这个差距的来源。至少考虑三种候选：(i) 可比配对本身更少；(ii) 组合步骤额外损失了信息；(iii) 组合步骤的 AND 结构改变了出错的方式（把"答错"变成"答不出"）。设计能把这三者分开的对照，不要只做两分类。
若两条边界重合，那么"二维更难"的说法就需要重新表述——请给出你认为正确的表述。

D.3（6 分）有没有一个控制变量

定义可解对比例 R = 闭包能补出的 held-out 配对占全部 held-out 配对的比例（这是一个可以直接测量的量，不依赖准确率）。

关键检验：把 D.1 中所有格点的准确率对 R 作图。

不同的 (N, p, ε) 组合是否塌缩到同一条曲线上？
与 A.4 对照：把 A.4 的解析式按你的秩距离分布加权，算出 ε = 0 时 R 的理论值，与 D.1 实测的 R 比较。两者一致吗？若一致，那么相图在 ε = 0 这一层是可以预测的，不必扫；请说明这对 ε > 0 的那几层意味着什么。
若塌缩：你就找到了控制这件事的单一变量，请给出这条曲线的形式，并说明它对一维、二维是否是同一条。
若不塌缩：残差里还有什么？ 请找出至少一个额外的控制变量，并证明加上它之后塌缩变好（给出定量的拟合优度对比，不要只看图）。

这一小题是整份作业里最难、也最有价值的一题。允许得出否定结论——“不存在这样的单一变量”，只要证据扎实，同样满分。

D.4（5 分）噪声的结构，还是噪声的大小？

前面把噪声当作 iid：每条边以相同概率 ε 反转。但真实的成对观测通常不是这样——两个实力接近的个体比一次，结果接近抛硬币；实力悬殊的两个，几乎不会爆冷。也就是说，错误率依赖于两者的秩距离。

请实现第二种噪声模型：

P(反转 | 秩距离 d) = f(d)，且 f 单调递减，例如 f(d) = 0.5 * exp(-d / λ)

这里有一个陷阱，请务必想清楚再动手。“把两种噪声模型调成可比"有两种做法，它们不等价：

对齐噪声函数：一维和二维用同一个 f(d)（即同一套"实力差 → 胜负概率"的规律）；
对齐实测错误率：分别调 λ，使两种结构实际产生的边错误率相同。

两种对齐必须都做，并分别报告。 然后回答：

在对齐噪声函数时，一维与二维实测的边错误率相同吗？若不同，差多少、为什么？（提示：把 N 个个体摆成二维网格后，每个轴上还剩几个等级？这些等级之间的秩距离典型是多大？）
在对齐实测错误率时，两种噪声模型的可行域边界还有差别吗？
综合两者：“二维更难"这件事，多大程度上来自结构本身，多大程度上来自二维改变了观测的噪声水平？ 这是本小题的核心结论，请用数字回答。

哪一种对齐才是"公平"的，取决于你要回答什么问题——请在报告中明确说明你的立场并给出理由。选错对齐方式会让你看不见真正的效应，这一点本身就是考点。

**这一小题可能会推翻 D.2 的结论。**如果发生这种情况，请在报告中明确写出来，并说明 D.2 的结论应当被限制在什么条件下——这比把两节结论强行调和更有价值。

6. 模块 E：真实概念空间上的复核（15 分）

前面全部用合成结构。合成结构的风险是：它可能比真实结构更好推（真值是随机全序，链条干净）。本模块检验结论在真实语义结构上是否还成立。

E.1（6 分）构建

数据来源（照此执行，以便跨组可比）：使用公开的 GloVe 6B / 300 维词向量，官方下载地址 https://nlp.stanford.edu/data/glove.6B.zip（约 820 MB，解压后取 glove.6B.300d.txt）。若下载不便，也可用 gensim.downloader.load("glove-wiki-gigaword-300") 获得同一份向量；用其他公开词向量亦可，但必须在报告中写明具体是哪一份。

规格：取 4 个语义类别（如动物 / 工具 / 食物 / 交通工具），每类 9 个词，共 N = 36 个节点。选 36 正是为了对上模块 D 的 N = 36 那一列，E.3 可以直接比，不必补跑。词表请写进代码或单独一个文件并提交，不要只在报告里罗列。

复现提示：不要把 820 MB 的原始文件提交到仓库。写一个脚本，从原始文件中抽出你那 36 个词的向量存成一个小文件（几十 KB）并提交，同时在 README 里写明原始文件的来源与抽取脚本。这样助教无需下载 820 MB 也能重跑。

节点为词，边权为余弦相似度。在这个图上定义一个有序结构（例如沿某个语义方向的投影给出秩），并按 §1 的通用协议采样观测、加噪、推断。

报告要求：说明你如何定义秩、为什么这个定义是合理的、以及它对结论的敏感性（换一个定义，结论变吗？）。

E.2（6 分）碎片化

真实语义空间是高度连通的：即使不加任何额外结构，度量本身就能跨类别到达。因此闭包的优势可能在真实空间上被稀释。

以概率 q 删除跨类别的边，把 q 从 0 扫到 1，量化"闭包相对查表的优势"随 q 的变化。

报告要求：优势是随 q 单调上升吗？如果是，说明了什么？如果不是，转折在哪、为什么？

E.3（3 分）与合成结构对照

把 E.2 的曲线与模块 D 中同等 (N, p, ε) 的合成结果画在一起。注意先对齐 N：E.1 取了多少个词，就用 D 相图上同样 N 的那一列作对照；若词数不在 D 的 N 网格上，请补跑该 N 的合成格点，不要拿相邻的 N 硬比。

真实结构比合成结构更好推还是更难推？ 给出定量比较，并解释差异来源。

7. 模块 F：产出可证伪的人类预测（10 分）

前六个模块全是模型。本模块要求你把模型变成可以被人类数据推翻的东西。本模块不收集任何人类数据。

产出一张预测表，每行一条预测，含以下列：

列	要求
可观测量	必须是行为实验能直接测到的（准确率、反应时、选项使用率…）
实验条件	具体到可照着编程序
关系图模型的预测	带数值和不确定度，来自你自己跑的模型
查表模型的预测	同上
（可选）网络的预测	若模块 B 的网络能对该条件作出预测，一并列出
判据	什么样的结果支持前者、什么样的结果支持后者、什么样的结果两者都推翻
所需被试量	给出功效分析或明确依据

至少 5 行，且其中至少 2 行必须是两个模型预测方向相反的（同向但幅度不同的预测区分力弱，只能算半行）。

加分（+3）：给出一条预测，是两个模型都无法解释的现象——即你认为人类会做出的、但你的任何模型都预测不出的行为。说明你的理由。

8. 结果汇交格式（不合格则模块 D 不计分）

所有小组的模块 D 结果将汇总比较，格式必须一致。提交 UTF-8 编码的 phase_diagram.csv：

列名	说明
`group_id`	组号
`structure`	`chain1d` / `grid2d`
`n_items`	个体数
`p`	观测密度
`eps`	总体边错误率
`noise_model`	`iid` / `rank_dependent`（见 D.4）
`combine_order`	`closure_then_combine` / `combine_then_closure`（见 §1 通用协议）
`n_repeats`	单个种子内的重复次数（须 ≥ 200）
`n_seeds`	独立种子数（须 ≥ 3）
`acc_map_mean`, `acc_map_sd`	关系图 agent，跨种子的均值与标准差
`acc_flat_mean`, `acc_flat_sd`	查表 agent，同上
`resolvable_frac_mean`, `resolvable_frac_sd`	可解对比例 R，同上
`acc_nn_mean`, `acc_nn_sd`	神经网络 agent（见 B.4）；未在该格点评测则留空
`seeds`	所有种子，逗号分隔，如 `42,43,44`
`git_commit`	产生本行的代码版本

代码仓库中须有一个脚本，从零重跑即可生成这张表，且与提交的表逐行一致（允许浮点末位差异）。助教会抽查复现。

9. 评分标准

项	分值	要点
A 基线与签名	15	A.4 必须有解析成分，不能只有模拟
B 可学习性边界	20	B.3 的对照能真正相互分离，不是罗列可能性
C 预算分配	15	给出可用于实验设计的具体建议
D 相图	25	D.3 的塌缩检验是否定量；D.4 的两种对齐方式都要做并分别报告；否定结论同样满分
E 真实概念空间	15	秩定义的敏感性分析不可省
F 人类预测	10	至少 2 行方向相反的预测
F 加分	+3	一条两个模型都解释不了的预测；总分上限 100，即加分只用于补足其他模块的失分
报告与英文汇报	乘数	写作与汇报质量作为 0.8–1.0 的乘数作用于总分；其中「模块间一致性检查」一节占该乘数评定的 1/3

扣分项：任一数字少于 3 个种子（该模块 0 分）；只报告最好的种子而不报全部（该模块 0 分）；图无坐标轴标签或图例（−3/图）；结论与自己的数据矛盾（−10）；未说明二维的组合次序（−5）；代码无法一键复现（该模块 0 分）；汇交表与重跑结果不一致（模块 D 0 分）。

关于否定结论：模块 B.3、D.2、D.3、D.4 与 E.2 允许并鼓励得出否定结论。“我做了，它不成立”，只要证据扎实（重复次数、误差棒、排除了平凡解释），与肯定结论同分。把不成立的结果修饰成成立的，按学术不端处理。

学术诚信：允许使用 AI 工具，但须在报告末尾说明用于哪一步、以及你如何核验其输出；未说明而被发现按学术不端处理。

10. 时间表（建议）

周	里程碑
第 1 周	分工；完成模块 A（含 A.4 解析）；搭好通用实验框架与并行跑图脚本；估算模块 D 的总机时
第 2 周	模块 B 与 C；同时启动模块 D 的扫描（这是最耗机时的一环，越早开始越好）
第 3 周	模块 D 收敛，做 D.2 / D.3 / D.4；B.4（需要 D.1 的格点，必须排在扫描之后）；模块 E
第 4 周	模块 F；补跑缺口；写报告；录汇报

先估机时再动手。模块 D 是 6（N）× 5（p）× 4（ε）× 2（结构）= 240 个格点，每格 ≥ 3 个种子 × ≥ 200 次重复，即 ≥ 144,000 次单次运行；D.4 的两种对齐各需再来一套。请在第 1 周就测出单次运行耗时、乘出总量、并做并行化（multiprocessing 或作业数组均可），不要等到第 3 周才发现跑不完。

提示：闭包用布尔矩阵运算实现，比三重 Python 循环快两到三个数量级——这一步做不做，直接决定这个模块能不能完成。

参考文献

[1] Bryant, P. E., & Trabasso, T. (1971). Transitive inferences and memory in young children. Nature, 232(5311), 456–458.

[2] Tolman, E. C. (1948). Cognitive maps in rats and men. Psychological Review, 55(4), 189–208.

[3] Behrens, T. E. J., Muller, T. H., Whittington, J. C. R., Mark, S., Baram, A. B., Stachenfeld, K. L., & Kurth-Nelson, Z. (2018). What is a cognitive map? Organizing knowledge for flexible behavior. Neuron, 100(2), 490–509.

[4] O’Keefe, J., & Dostrovsky, J. (1971). The hippocampus as a spatial map: preliminary evidence from unit activity in the freely-moving rat. Brain Research, 34(1), 171–175.

[5] Constantinescu, A. O., O’Reilly, J. X., & Behrens, T. E. J. (2016). Organizing conceptual knowledge in humans with a gridlike code. Science, 352(6292), 1464–1468.

[6] Garvert, M. M., Dolan, R. J., & Behrens, T. E. J. (2017). A map of abstract relational knowledge in the human hippocampal–entorhinal cortex. eLife, 6, e17086.

[7] Park, S. A., Miller, D. S., Nili, H., Ranganath, C., & Boorman, E. D. (2021). Inferences on a multidimensional social hierarchy use a grid-like code. Nature Neuroscience, 24(9), 1292–1301.

[8] Kumaran, D., Banino, A., Blundell, C., Hassabis, D., & Dayan, P. (2016). Computations underlying social hierarchy learning: distinct neural mechanisms for updating and representing self-relevant information. Neuron, 92(5), 1135–1147.

[9] Stachenfeld, K. L., Botvinick, M. M., & Gershman, S. J. (2017). The hippocampus as a predictive map. Nature Neuroscience, 20(11), 1643–1653.

[10] Bellmund, J. L. S., Gärdenfors, P., Moser, E. I., & Doeller, C. F. (2018). Navigating cognition: spatial codes for human thinking. Science, 362(6415), eaat6766.

CogPath: 人脑功能通路启发的多模态视频理解

Thu, 20 Aug 2026 00:00:00 +0000

项目背景与目标

人类视觉系统对不同类型的信息存在功能分化。经典研究将腹侧通路与物体识别和视觉知觉联系起来，将背侧通路与空间信息和视觉引导动作联系起来 [Goodale & Milner, 1992]；后续研究进一步提出与动态社会知觉相关的第三视觉通路 [Pitcher & Ungerleider, 2021]，以及偏向动作和物体识别的两条“what”通路 [Wurm & Caramazza, 2022]。场景加工也涉及多个具有功能选择性的皮层区域 [Epstein & Baker, 2019]。

自然视觉 fMRI 研究表明，自然电影能够诱发跨被试共享的皮层活动模式 [Hasson et al., 2004]；大量物体和动作类别则在皮层上形成连续的语义组织，并表现出一定程度的跨个体一致性 [Huth et al., 2012]。这些结果提示，视觉内容与皮层功能组织之间存在结构性的对应关系。

与此相比，许多视频多模态模型采用共享的视觉表征前端，再通过时序建模、跨模态融合或语言模块完成不同内容的理解。已有多专家和条件路由方法主要从任务分工、模态差异或计算效率出发，而较少直接利用人类视觉功能组织作为模型的结构先验。

基于以上背景，本研究重点关注“人脑视觉系统中的功能分化，能否成为视频模型的有效归纳偏置”这一问题，具体探索三个问题：

视频理解中的功能分化是否呈现出与人类视觉功能组织相似的结构；
基于功能分化的动态多专家架构是否优于共享编码器和纯学习式路由；
如果有效，收益究竟来自多专家结构本身，还是来自脑数据提供的结构化先验。

这一思路与 neuroconnectionist research programme 一致：将神经科学中的理论组织转化为可计算、可检验的模型假设，并通过模型比较检验其解释力 [Doerig et al., 2023]。

核心任务设计

任务一：内容类型与功能通路占据分析（25分）

目标：建立“视频内容属性 → 皮层功能选择性”的统计映射，为模型提供神经科学先验。
任务要点：
- 将公开皮层功能分区与自然视觉 fMRI 数据对齐；
- 使用现成视觉/视频模型自动提取内容属性，尽量避免人工标注；
- 分析物体、场景/空间、动作、动态社会信息等内容属性与不同功能区域响应之间的关系；
- 通过跨被试分析和交叉验证检验稳定性；
- 控制低级视觉特征、运动能量、镜头切换等潜在混杂因素。
评分维度：数据管线（10分），稳定性检验（10分），混杂控制（5分）。

任务二：功能专家视觉前端与动态路由（30分）

目标：构建受人类视觉功能组织启发的多专家视觉前端，并检验动态路由是否带来实际收益。
任务要点：
- 设置 object、spatial/scene、action、social-dynamic 等功能专家；
- 增加 language-semantic expert 作为跨模态工程对照，不将其宣称为与上述视觉通路同构的独立脑通路；
- 为各专家配置现成开源视觉编码器，通过统一接口接入主干；
- 构建随视频内容和任务变化的动态路由器；
- 加入时间持续性约束，避免路由在 token 级频繁切换；
- 冻结主干，仅训练路由器、投影层和轻量适配器。
评分维度：架构实现（12分），路由器与持续性约束（10分），可复现性（8分）。
加分项：输出专家占据时间轴，并与独立估计的人脑功能选择性进行对照可视化（最多额外10分）。

任务三：脑先验的作用与消融（25分）

目标：区分性能提升来自多专家架构本身，还是来自脑数据提供的结构化先验。
任务要点：
- 比较共享单编码器、多专家无脑先验、多专家+数据驱动路由、多专家+随机脑先验、多专家+真实脑先验等设置；
- 尽可能保持参数量、训练数据、训练步数和推理预算一致；
- 对脑先验强度进行分级消融；
- 分别分析脑先验对专家划分、路由稳定性和任务性能的影响；
- 若脑先验没有额外收益，完整报告负结果。
评分维度：对照设计（10分），等算力比较的严谨性（10分），结果分析（5分）。

任务四：评测与跨内容泛化（20分）

目标：验证功能结构化是否带来真实且可泛化的收益。
任务要点：
- 在标准视频理解评测集上进行评测；
- 按内容属性分层报告结果；
- 检验模型是否会随内容变化合理调整专家权重；
- 检验低社会性内容下 social expert 是否相应降权；
- 在固定推理预算下比较性能与计算效率；
- 在训练分布之外进行跨内容泛化测试。
评分维度：评测完整性（8分），分层分析（7分），泛化与效率（5分）。

数据集与技术栈

脑影像数据：公开皮层功能分区图谱、公开自然视觉/自然电影 fMRI 数据集。具体使用的数据集将单独引用其原始数据论文及数据说明。
视频数据：公开视频理解训练与评测集，优先使用已有内容属性或可由模型自动提取属性的数据集，减少人工标注。
技术栈：脑影像预处理与统计分析工具、开源视觉/视频编码器、开源视频多模态模型；以冻结主干为基础，外挂多专家和轻量动态路由器。

评估标准与预期成果

内容类型与功能系统占据分析（25分）
功能结构化视觉前端与动态路由（30分）
脑先验的作用与消融（25分）
评测与跨内容泛化（20分）

总分：100分（加分项计入总分，超过100分按100分计）

预期交付物

内容属性与皮层功能选择性的分析结果与可视化；
功能结构化视频模型原型及严格匹配的基线；
脑先验强度的对照实验与消融报告；
专家路由随视频时间变化的可视化；
模型路由与脑数据之间的表征对应分析；
可复现代码与技术文档。

参考文献

Bi, Y., Wang, X., & Caramazza, A. (2016). Object domain and modality in the ventral visual pathway. Trends in Cognitive Sciences, 20(4), 282–290. https://doi.org/10.1016/j.tics.2016.02.002

Goodale, M. A., & Milner, A. D. (1992). Separate visual pathways for perception and action. Trends in Neurosciences, 15(1), 20–25. https://doi.org/10.1016/0166-2236(92)90344-8

Hasson, U., Nir, Y., Levy, I., Fuhrmann, G., & Malach, R. (2004). Intersubject synchronization of cortical activity during natural vision. Science, 303(5664), 1634–1640. https://doi.org/10.1126/science.1089506

Huth, A. G., Nishimoto, S., Vu, A. T., & Gallant, J. L. (2012). A continuous semantic space describes the representation of thousands of object and action categories across the human brain. Neuron, 76(6), 1210–1224. https://doi.org/10.1016/j.neuron.2012.10.014

Epstein, R. A., & Baker, C. I. (2019). Scene perception in the human brain. Annual Review of Vision Science, 5, 373–397. https://doi.org/10.1146/annurev-vision-091718-014809

Pitcher, D., & Ungerleider, L. G. (2021). Evidence for a third visual pathway specialized for social perception. Trends in Cognitive Sciences, 25(2), 100–110. https://doi.org/10.1016/j.tics.2020.11.006

Wurm, M. F., & Caramazza, A. (2022). Two “what” pathways for action and object recognition. Trends in Cognitive Sciences, 26(2), 103–116. https://doi.org/10.1016/j.tics.2021.10.003

Doerig, A., Sommers, R. P., Seeliger, K., Richards, B., Ismael, J., Lindsay, G. W., Kording, K. P., Konkle, T., van Gerven, M. A. J., Kriegeskorte, N., & Kietzmann, T. C. (2023). The neuroconnectionist research programme. Nature Reviews Neuroscience, 24(7), 431–450. https://doi.org/10.1038/s41583-023-00705-w

MindGrid: 基于心理理论的多智能体协作推理仿真

Thu, 20 Aug 2026 00:00:00 +0000

项目背景与目标

人在协作任务中不仅需要观察环境，还需要推测他人的目标、知识状态和下一步行动。例如，两个人一起整理房间时，一个人可能会根据同伴正在走向的位置，推断其想要拿取的物体，并提前让路或递交工具。这类能力通常被称为心理理论（Theory of Mind），也是认知推理、社会协作和人机交互中的核心问题。

当前许多智能体评测环境主要关注单个智能体的任务完成率，较少显式考察智能体是否能够表示并利用“他人可能知道什么、想做什么、会怎么行动”。本项目旨在构建一个轻量级多智能体网格环境 MindGrid，要求智能体在部分可观测、目标不完全共享、需要沟通或协作的任务中完成推理决策。学生需要实现基本环境、构建不同类型的智能体，并分析心理理论建模是否能提升协作表现。

本项目适合作为本科生小组课程作业。重点不在于开发大型平台，而在于清楚定义任务、实现可运行 baseline、给出可解释的实验比较。

核心任务设计

任务一：多智能体协作环境构建（20分）

构建一个简单的二维网格世界，用于模拟两个或多个智能体在部分可观测条件下完成协作任务。环境可以自行实现，也可以基于 MiniGrid、PettingZoo、Overcooked-AI 等已有环境修改。

环境至少应支持以下元素：

多个智能体，每个智能体拥有自己的位置、视野和可执行动作。
多个物体或目标点，例如钥匙、箱子、门、工具、任务区域等。
部分可观测机制，例如智能体只能看到局部视野，或只能知道部分物体状态。
一个需要协作才能完成的任务，例如一名智能体开门、另一名智能体搬运物体，或两名智能体需要交换信息后选择正确目标。

期望产出：

环境代码和运行说明。
至少 3 个不同难度的任务场景。
每个场景的截图或录屏。
对状态空间、动作空间、奖励函数和任务成功条件的说明。

评分维度：

环境可运行性（8分）
任务设计清晰度（6分）
部分可观测与协作机制设计（4分）
可视化或日志记录（2分）

任务二：协作推理智能体实现与对比（40分）

在任务一的环境中实现并比较不同类型的智能体。学生至少需要完成一个规则型 baseline 和一个学习或语言模型驱动的智能体。

可选智能体包括：

规则型智能体：基于手写策略完成导航、交互和简单协作。
强化学习智能体：使用 PPO、DQN、A2C 等方法训练策略。
大语言模型智能体：将环境状态转化为自然语言描述，由 LLM 输出动作或高层计划。
心理理论智能体：维护对其他智能体目标、可见信息或下一步动作的估计，并据此调整自己的策略。

建议比较以下问题：

只根据自身观察行动的智能体，与显式建模同伴状态的智能体相比，表现有何差异？
在部分可观测程度增加时，不同智能体的任务成功率如何变化？
加入简单沟通机制后，智能体是否更容易完成任务？

期望产出：

至少两类智能体的实现代码。
至少一个定量评估指标，例如成功率、平均完成步数、平均奖励或沟通次数。
至少 20 次运行结果的统计表格。
对典型成功和失败案例的简要分析。

评分维度：

baseline 实现完整性（10分）
学习型或 LLM 智能体实现（10分）
心理理论变量或同伴建模设计（10分）
实验对比与结果分析（10分）

任务三：开放探索与改进（40分）

在完成前两项任务的基础上，选择一个开放问题进行深入探索。目标是提出一个具体改进，并用实验说明它是否有效。

可选方向包括：

更好的同伴建模：预测同伴目标、视野、动作或错误信念。
更好的沟通策略：限制通信轮数或消息长度，研究智能体如何选择最有用的信息。
更复杂的任务组合：加入诱饵目标、动态障碍、临时角色分工或任务冲突。
更强的规划能力：结合搜索、树搜索、LLM planning 或 hierarchical policy。
更好的可解释性分析：可视化智能体对同伴目标的估计过程，分析失败原因。

本任务不要求一定超过所有 baseline，但需要清楚说明改进方法、实验设置和结果变化。如果改进失败，也可以得分，但需要分析失败原因。

期望产出：

一个明确的开放问题。
一个可运行的改进方法。
与任务二 baseline 的对比实验。
对结果的解释，包括有效、无效或不稳定的原因。

评分维度：

问题定义清晰度（8分）
方法设计合理性（12分）
实验对比充分性（12分）
结果解释与反思（8分）

数据集与技术栈

推荐环境：

MiniGrid / PettingZoo：用于快速构建多智能体网格任务。
Overcooked-AI：用于协作烹饪任务和多智能体协作 baseline。
自建 Python 网格环境：适合做轻量级课程项目。

推荐技术栈：

Python 3.9+
NumPy / Pandas：环境状态、实验日志和结果分析。
PyTorch / Stable-Baselines3：强化学习 baseline。
OpenAI API / transformers / LangChain：LLM 智能体和自然语言推理模块。
Matplotlib / Seaborn：结果图表。
Streamlit / Gradio：可选，用于展示环境运行过程。

评估标准与预期成果

评分分配（100分）：

任务一：20分，完成多智能体协作环境和任务场景。
任务二：40分，实现并比较不同协作推理智能体。
任务三：40分，完成一个开放探索方向并进行实验分析。

预期交付物：

完整代码：环境、智能体、实验脚本和 README。
实验报告：包括背景、任务定义、方法、实验结果、案例分析和不足反思。
可视化材料：任务场景截图、运行录屏或轨迹可视化。
结果文件：实验日志、统计表格和主要图表。

加分项：

构建网页或交互式界面展示智能体运行过程，最多额外 10 分。
加入人类参与实验，例如让人类观察者判断智能体意图，并与模型估计对比，最多额外 10 分。
将任务扩展到 3 个以上智能体或更复杂的开放环境，最多额外 10 分。

最终成绩超过 100 分时按 100 分计算。

参考文献

Baker, C. L., Saxe, R., & Tenenbaum, J. B. (2009). Action understanding as inverse planning. Cognition, 113(3), 329-349. https://doi.org/10.1016/j.cognition.2009.07.005
Rabinowitz, N. C., Perbet, F., Song, H. F., Zhang, C., Eslami, S. M. A., & Botvinick, M. (2018). Machine theory of mind. In Proceedings of the 35th International Conference on Machine Learning (pp. 4218-4227). PMLR. https://proceedings.mlr.press/v80/rabinowitz18a.html
Carroll, M., Shah, R., Ho, M. K., Griffiths, T. L., Seshia, S. A., Abbeel, P., & Dragan, A. D. (2019). On the utility of learning about humans for human-AI coordination. In Advances in Neural Information Processing Systems 32. https://papers.nips.cc/paper_files/paper/2019/hash/f5b1b89d98b7286673128a5fb112cb9a-Abstract.html
Lowe, R., Wu, Y., Tamar, A., Harb, J., Abbeel, P., & Mordatch, I. (2017). Multi-agent actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information Processing Systems 30. https://papers.nips.cc/paper_files/paper/2017/hash/68a9750337a418a86fe06c1991a1d64c-Abstract.html
Chevalier-Boisvert, M., Willems, L., & Pal, S. (2018). Minimalistic Gridworld Environment for Gymnasium. GitHub repository. https://github.com/Farama-Foundation/Minigrid
Terry, J. K., Black, B., Grammel, N., Jayakumar, M., Hari, A., Sullivan, R., Santos, L. S., Dieffendahl, C., Horsch, C., Perez-Vicente, R., Williams, N. L., Lokesh, Y., & Ravi, P. (2021). PettingZoo: Gym for multi-agent reinforcement learning. In Advances in Neural Information Processing Systems 34. https://proceedings.neurips.cc/paper/2021/hash/7ed2d3454c5eea71148b11d0c25104ff-Abstract.html

OpenMic: 基于多智能体框架的智能脱口秀生成系统

Thu, 20 Aug 2026 00:00:00 +0000

项目背景与目标

随着人工智能在创意内容生成领域的突破，AI辅助的娱乐内容创作成为研究热点。Stand-up comedy作为高度依赖语言技巧、时机把握和情感表达的艺术形式，对AI系统提出独特挑战。本项目要求学生构建基于AutoGen多智能体框架的端到端脱口秀生成系统（Wu et al., 2024），实现从主题输入到专业语音表演的完整流程，深入探索多模态AI协作在创意内容生成中的应用。

系统核心目标是接收用户输入的生活主题（如"校园糗事"、“学霸日常”），自动生成符合中文stand-up comedy特点的3-5分钟表演内容，并转换为具有专业表演效果的语音输出。学生需掌握多智能体系统设计、中文幽默内容生成、语音合成优化等核心技术。

核心任务设计

任务一：AutoGen多智能体系统架构（30分）

使用AutoGen框架构建5个核心智能体的协作系统，通过GroupChat模式实现智能体间的有序对话流程（Wu et al., 2024）。关键智能体包括：ComedyDirector负责整体策略制定和风格控制；JokeWriter专注核心内容创作，基于CFunSet数据集生成setup-punchline结构段子（Yu et al., 2025）；AudienceAnalyzer进行受众适配分析；PerformanceCoach设计语音表达策略和表演标记；QualityController负责内容评估和质量控制。

技术实现采用AutoGen的ConversableAgent和GroupChatManager，设计清晰的角色定义和交互协议（Wu et al., 2024）。工作流程遵循策略制定→受众分析→内容创作→表演指导→质量控制的循环优化模式，确保生成内容的专业性和一致性。

任务二：基于CFunSet的中文幽默内容生成（25分）

深度利用CFunSet中文幽默数据集，构建专门的中文stand-up comedy生成算法（Yu et al., 2025）。核心技术包括主题扩展算法，将简单输入转化为丰富表演素材；Setup-Punchline生成器，学习CFunSet中铺垫-包袱的中文幽默结构模式，实现包袱的延时抖落和多层回调；语言风格引擎，保持口语化表达和网络流行语特色。

推荐使用DeepSeek-V3大模型（DeepSeek-AI, 2024），通过LoRA微调（Hu et al., 2021）或精心设计的prompt engineering实现风格迁移。建立量化评估机制，包括幽默度、文化适配度、结构完整性等关键指标，确保生成内容符合中文stand-up comedy标准。

任务三：专业级语音合成与表演优化（30分）

这是项目的核心创新点，实现具有专业表演效果的语音输出。技术架构包含三个关键模块：情感语音合成模块使用ChatTTS或VALL-E X（Wang et al., 2023; Zhang et al., 2023）实现多情感控制，根据笑点类型动态调整语调；动态节奏控制系统精确控制铺垫语速、包袱前停顿（0.8秒）、重音强调和笑点后停顿（2秒）；语气词处理算法智能插入"呃"、“那个"等自然语气词，平衡自然度和表演感。

语音质量评估采用MOS标准（ITU-T, 1996）结合自定义表演效果评估，包括停顿合理性、重音准确性、情感表达度等维度。系统要求实现16kHz采样率的高质量音频输出，语音合成实时率大于1.0x，确保用户体验流畅。

任务四：系统集成与用户体验（15分）

开发完整的用户界面，支持主题描述输入、表演风格选择（观察类、自嘲类、吐槽类）、时长设定和目标受众选择。系统需实时展示智能体协作过程，提供文字脚本和语音播放功能，支持重新生成和导出功能。技术栈推荐使用Streamlit（Streamlit Inc., 2019）快速搭建前端，FastAPI（Ramirez, 2018）作为后端API，确保内容生成响应时间小于30秒，系统连续运行稳定无崩溃。

数据集与技术栈

核心数据集：CFunSet中文幽默数据集提供相声、小品、网络段子等多种幽默形式，用于学习中文幽默模式和语言特征（Yu et al., 2025）。辅助数据集包括脱口秀节目文本和AISHELL中文语音数据。

推荐技术栈：AutoGen多智能体框架（Wu et al., 2024），DeepSeek-V3中文大模型（DeepSeek-AI, 2024），ChatTTS/VALL-E X语音合成（Wang et al., 2023; Zhang et al., 2023），Streamlit/FastAPI系统框架（Streamlit Inc., 2019; Ramirez, 2018）。开发环境要求Python 3.9+，主要依赖包括autogen、transformers、librosa、soundfile等。

评估标准与预期成果

评分分配（100分）：AutoGen多智能体系统实现（30分）重点考察智能体设计合理性和协作效果；中文幽默内容生成（25分）评估CFunSet利用效果和setup-punchline结构质量；语音合成表演优化（30分）关注语音自然度和表演技巧；系统完整性和用户体验（15分）考察界面友好度和系统性能。

预期交付物：完整系统代码（GitHub仓库）、5个不同主题的3-5分钟音频样本、技术文档包含架构设计和评估结果、10分钟系统演示视频。现场演示要求展示从"我的网购经历"输入到专业脱口秀语音输出的完整流程，体现多智能体协作优势和中文幽默特色。

参考文献

DeepSeek-AI. (2024). DeepSeek-V3 Technical Report. arXiv preprint arXiv:2412.19437.

Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2021). LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685.

ITU-T. (1996). Methods for subjective determination of transmission quality. ITU-T Recommendation P.800.

Ramirez, S. (2018). FastAPI framework, high performance, easy to learn, fast to code, ready for production. Retrieved from https://fastapi.tiangolo.com/

Streamlit Inc. (2019). Streamlit — A faster way to build and share data apps. Retrieved from https://streamlit.io/

Wang, C., Chen, S., Wu, Y., Zhang, Z., Zhou, L., Liu, S., Chen, Z., Liu, Y., Wang, H., Li, J., He, L., Zhao, S., & Wei, F. (2023). Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111.

Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., … & Wang, C. (2024). Autogen: Enabling next-gen LLM applications via multi-agent conversations. In Proceedings of the First Conference on Language Modeling.

Yu, Z., Hu, X., & Wan, X. (2025). CFunModel: A “Funny” Language Model Capable of Chinese Humor Generation and Processing. arXiv preprint arXiv:2503.20417.

Zhang, Z., Zhou, L., Wang, C., Chen, S., Wu, Y., Liu, S., Chen, Z., Li, J., Liu, Y., Wang, H., Li, J., He, L., Zhao, S., & Wei, F. (2023). Speak foreign languages with your own voice: Cross-lingual neural codec language modeling. arXiv preprint arXiv:2303.03926.

RoboLoc: 基于MuJoCo仿真的机器人足球场自定位系统

Thu, 20 Aug 2026 00:00:00 +0000

项目背景与目标

在 RoboCup 机器人足球比赛中，机器人需要在没有外部定位设备辅助的情况下，仅依靠自身传感器持续判断自己在足球场中的位置和朝向。准确的自定位是机器人完成自主行走、球场导航、寻找足球、战术站位和团队协作的基础。当机器人长时间行走时，单纯依赖关节编码器或里程计会产生累计误差；而机器人相机虽然可以观察到足球场中的白线、球门和其他场地特征，但视觉观测本身又存在噪声、遮挡和误匹配。因此，一个真正能够用于机器人足球的自定位系统通常需要融合机器人自身运动信息和环境视觉信息，并通过状态估计方法持续修正机器人位置。

本项目要求学生使用 MuJoCo 搭建一个简化的 RoboCup 足球机器人仿真环境，并在仿真机器人上实现完整的自定位系统。学生需要自行完成机器人模型、足球场环境、相机和相关传感器的配置，使机器人能够在足球场中运动。同时，系统需要利用机器人自身的运动信息和相机获得的足球场视觉信息，对机器人在足球场中的位置和朝向进行实时估计。

项目不要求学生单纯训练一个从图像直接预测机器人坐标的神经网络，而是希望通过 MuJoCo 仿真建立一个具有明确物理意义的机器人系统，并研究“运动模型 + 视觉观测 + 状态估计”这一经典机器人定位框架。学生需要通过实验理解机器人为什么会产生定位漂移、视觉信息如何帮助机器人修正位置，以及不同状态估计方法在 RoboCup 足球场环境中的性能差异。

本项目的最终目标是构建一个能够在 MuJoCo 中运行的 RoboCup 机器人自定位系统。机器人在足球场中运动时，系统能够实时输出机器人估计的二维位置和朝向，并在足球场地图中显示真实轨迹、估计轨迹以及当前定位结果。

系统核心目标

本项目主要围绕 MuJoCo 仿真、机器人感知和自定位三个部分展开。

首先，学生需要在 MuJoCo 中建立一个简化的 RoboCup 足球机器人模型和足球场环境。机器人至少需要具有能够产生运动的身体模型、关节、相机以及用于获取自身运动信息的传感器。足球场需要按照给定的尺寸和场地标记进行建模，使机器人能够在仿真环境中观察到真实的足球场结构。

其次，学生需要从 MuJoCo 中获取机器人自身的运动信息和视觉信息。运动信息可以来自机器人关节状态、机身位姿、IMU 或自行计算的里程计；视觉信息则来自机器人头部或身体上的 RGB 相机。学生需要从相机图像中识别足球场中的有效特征，例如场地边线、中线、中心圆、禁区线、球门等，并将这些信息转换成能够用于定位的观测。

最后，学生需要设计状态估计系统，将机器人运动信息和视觉观测进行融合。系统至少需要实现一种基于概率状态估计的方法，并建议进一步实现 Particle Filter，与 EKF 等方法进行比较。通过连续仿真实验，分析不同方法的定位精度、稳定性和计算开销。

核心任务设计

任务一：MuJoCo 足球机器人与比赛环境建模（35分）

学生首先需要使用 MuJoCo 建立一个简化的 RoboCup 足球机器人仿真环境。机器人可以根据课程提供的基础模型进行修改，也可以自行建立简化的人形机器人模型。模型不要求完全复现真实 RoboCup 机器人，但需要具有合理的身体结构、关节和质量参数，并能够在足球场中完成基本运动。

机器人至少需要配置一个用于足球场观察的 RGB 相机，并能够从 MuJoCo 中获取机器人关节位置、关节速度、机身位姿和相机图像等信息。鼓励学生进一步配置 IMU 等传感器，并根据机器人运动信息计算简单的 odometry。

足球场需要在 MuJoCo 中建立完整的二维几何环境，包括场地边界、中心线、中心圆、禁区以及球门等主要结构。足球场中的关键线条和 landmark 应当具有明确的尺寸和全局坐标，使后续视觉定位可以利用这些信息。

学生需要设计一个能够让机器人在足球场中连续运动的实验场景。例如让机器人沿预先设定的路径行走，或者通过简单的控制器让机器人在足球场中自主移动。机器人运动轨迹由 MuJoCo 中的真实状态作为 Ground Truth，用于后续评价定位算法。

这一阶段重点考察学生是否能够正确理解 MuJoCo 中的机器人模型、坐标系、传感器和仿真状态，并建立一个能够支持后续定位算法的完整实验环境。

任务二：视觉感知与机器人运动信息提取（30分）

在建立 MuJoCo 仿真环境之后，学生需要实现机器人定位所需要的感知模块。机器人相机获得的图像应当能够反映足球场中的主要视觉结构，学生需要从图像中提取能够用于定位的信息。

基础要求可以采用传统计算机视觉方法，例如颜色分割、边缘检测、直线检测和几何特征提取。学生可以首先识别足球场中的白色场地线，再进一步判断这些线在机器人视野中的位置和方向。对于希望进一步探索的学生，也可以使用轻量级目标检测或语义分割网络辅助识别足球场 landmark。

与此同时，需要从 MuJoCo 中获取机器人自身的运动信息。可以利用机器人真实状态计算 Ground Truth，也可以根据关节和 IMU 等传感器信息构造机器人自己的 odometry。需要特别注意，Ground Truth 只能用于评价，不能直接作为定位算法的输入。

最终形成两类信息：一类是机器人认为自己“走了多远”的运动信息，另一类是机器人通过相机观察到的足球场信息。后续定位模块需要利用这两类信息估计机器人当前的位置。

任务三：机器人自定位与状态估计（35分）

学生需要在 MuJoCo 中实现机器人自定位算法，使机器人能够根据自身运动和视觉观测持续估计当前位置。

首先建议实现一个仅依赖运动信息的 baseline。机器人根据自己的 odometry 不断更新位置，并与 MuJoCo 中的真实 Ground Truth 进行比较。通过这个实验，学生应该能够观察到随着机器人运动时间增加，单纯依赖运动信息会逐渐产生累计漂移。

随后加入视觉观测，通过 EKF、UKF 或 Particle Filter 等方法实现视觉与运动信息的融合。其中至少需要完成一种概率状态估计方法；推荐使用 Particle Filter，因为 RoboCup 足球场中存在大量几何对称结构，机器人可能同时存在多个可能的位置假设，而粒子滤波能够自然表示这种不确定性。

Particle Filter 中的每一个粒子代表机器人可能的位置和朝向。机器人运动时，粒子根据运动模型进行传播；获得新的视觉观测后，根据该粒子位置与实际足球场结构之间的一致程度调整粒子权重。经过多次观测之后，大量粒子应该逐渐集中到真实机器人所在的位置附近。

学生还需要研究机器人在不知道初始位置的情况下能否通过视觉逐渐完成全局定位。可以让粒子在整个足球场中随机初始化，然后让机器人开始运动，通过连续视觉观测观察定位结果如何从一个高度不确定的状态逐渐收敛。

最终需要在 MuJoCo 中实时显示机器人真实位置和估计位置，并能够观察机器人运动过程中定位结果的变化。

任务四：定位性能分析与可视化

学生需要对完成的定位系统进行完整的实验评价。系统至少需要在足球场二维地图中显示机器人 Ground Truth 轨迹和估计轨迹，并显示当前机器人位置、朝向以及定位误差。

对于 Particle Filter，还需要可视化当前粒子分布，使用户能够直观看到机器人在不同时间点的定位置信度。例如可以展示机器人刚开始定位时粒子分布较为分散，随着视觉信息不断加入，粒子逐渐集中到真实位置附近。

学生需要至少比较以下几种情况：

仅使用 Odometry、视觉 + 状态估计、全局初始化后的 Particle Filter。

如果实现了 EKF/UKF，也可以进一步比较不同状态估计方法。

鲁棒性测试与评价指标

为了测试定位系统在实际 RoboCup 场景中的可靠性，需要在 MuJoCo 中人为增加一定程度的传感器误差和视觉误差。

可以通过增加 odometry 噪声、IMU 噪声、视觉检测误差以及短时间视觉失效等方式模拟真实机器人比赛中的情况。学生不需要针对每一种情况单独设计复杂实验，而是统一比较定位系统在不同噪声条件下的性能。

主要测试指标包括：

Position Error： 机器人估计位置与 MuJoCo Ground Truth 之间的平均距离误差。

Heading Error： 机器人估计朝向与真实朝向之间的角度误差。

RMSE： 衡量整个运动轨迹上的综合定位误差。

Localization Recovery Time： 当机器人定位发生较大偏差或者视觉短时间失效后，系统重新恢复到正常定位范围所需要的时间。

Computation Time： 单次定位更新所需要的计算时间，用于分析算法是否具有实时运行的可能性。

最终需要比较不同方法在定位精度、鲁棒性和计算效率上的差异。

预期成果

学生最终需要提交一个完整的 MuJoCo RoboCup 自定位系统，包括机器人和足球场仿真模型、机器人运动控制程序、视觉感知程序、定位算法以及结果可视化程序。

系统运行时，机器人能够在 MuJoCo 足球场中运动，并实时输出自身估计的位置和朝向。系统应当能够同时显示机器人真实轨迹和估计轨迹，并能够展示定位误差以及定位状态。

学生还需要提交一份技术报告，报告中需要说明机器人模型设计、传感器配置、视觉感知方法、状态估计方法、实验结果以及误差来源。

评分标准

模块	分数	主要评价内容
MuJoCo机器人与足球场建模	35	机器人模型、相机、传感器、足球场、运动控制
视觉与运动信息提取	30	场地特征检测、运动信息、观测构建
自定位算法	25	状态估计、Particle Filter / EKF、定位效果
实验分析与可视化	10	轨迹、误差、粒子分布、技术报告
总计	100

完成基础任务后，可以通过实现视觉深度学习、传感器融合、全局重定位、动态遮挡处理等进阶功能获得额外加分。

参考文献

https://docs.b-human.de/coderelease2023/perception/localization-features/
Hou, R., Zhu, M., Nam, H., Fernandez, G. I., & Hong, D. W. (2026, July). Fast and robust localization for humanoid soccer robot via iterative landmark matching. In 2026 23rd International Conference on Ubiquitous Robots (UR) (pp. 163-170). IEEE.
Nadiri, F., Banirostam, T., & Rad, A. B. (2025). On a novel localization methodology for humanoid soccer robots via sensor fusion and perspective transformation. International Journal of Intelligent Robotics and Applications, 9(4), 1577-1593.

Syllabus and Logistics

Thu, 20 Aug 2026 00:00:00 +0000

Syllabus

Week 01/Lecture 01 (09.07): Introduction & Logistics
Week 02/Lecture 02 (09.14): Affordance, Functionality, and HOIs (Part 1)
Week 03/Lecture 03 (09.21): Affordance, Functionality, and HOIs (Part 2)
Week 04/Lecture 04 (09.28): Intuitive Physics
Week 05: National Holidays
Week 06/Lecture 05 (10.12): Causality
Week 07/Paper Presentation 1 (10.19): Physical Commonsense (Attendance Required)
Week 08/Midterm (10.26)
Week 09/Lecture 06 (11.02): Project Miterm Presentation (Attendance Required)
Week 10/Lecture 07 (11.09): Tool, Mirroring, and Imitation (Part 1)
Week 11/Lecture 08 (11.16): Tool, Mirroring, and Imitation (Part 2)
Week 12/Lecture 09 (11.23): Intro to Social Commonsense
Week 13/Lecture 10 (11.30): LLM Inference
Week 14/Paper Presentation 2 (12.07): Social Commonsense (Attendance Required)
Week 15 (12.14): Project Final Presentation 1 (Attendance Required)
Week 16 (12.21): Project Final Presentation 2 (Attendance Required)
Week 19 (01.11): Project Report Due (Hard DDL)

Grading

required attendance: 10%
midterm: 10%
4 labs (out of 6): 20%
5 quiz: 10%
paper presentation (in English): 15%
project presentation (in English): 15%
project report (in English and LaTeX): 20%

Project Report

team: 4-5 students
duration: the entire semester
topic: choose one listed on the project page (see also the instructions); topics are exclusive among teams; reserve online
delivery:
1. technical report in English in top conference quality that includes
- insights of the problem
- related work / literature review
- major components of the systems of algorithms
- technical details, e.g., learning, training, parameters
- experimental results
1. publicly available code, e.g., on GitHub

Paper Presentation

team: 4-5 students
duration: up to the presentation session
topic: choose papers listed in one of the 10 topics and additional papers at your own choice
note: no need to cover all the listed papers, but need to convey a coherent message during the presentation
delivery: presentation, in English

Project Presentation

duration: up to the presentation session
content: present the full arc of your semester project, including
- an overview of the project topic and your team’s research question
- division of work among team members
- technical implementation: methods, models, and key design decisions
- challenges encountered and how you addressed them
- results, findings, and anything worth sharing with the class
note: the presentation should tell a coherent story from motivation to outcome; highlight what was non-obvious or surprising
delivery: presentation, in English
LaTeX Templates

TowerCollapse: 基于神经网络的塔倒塌预测模型

Thu, 20 Aug 2026 00:00:00 +0000

项目背景与目标

在现实世界中，预测结构是否稳定（例如一座积木塔是否会倒塌）不仅是一个趣味问题，更是涉及建筑工程、机器人操作、灾害预警等多个领域的关键挑战。人类依靠视觉直觉和经验，往往能在瞬间判断一座塔是否稳固，甚至能预估它倒塌的方向和时间。这种预测能力本质上源自于对物理规律的隐性学习与因果推理能力的结合。

然而，当前的人工智能模型在这类任务上仍存在明显差距。传统的计算方法依赖显式物理引擎模拟，虽然准确，但计算代价高昂、泛化性不足；而端到端的深度学习模型在视觉识别上表现出色，却缺乏物理推断与因果理解能力。如何结合感知、物理建模与推理，让智能体具备“直觉物理感知”，已成为人工智能迈向通用智能的重要一步。

本项目旨在基于深度神经网络，开发一个塔倒塌预测平台 TowerCollapse。该平台将在不确定、动态、物理约束的复杂环境下，推动智能体学习和推理的协同能力：

视觉输入建模：通过卷积神经网络（CNN）快速理解塔的结构布局；

时序推断机制：结合序列建模（RNN/Transformer），预测塔在未来的演化轨迹；

多任务输出：既能判断塔是否会倒塌（分类任务），又能预测倒塌发生的过程模式（时序预测任务）。

最终目标是构建一个可端到端训练的神经网络预测器，让人工智能能够像人类一样 “看一眼就知道塔会不会倒”，并具备一定的因果推断与物理直觉。这不仅在虚拟仿真与认知科学研究中具有学术意义，也在机器人操作、自动化建模、建筑安全、灾害模拟等应用场景中具备重要价值。

核心任务设计

任务一：高复杂度塔结构仿真数据集构建（30分）

TowerCollapse 使用 Blender 生成塔结构及其倒塌过程（会提供参考代码），提供图像序列与标签。支持以下场景：

基于方块积木的塔构建，包含高度、层数、方块大小的变化。
模拟不同材料（木块、金属、石块）的属性差异。
数据集提供图像帧序列 + “倒塌/不倒塌”二分类标签，以及倒塌过程序列。

交付内容：
一个小规模数据集（如 1k 样本），包含不同条件下的塔结构图像与标签；提供一个脚本或录屏展示数据生成过程。

任务二：倒塌过程模型训练（30分）

基于时序建模的预测模型

搭建一个 CNN + RNN/LSTM/Transformer 的模型，用于处理塔倒塌的时序预测。
模型输入为初始图像，输出为未来时刻的图像序列。
展示模型在判断塔是否倒塌任务上的表现，并与静态 CNN 进行对比分析。

任务三：搭建积木塔（40分）

基于预测模型的规划任务

任务目标：利用任务二中训练好的“倒塌过程预测模型”，设计一个能够主动规划积木摆放顺序与位置的系统，使积木塔能够在限定高度下保持稳定。

任务设置：

初始状态：给定一组不同尺寸与形状的积木（如长方体、立方体、不规则形状），以及一个空白平面。
规划目标：在不超过最大层数的条件下（例如 10 层），尽可能搭建更高的积木塔，并且预测其是否会倒塌。
约束条件：每次只能在已有结构的顶部放置一个新的积木，且积木必须稳定放置。

方法设计：

状态建模：使用相机图像或状态描述（积木位置、角度、形状）作为输入。借助任务二中的预测模型，对新结构的稳定性进行前向预测。
动作选择：给定若干候选动作（如不同位置、角度放置新积木），预测模型评估其未来倒塌风险。采用规划/搜索方法（如 Monte Carlo Tree Search, Beam Search）或强化学习方法，选择最优动作。
反馈修正：如果预测结果显示高风险，系统需重新选择放置方式，直至找到相对稳定的方案。

评估指标： 需要讲规划出来的搭建方式在任务一中的Blender物理模拟器中验证。

成功率：在规定高度下成功搭建稳定塔的次数比例。
平均高度：搭建过程中最终积木塔的平均高度。
规划效率：完成一次积木塔搭建所需的规划步骤与时间。

对比实验：

无预测基线：随机放置积木，或仅使用静态启发式（如面积最大化放置）。
仅根据初始场景预测基线：只根据初始图像预测是否倒塌的模型。
动态预测模型：结合任务二训练的时序预测模型，进行动态规划。

数据集与技术栈

核心数据集：TowerCollapse（自建数据集，支持图像、序列标签）。
推荐技术栈：
- 数据生成：Blender, Python 脚本
- 模型训练：PyTorch/TensorFlow
- 可视化：Matplotlib/Seaborn

评估标准与预期成果

评分分配（100分）：

任务一：30分 —— 成功生成并展示数据集
任务二：30分 —— 神经网络实现、实验复现、对比分析
任务三：40分 —— 搭建过程算法设计与展示

预期交付物：

完整代码（数据生成脚本 + 模型训练脚本 + readme file）
详细实验报告（含背景、方法、结果与分析）
模型预测结果截图/录屏示例
创新思路与未来改进方向

参考文献

Lerer A, Gross S, Fergus R. Learning physical intuition of block towers by example. International conference on machine learning, 2016.
Groth O, Fuchs F B, Posner I, et al. Shapestacks: Learning vision-based physical intuition for generalised object stacking, 2018.

VLM中的物体绑定记忆与错误机制分析

Thu, 20 Aug 2026 00:00:00 +0000

项目背景与目标

现代 VLM 已经能够准确识别物体及其颜色、形状、纹理等属性，但识别出这些属性并不意味着模型能够正确记住“哪些属性属于同一个物体”。认知科学中的 memory conjunction 研究发现，人类有时会记住多个刺激的组成特征，却错误地把来自不同刺激的特征重新组合起来，从而产生虚假的“见过”判断 [Reinitz et al., 1992]。

类似现象也可能出现在 VLM 中。例如，模型先后看到多个具有不同颜色、形状和纹理的物体，测试时出现一个从未完整出现过、但其颜色、形状和纹理都分别出现过的新物体。如果模型主要依赖各个特征是否熟悉，而没有保存稳定的物体绑定关系，就可能错误地认为该物体曾经出现。

本项目旨在研究 VLM 是否能够在多图像上下文中保持稳定的物体绑定记忆，以及上下文距离和视觉干扰等因素如何影响绑定错误的发生。在行为实验的基础上，项目还将使用已有的表示探针方法，对模型不同层的内部表示进行初步分析，探索模型能否分别表示单独的物体特征以及这些特征之间的绑定关系。

核心任务设计

任务一：构建并测试 VLM 物体绑定记忆任务（30分）

参考经典 memory conjunction 实验，构建一个可控的视觉物体记忆任务。图像生成可以借鉴 CLEVR [Johnson et al., 2017] 中程序化生成、精确控制视觉属性的思路，自行生成由不同颜色、形状和纹理组合而成的简单单物体图像。

每次实验首先向 VLM 展示若干学习图片，随后要求模型判断测试物体是否曾经完整出现。测试物体应至少包括真正出现过的目标项（target）、所有组成特征都分别出现过但完整组合从未出现过的关键重组诱饵（critical conjunction lure），以及只包含部分已见特征的 near lure 和基本不包含已见特征的 unrelated lure。

选择至少两个 VLM 进行测试，比较不同类型测试物体的正确率和错误接受率，分析模型是否存在“能够正确识别真正见过的物体，却错误接受由已见特征重新组合得到的新物体”的现象。

期望产出：

可重复生成的物体绑定记忆测试数据；
至少两个 VLM 的测试结果；
不同测试物体类型的准确率和错误接受率比较。

任务二：研究上下文与视觉干扰对绑定错误的影响（30分）

在任务一的基础上，进一步研究不同实验条件如何影响绑定错误。学生从上下文距离、干扰物体与学习物体的特征重叠程度、组成特征的重复次数等因素中选择至少两个进行实验。

例如，可以在学习图片和测试图片之间插入不同数量的无关内容，观察上下文距离增加后绑定错误是否更加明显；也可以比较普通无关干扰图片和与学习物体共享颜色、形状或纹理的干扰图片，判断具有相同特征的物体是否会产生更强的记忆干扰。

实验应保证不同条件之间除研究变量外尽量保持一致，并同时关注目标项和各类新物体的表现，从而区分一般性的遗忘和特定的绑定错误。

期望产出：

至少两个实验因素的对照实验；
不同条件下绑定错误变化的统计结果和可视化；
不同 VLM 对实验变量敏感性的比较和分析。

任务三：使用表示探针初步分析绑定错误的内部机制（20分）

行为结果只能说明 VLM 是否发生了绑定错误，并不能直接说明模型内部是否保存了正确的物体绑定信息。本任务选择一个可以获得中间层表示的开源 VLM，使用线性探针（linear probing）对不同网络层中的信息进行分析。

学生可以参考 Huang et al. [2026] 在 Formalizing the Binding Problem 中提出的方法及其公开代码。该工作通过冻结模型并训练简单探针，分析模型表示中是否包含单独的视觉特征以及特征之间的绑定信息，并提供了数据构造、模型表示提取和探针训练的完整实现。

学生需要分别训练探针预测物体的颜色、形状或纹理，以及不同特征是否来自同一个物体，并比较这些信息在模型不同层中的可解码程度。重点分析是否存在“单独特征仍然能够被准确解码，但物体绑定信息明显较弱”的情况，并比较模型回答正确和发生绑定错误时的内部表示差异。

本任务不要求完整定位绑定错误的因果机制，只需要利用已有方法得到一个能够补充行为实验的内部表示分析结果。

期望产出：

基于已有公开方法实现的表示探针实验；
不同网络层中单独特征信息和绑定信息的比较；
对行为错误与内部表示之间关系的初步分析。

任务四：结果分析与研究报告（20分）

综合前三项任务撰写项目报告，总结 VLM 在物体绑定记忆任务上的主要表现，并分析上下文距离、视觉干扰等因素如何影响错误发生。

结合表示探针结果，讨论模型的错误更可能与物体绑定信息本身较弱有关，还是可能发生在后续的信息使用过程中。不要求得到唯一的机制结论，但需要明确说明现有实验能够支持哪些结论、还存在哪些无法排除的解释，以及未来可以通过哪些实验进一步验证。

报告应包括数据构造方法、行为实验结果、不同实验条件的比较、表示探针结果、典型成功与失败案例，以及当前研究的主要局限和可能的改进方向。

数据集与技术栈

数据集：参考 CLEVR 的程序化生成方式，自行构建由颜色、形状和纹理组合而成的简单物体图像。
模型：Qwen、LLaVA 等能够处理多图像输入的 VLM；机制实验选择能够获取中间层表示的开源模型。
数据生成与实验：Python、Pillow、PyTorch、Hugging Face Transformers。
机制分析：参考 Huang et al. [2026] 的 feature-binding probing 方法及公开代码。

评估标准与预期成果

VLM 物体绑定记忆任务构建与行为测试（30分）
上下文与视觉干扰实验（30分）
表示探针机制分析（20分）
研究报告（20分）

总分：100分

预期交付物：

数据生成、模型测试和结果分析代码；
物体绑定记忆测试数据；
行为实验与表示探针实验结果；
项目研究报告。

参考文献

Reinitz, M. T., Lammers, W. J., & Cochran, A. P. S. (1992). Memory-conjunction errors: Miscombination of stored stimulus features can produce illusions of memory. Memory & Cognition, 20(1), 1-11.
Wheeler, M. E., & Treisman, A. M. (2002). Binding in short-term visual memory. Journal of experimental psychology: General, 131(1), 48.
Johnson, J., Hariharan, B., Van Der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., & Girshick, R. (2017). Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2901-2910).
Campbell, D., Rane, S., Giallanza, T., De Sabbata, N., Ghods, K., Joshi, A., … & Webb, T. (2024). Understanding the limits of vision language models through the lens of the binding problem. Advances in Neural Information Processing Systems, 37, 113436-113460.
Huang, L., Li, Y., Salehi, S., Chang, Y., Soni, A., & Kording, K. P. (2026). Formalizing the Binding Problem. arXiv preprint arXiv:2606.03976.

基于Virtual Tool Game实现类人的工具使用学习算法

Thu, 20 Aug 2026 00:00:00 +0000

Background

We have introduced the virtual tool game, which explores the mechanics of human physical problem-solving using the Sample, Simulate, Update (SSUP) model in a 2D virtual tool-use setting. The model suggests that human flexibility comes from simulating the effects of hypothetical actions, while efficiency arises from relying on rich action priors that are continuously updated with real-world observations.

Instructions

In this project, you will implement the SSUP algorithm to reproduce the main findings from the Virtual Tools paper on human tool-use in problem-solving. Its core task is to use one of several tools to move a red object into a green goal area within a 2D physics environment. Your task is to complete the missing SSUP algorithm within the provided game framework and analyze its performance against human data. Through this, you will gain insights into how mental simulation and prior knowledge enable flexible and efficient tool use.

This process will help you explore how action priors and mental simulation contribute to problem-solving flexibility and efficiency. While the paper reports certain technical details (e.g., Gaussian policy, Policy Gradient algorithm), you may use other implementations or algorithms, provided you justify your choices. You are not required to reproduce results exactly, but a thoughtful and reasonable analysis is expected.

Task 1: Setup and Familiarization (20 pts)

Get yourself prepared to the project.

Start by reading the provided paper to understand the theoretical foundation of the SSUP model. Include a brief introduction to the Virtual Tool Game and a summary on the core of the algorithm. (15 pts)
Clone the provided codebase and set up the development environment according to the repository instructions. The game framework is available in this repo, and note that it includes the game itself but does not contain the SSUP algorithm. Include the screenshots of successful installation. (5 pts)

Task 2: SSUP Implementation (50 pts)

Implement the SSUP algorithm and run it across different game levels. It is recommended to wrap the environment with standard gym-like APIs so that it can be compatible with standard RL wrappers like Stable Baseline.

Implement the SSUP algorithm based on your understanding according to the paper. Key procedures include:
1. Sample: Proposing new actions based on a structured, object-based prior. (10pt)
2. Simulate: “Mentally” simulating the outcomes of proposed actions using a noisy physics engine. The reward is based on how much the action reduces the distance between the red object and the green goal. (10 pt)
3. Update: Adjusting beliefs about high-value actions based on both simulated and real-world outcomes. The paper uses a Gaussian mixture model policy for this step, but you may adopt alternatives if you can justify your choice. (20pt)
4. Environment setup: implement logics on environment initialization, stepping, reset, the computation of key metrics, and logging. (10 pt)

Task 3: Result Analysis (30 pts)

Following Fig. 4: Visualize the results (e.g., selected actions, model beliefs, performance, etc.) and analyze them. Your analysis could include but is not be limited to the questions below. Include the necessary visualizations, analysis and discussions in the report. (20 pts)
1. Does the algorithm learn effectively through the SSUP process?
2. How quickly does your model converge?
3. How does it perform across different levels? What are the primary failure modes and their potential causes?
Following Fig. 5: Compare your implementation’s results with the human data (see this page). Use linear regression to assess how closely the performances align. Include the necessary visualizations, analysis and discussions in the report. (10 pts)

Deliverable

A report with details mentioned above.
Codebase of your project, which includes:
1. The source code of your project, with necessary comments on the code.
2. A README.md file that illustrates the repo structure and startup script to replicate the results in your paper.
3. If you build upon the Virtual Tool Game repo, you have to either open source the repo on GitHub with your commits preserved, or provide a git diff log file that illustrates the modifications you have made in the zip file you submit. See this page for how to generate the git diff log, which you should dump into a text file.

References

Allen, K. R., Smith, K. A., & Tenenbaum, J. B. (2020). Rapid trial-and-error learning with simulation supports flexible tool use and physical reasoning. Proceedings of the National Academy of Sciences, 117(47), 29302-29310.

基于人体视频的人形机器人实时动作生成与控制

Thu, 20 Aug 2026 00:00:00 +0000

项目背景与目标

随着人形机器人技术的发展，如何通过自然、低成本的方式实时控制机器人，逐渐成为具身智能领域的重要研究方向。传统遥操作系统通常依赖动作捕捉设备、惯性传感器或专业控制器，部署成本较高，也难以充分利用大规模人体视频数据。

第三视角人体视频包含丰富的姿态和运动信息，但从视频直接预测机器人动作仍面临深度信息缺失、人体遮挡、相机运动以及人体与机器人结构差异等问题。此外，视觉模型生成的运动学动作未必满足机器人的动力学和平衡约束，需要通过动作追踪控制器才能稳定部署到机器人平台。

本项目旨在构建一个基于第三视角人体视频的人形机器人实时控制系统，实现从视频输入、机器人运动学动作生成到仿真及真机执行的完整流程。具体目标包括：

整合多个大规模人体视频数据集，构建视频—机器人动作配对数据；
训练监督学习模型，实现从人体视频到机器人逐帧关节旋转的预测；
训练通用动作追踪控制器，并完成仿真与真实机器人部署；
探索生成动作的物理可行性优化方法，提高系统的稳定性与动作质量。

核心任务设计

任务一：视频—机器人动作数据集构建（35分）

整合BEDLAM、BEDLAM2.0和Human3.6M数据集，完成视频、人体模型、相机参数、帧率、坐标系及动作表示的统一。使用GMR动作重定向算法，将人体动作批量转换为目标人形机器人的关节动作，使重定向结果满足机器人的骨骼结构、关节自由度和关节限位等基本约束。

进一步完成异常动作过滤、视频与动作时间同步、训练集划分和动作可视化，形成统一的视频—人体动作—机器人动作数据处理管线。

任务二：视频驱动的机器人动作生成（35分）

构建基于监督学习的机器人运动学动作生成模型，以第三视角人体视频或连续视频帧为输入，输出机器人每一帧的全身关节旋转、根节点姿态及必要的运动状态。

模型需要学习人体外观、空间姿态和时序运动之间的关系，并处理人体遮挡、动态相机、复杂背景和快速动作等情况。训练过程中应保证预测动作的准确性、连续性和完整性，减少关节抖动、姿态漂移及动作突变。

同时，实现离线视频和摄像头视频流两种推理方式，并对模型进行推理加速，使其满足实时控制需求。

任务三：动作追踪控制器与机器人部署（30分）

参考SONIC等人形机器人全身动作追踪方法，基于任务一生成的机器人动作数据训练通用动作追踪控制器。控制器接收模型预测的运动学动作，并输出机器人可执行的控制指令，使机器人能够在动力学环境中稳定复现目标动作。

在Isaac Lab或MuJoCo中搭建目标人形机器人模型，验证控制器对不同动作、输入噪声和预测误差的鲁棒性。随后集成视频动作生成模型与动作追踪控制器，形成“人体视频—机器人关节动作—底层控制指令”的完整实时系统。

最终将系统部署至真实人形机器人，完成通信接口、实时推理、状态反馈和安全保护等模块，并比较仿真与真机环境中的动作追踪精度、稳定性和端到端延迟。

任务四：动作物理可行性优化（可选加分项，20分）

使用强化学习对模型生成的机器人运动学动作进行优化，在保持原始人体动作特征的同时，减少脚底滑动、关节超限、身体失衡和动作不连续等问题。

将优化后的动作作为追踪控制器的参考动作或新的监督训练数据，并比较优化前后动作在仿真环境中的物理可执行性、动作质量和追踪成功率。

数据集与技术栈

本项目将使用BEDLAM、BEDLAM2.0和Human3.6M数据集，覆盖合成与真实、多视角与动态相机等不同视频条件。使用GMR完成人体动作到人形机器人动作的批量重定向，并根据目标机器人平台构建统一的动作表示。

开发环境采用Python 3.10+、PyTorch 2.0+和CUDA 12.0+，使用OpenCV完成视频处理，使用NumPy和Pandas完成数据整理。机器人控制与策略训练采用Isaac Lab或MuJoCo，实验记录采用Weights & Biases。真机部署还需集成摄像头输入、GPU实时推理、机器人通信和安全控制模块。

预期成果

预期技术成果包括：

统一的多源人体视频与动作数据处理框架；
基于GMR构建的视频—机器人动作配对数据集；
视频驱动的机器人运动学动作生成模型；
支持实时推理的视频流控制系统；
通用人形机器人动作追踪控制器；
仿真环境与真实机器人上的动作执行演示。

预期学术成果包括一份详细描述数据构建、模型设计、控制策略及部署结果的技术报告，公开或可复现的项目代码，以及可能的机器人、计算机视觉或具身智能相关会议论文投稿。

参考文献

Black, M. J., Patel, P., Tesch, J., and Yang, J. “BEDLAM: A Synthetic Dataset of Bodies Exhibiting Detailed Lifelike Animated Motion.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.

Tesch, J., et al. “BEDLAM2.0: Synthetic Humans and Cameras in Motion.” Advances in Neural Information Processing Systems, 2025.

Ionescu, C., Papava, D., Olaru, V., and Sminchisescu, C. “Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments.” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2014.

Araujo, J. P., et al. “Retargeting Matters: General Motion Retargeting for Humanoid Motion Tracking.” 2026.

Luo, Z., et al. “SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control.” 2026.

基于多模态生成模型的人形机器人动作生成与控制

Thu, 20 Aug 2026 00:00:00 +0000

项目背景与目标

当前，人形机器人研究正处于从被动控制向自主智能转型的关键阶段。传统的人形机器人控制研究主要集中于底层运动控制，包括预编程舞蹈动作、精确关节力矩控制以及遥操作系统等。这些方法虽然能够在特定场景下取得良好效果，但通常缺乏对复杂环境的自主感知、理解与任务执行能力（Brohan et al., 2023）。

近年来，大语言模型（Large Language Model，LLM）和视觉—语言模型（Vision-Language Model，VLM）的快速发展，为提升机器人智能水平带来了新的机遇。GPT系列、LLaMA等模型展现出较强的语言理解与推理能力，而CLIP、BLIP等多模态模型则实现了视觉信息与自然语言之间的有效对齐。这些技术突破为机器人通过自然语言指令和视觉感知理解并执行复杂任务提供了新的技术路径。

本项目旨在构建一个基于多模态生成模型的人形机器人自主动作生成与执行系统，实现从高层任务描述到底层动作控制的完整流程。具体目标包括：

基于预训练动作生成模型，实现从自然语言指令到人体动作序列的生成；
实现动作重定向算法，将人体动作转换为符合人形机器人运动学约束的动作序列；
通过模仿学习实现生成动作的控制与执行，并在仿真环境中验证系统的有效性。

核心任务设计

任务一：动作序列生成与文本对齐（40分）

基于HumanML3D数据集，完成文本描述与人体动作序列的筛选、整理和对齐。构建包含基础动作（如“抬手”“转身”“抓取”等）的动作词汇表，设计适合机器人执行的文本—动作配对数据集，数据集应包含至少500组样本。在此基础上，实现从自然语言描述到人体动作序列的生成模型。

任务二：动作重定向（30分）

实现人体动作到人形机器人动作的重定向算法，将人体动作序列转换为符合机器人关节限位、运动学结构和平衡约束的动作序列。

任务三：模仿学习与仿真验证（30分）

通过模仿学习将重定向后的动作序列转换为机器人控制指令，并在仿真环境中进行验证。在Isaac Lab或MuJoCo仿真环境中搭建人形机器人模型，基于已有数据集中的动作序列，采用行为克隆等方法训练控制策略，使机器人能够复现前述任务中生成并完成重定向的动作。

任务四：高级任务设计（可选加分项，20分）

在基础动作生成与执行系统的基础上，设计并实现物体操作、自主导航或人机交互等高级任务。

数据集与技术栈

本项目将使用HumanML3D数据集。该数据集包含14,616个动作序列和44,970条文本描述，可用于训练和评估文本驱动的人体动作生成模型。

开发环境采用Python 3.8+、PyTorch 1.12+和CUDA 11.0+，并使用GPU进行模型训练与推理加速。核心框架包括用于加载和微调预训练模型的Transformers，以及用于搭建机器人仿真与强化学习环境的Isaac Lab。其他工具包括用于数据处理的NumPy和Pandas、用于图像及视频处理的OpenCV，以及用于实验记录与管理的Weights & Biases。

预期成果

预期技术成果包括：

经过筛选和整理的文本—动作配对数据集；
训练完成的动作生成模型权重及推理代码；
人体动作到人形机器人动作的重定向算法；
基于模仿学习的机器人动作控制策略；
仿真环境中的动作执行演示视频。

预期学术成果包括一份详细描述系统设计、算法实现和实验结果的技术报告，代码，以及可能的会议论文投稿。

参考文献

Guo, C., et al. “Generating Diverse and Natural 3D Human Motions from Text.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.

Jiang, N., et al. “Scaling Up Dynamic Human-Scene Interaction Modeling.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.

Van den Oord, A., and Vinyals, O. “Neural Discrete Representation Learning.” Advances in Neural Information Processing Systems, 2017.

基于机器学习的移动机器人运动规划方法

Thu, 20 Aug 2026 00:00:00 +0000

项目背景与目标

移动机器人全身操作需要协调移动底盘、机械臂与夹爪，并处理关节限位、环境碰撞和任务约束。相比单独规划机械臂或移动底盘，全身移动操作具有更高的状态维度和更复杂的运动约束。

本项目支持 MuJoCo 和 Isaac Lab 两条仿真实现路线，学生可以任选其一。学生可以使用仿真平台已有的场景和任务，也可以自行设计抓取、搬运、放置、开关抽屉或柜门等移动操作任务。

学生首先需要使用一种传统规划或控制方法生成任务轨迹。方法不作统一限制，可以采用模型预测控制、轨迹优化、采样规划、全身逆运动学或多种方法的组合。学生也可以通过脚本控制或遥操作等方式采集示范轨迹。在此基础上，学生需要采集机器人状态、动作和视觉观测等数据，构建任务示范数据集。

随后，学生需要使用一种基于行为克隆的学习方法对示范数据进行训练，例如 Diffusion Policy、3D Diffusion Policy（DP3）或 Action Chunking with Transformers（ACT），并在仿真环境中进行推理和任务验证。

本项目的目标是构建一套完整的移动机器人全身操作流程，包括任务设计、轨迹生成、示范数据采集、行为克隆训练以及仿真验证，并比较不同方法在任务成功率、执行效率和泛化能力等方面的表现。

核心任务设计

任务一：安装并测试仿真环境（10分）

学生可从以下两条路线中任选其一。

MuJoCo 路线：

打开 RoboCasa 项目网站，下载并安装。RoboCasa 基于 robosuite 和 MuJoCo，按照项目提供的安装说明配置环境即可。

安装完成后，运行 RoboCasa 提供的移动操作示例，并进行截图记录。

Isaac Lab 路线：

打开 LW-BenchHub 项目网站，下载并安装。LW-BenchHub 基于 Isaac Lab–Arena 和 Isaac Sim，提供了厨房场景、机器人和操作任务。

安装完成后，运行 LW-BenchHub 提供的移动操作示例，并进行截图记录。

运行示例时，应确认能够正常加载机器人、场景和交互物体，并能够读取机器人状态、相机图像以及向移动底盘、机械臂和夹爪发送控制指令。

期望产出：仿真环境测试代码和运行截图。

任务二：安装并测试轨迹生成与模仿学习算法（20分）

学生可以根据所选任务，自行选择轨迹和示范数据的生成方法，包括模型预测控制、轨迹优化、采样规划、全身逆运动学、脚本控制或遥操作等。

可参考以下机器人规划与建模代码库：

随后，从以下行为克隆方法中任选一种，下载并安装对应的代码库：

安装完成后，学生可以按照代码库提供的说明下载示例数据，使用项目提供的专家策略或脚本生成示范数据，也可以自行生成数据或通过遥操作采集数据。

学生需要生成或采集至少一条可以在仿真环境中回放的示例轨迹，并使用示例数据运行一次训练和推理。

期望产出：示例轨迹、轨迹生成或遥操作代码、轨迹回放视频，以及模仿学习算法的训练和推理截图。

任务三：在仿真环境中构建任务，采集移动机器人的运动轨迹和视频（30分）

目标是在所选仿真平台中搭建任务环境，生成或采集移动机器人的全身操作轨迹，同时采集视频数据。场景既可从 RoboCasa 或 LW-BenchHub 提供的已有任务中选取并进行修改，也可由学生自行构建。

任务由学生自定义，例如简单的“抓取—放置”任务，明确物体的初始状态和目标状态即可。任务应同时使用移动底盘、机械臂和夹爪。

本任务主要分为两个步骤：生成或采集轨迹，与采集轨迹对应的视频。

在轨迹生成阶段，学生可以使用模型预测控制、轨迹优化、采样规划、全身逆运动学等方法，根据自己定义的物体初始状态和目标状态，控制移动机器人完成任务。学生也可以使用键盘、手柄、SpaceMouse 或 VR 设备，通过遥操作采集轨迹。物体的抓取姿态可手动指定，不必自动生成。本阶段成功的话，应该可以在图形界面中观察机器人完成抓取、移动和放置过程。

在视频采集阶段，学生需要使用仿真环境提供的 camera 功能，采集轨迹对应的 RGB、RGB-D 或点云数据。采集的数据格式应与任务二中选择的模仿学习算法相匹配。

期望产出：

场景与任务定义：构建一个仿真场景，在场景中设计机器人需要执行的具体任务，明确交互物体、物体的初始状态以及目标状态。
生成或采集的轨迹：针对上述场景与任务定义，生成或采集机器人完成任务的全身操作轨迹。
采集的轨迹视频：根据采集到的轨迹，录制仿真环境下的动作视频或图像序列。

加分项：构建多个场景和任务，统计和报告每个场景的几何特征，包括障碍物数量、类型、尺寸和位置等，并在这些场景上进行数据采集。（最多额外10分）

加分项2（难度非常大）：尝试完成开门任务并采集轨迹和视频。（最多额外30分）

任务四：使用采集的数据进行行为克隆训练和推理，测试效果（20分）

本任务的目标是使用任务三采集的数据进行模型训练和推理，并验证结果。本任务分为两个步骤：模型训练，推理仿真验证。

模型训练阶段，学生需要将采集的机器人状态、动作和视觉观测数据转换为任务二所选行为克隆算法使用的数据格式，并使用 Diffusion Policy、3D Diffusion Policy 或 ACT 进行训练。

推理仿真验证阶段，学生需要在仿真环境中运行训练得到的模型，根据机器人状态和视觉观测预测动作，控制移动机器人完成所定义的任务。

学生需要更改训练轨迹数量、观测数据类型、图像预处理方式或网络参数等变量，分析这些因素对训练和推理结果的影响。

期望产出：

训练得到的模型：使用任务三采集的数据完成训练。
任务验证代码：使用训练模型在仿真环境中进行推理和控制的代码。
任务验证实验结果：记录任务成功率，并对不同训练设置的结果进行比较。

加分项：增加使用另一种行为克隆算法，并比较不同算法的实验结果。（最多额外10分）

任务五：撰写报告，总结轨迹生成、数据采集、训练和推理过程中的主要问题（20分）

基于以上已完成的工作，总结轨迹生成和数据采集阶段的问题（比如：轨迹生成速度慢、遥操作数据质量不稳定、视频质量不好等），以及训练和推理过程中的各种问题（如：训练效果差、推理轨迹不稳定、任务成功率低等），结合相关文献调研，撰写一份报告。

报告应包括：

仿真环境和算法示例的运行截图；
场景和任务的具体定义；
轨迹生成或遥操作数据采集使用的具体方法；
采集的数据数量、数据格式和数据质量；
模型训练及验证阶段使用的参数；
轨迹生成、数据采集和模型推理阶段的主要瓶颈；
不同实验设置的结果对比；
潜在的改进思路。

期望产出：一份完整的项目报告，以及能够说明任务执行效果的成功和失败视频。

预期交付物

代码库：任务环境、轨迹生成或遥操作、数据采集、训练和推理流程的可复现代码。
数据：任务三采集的机器人状态、动作和视觉观测数据。
模型：任务四训练得到的模型及相关配置。
视频：轨迹生成、模型推理成功和失败的代表性视频。
报告：按照任务五要求撰写的项目报告。

最终评分将根据各环节完成情况打分。任务一至任务五共100分，额外加分计入总分，如果总分超过100分取100分。

参考资料

Introduction

Wed, 20 Aug 2025 00:00:00 +0000

Reading list

survey/perspective paper book YouTube

Required

[YouTube] A DARPA Perspective on Artificial Intelligence
[YouTube] The Exciting, Perilous Journey Toward AGI
[YouTube] Scaling Intelligence the Human Way
[YouTube] Sam Altman and Lex Fridman
[YouTube] Why AI Is Incredibly Smart and Shockingly Stupid
Scientific discovery in the age of artificial intelligence
How to Grow a Mind: Statistics, Structure, and Abstraction, Science 2011
Building machines that learn and think like people, BBS 2016
Dark, Beyond Deep: A Paradigm Shift to Cognitive AI with Humanlike Common Sense (Section 1 and 2), Engineering 2020

Lab 1: SMPL-X Human Body Modeling

Wed, 20 Aug 2025 00:00:00 +0000

Background

This assignment serves as a comprehensive recap of an important topic covered in our course: human body representation using SMPL-X. You will work with 3D human pose visualization and joint-to-smplx optimization that are widely used in computer vision and graphics research.

SMPL-X (SMPL eXpressive) is a expressive body model that represents the human body shape and pose through a parametric model. It extends the original SMPL model to include facial expressions and hand articulation, making it suitable for full-body human modeling applications.

In this assignment, you will implement key components for this topic, demonstrating your understanding of 3D human modeling.

Learning Objectives

By completing this assignment, you will:

Understand the SMPL-X parametric human body model
Gain hands-on experience with PyTorch and 3D graphics libraries

Environment Setup

Prerequisites

Python 3.8 or higher
Linux environment (recommended)
CUDA-capable GPU (recommended)

Installation

Unzip and navigate to the lab directory

Create and activate a conda virtual environment:

conda create -n core_lab1 python=3.8
conda activate core_lab1

Install Required Packages
```
pip3 install -r requirements.txt
```

Verify Installation:

python -c "import torch; import smplx; import trimesh; print('Environment setup successful')"

Assignment Tasks

Task 1: SMPL-X Pose Visualization (50 points)

Objective: Implement a function to visualize 3D human poses using the SMPL-X model.

Description: You will complete the visualize_smplx() function in framework.py. Given SMPL-X parameters (translation, body pose, and global orientation), your function should generate and display a 3D mesh visualization of the human pose and save it in .png format.

Key Requirements

Load the SMPL-X model correctly

Expected Output

3D mesh visualization saved as image .png file

Task 2: Joints to SMPL-X Pose Regression (50 points)

Objective: Implement an optimization function to estimate SMPL-X parameters from 3D joint positions.

Description: Sometimes you can only get the 3D joint positions of a human pose, therefore you need to convert the joint positions to SMPL-X parameters. Complete the optimize_smplx() function that takes 3D joint positions as input and optimizes SMPL-X parameters to best fit these target joints. This involves setting up an optimization loop with appropriate loss functions.

Key Requirements

Define appropriate loss function (joint position error)
Implement gradient-based optimization using PyTorch
Return optimized SMPL-X parameters

Expected Output

Optimized SMPL-X parameters (translation, body_pose, global_orient)
Final loss value (about 8e-5 after 100 iterations)
You should achieve a similar mesh rendering result as in Task 1

Implementation Guidelines

Code Structure

Only modify code within ## TODO sections in framework.py
Do not change function signatures or import statements
Maintain consistent tensor shapes and data types

Debugging Tips

Use breakpoint() to check intermediate results
Visualize intermediate results when possible

Hand-in Requirements

Submission Format

Submit only one file: framework.py
Platform: Gradescope autograder

File Requirements

Your framework.py must contain all implemented functions
Do not modify function signatures or imports
Include necessary comments explaining your implementation approach

Autograder Testing

Your submission will be tested against multiple test cases
Partial credit will be awarded for partially correct implementations
Timeout limit: 10 minutes in total

Academic Integrity

This is an individual assignment. While you may discuss general concepts with classmates, all submitted code must be your own work. Plagiarism will result in course failure.

Allowed

Discussing assignment requirements and clarifications
Sharing debugging strategies (without code)
Using official documentation and course materials

Not Allowed

Sharing or copying code solutions
Using external implementations without permission
Collaborating on actual implementation

Additional Resources

Documentation

Good luck with your assignment!

Basics

Sat, 10 Dec 2022 00:00:00 +0000

Math

MIT 18.01SC: Single Variable Calculus [Fall 2010]
MIT 18.02SC: Multivariable Calculus [Fall 2010]
MIT RES.18-007: Calculus Revisited: Multivariable Calculus [1971]
MIT 18.06SC: Linear Algebra [Fall 2011]
MIT RES.18-010: A 2020 Vision of Linear Algebra [Spring 2020]
Essence of linear algebra
Stanford EE364a: Convex Optimization
MIT 15.084J: Nonlinear Programming [Spring 2004]

CS

General

计算机教育中缺失的一课 [English (recommended)] [Chinese]
计算机自学指南

Python

Data Structure

Algorithm

MIT 6.006: Intro to Algorithms [Fall 2011] [Spring 2020]
MIT 6.046J: Design And Analysis Of Algorithms [Spring 2015]

Survival Skills

English

Mathematical English

Linux

鸟哥私房菜

Git

On Shit, Git!?! (explicit) or Dangit, Git!?! (clean)

LaTeX

Handouts

Mon, 07 Sep 2026 00:00:00 +0000

Slides

Google Drive

Labs

Google Drive

Lab 2: Linear Inverted Pendulum Model

Tue, 16 Sep 2025 00:00:00 +0000

Background

This assignment focuses on a fundamental concept in bipedal robotics: the Linear Inverted Pendulum Model (LIPM). You will implement and visualize a simplified 2D LIPM system that captures the essential dynamics of walking robots and is widely used in humanoid robotics research.

The Linear Inverted Pendulum Model (LIPM) simplifies bipedal walking dynamics by modeling the robot’s center of mass (CoM) as a point mass moving under gravity while constrained to a constant height. This abstraction enables real-time motion planning and control for walking robots by reducing the complex full-body dynamics to tractable equations.

In this assignment, you will implement key components of LIPM dynamics and foot placement strategies.

Learning Objectives

By completing this assignment, you will:

Understand the dynamics of the Linear Inverted Pendulum Model
Implement CoM trajectory computation for single and multiple steps
Learn capture point theory for stable foot placement

Environment Setup

Prerequisites

Python 3.8 or higher
Linux environment (recommended)

Installation

Unzip and navigate to the lab directory

Create and activate a conda virtual environment:

conda create -n core_lab2 python=3.8
conda activate core_lab2

Install Required Packages
```
pip3 install -r requirements.txt
```

Verify Installation:

python -c "import numpy; import matplotlib; print('Environment setup successful')"

Assignment Tasks

Task 1: Single Step LIPM Dynamics (50 points)

Objective: Implement the core dynamics of LIPM by computing the center of mass trajectory during a single step.

Description: You will complete the updateXtVt() function in LIPM.py. This function computes the current position and velocity of the center of mass based on the LIPM dynamics equations.

Testing Your Implementation

Run the following command to visualize a single step:

python test.py --task single_step

Expected Output

Smooth trajectory of the CoM during one step
Reference output video: ./expected_output/single_step.mp4

Task 2: Multi-Step Walking with Capture Point (50 points)

Objective: Implement orbital energy computation and capture point calculation for stable multi-step walking.

Description: Complete two functions that enable stable walking over multiple steps:

getOrbitalEnergy(): Computes the orbital energy of the LIPM system, a conserved quantity useful for stability analysis
updateCapturePoint(): Calculates the capture point, which determines the optimal foot placement location for maintaining balance

These functions work together to enable leg switching and continuous walking by determining where to place the next step iteratively.

Key Requirements

Correctly compute orbital energy based on current state
Implement capture point dynamics

Testing Your Implementation

Run the following command to visualize multi-step walking:

python test.py --task multi_step

Expected Output

Continuous walking motion over multiple steps
Reference output video: ./expected_output/multi_step.mp4

Implementation Guidelines

Code Structure

Only modify code within ## TODO sections in LIPM.py
Do not change function signatures or class structures

Debugging Tips

Verify energy conservation in your orbital energy calculation
Use breakpoint() to check intermediate results
Compare your output with the provided reference videos

Hand-in Requirements

Submission Format

Submit only one file: LIPM.py
Platform: Gradescope autograder

File Requirements

Your LIPM.py must contain all implemented functions
Do not modify function signatures or imports

Academic Integrity

This is an individual assignment. While you may discuss general concepts with classmates, all submitted code must be your own work. Plagiarism will result in course failure.

Allowed

Discussing theory and concepts
Sharing debugging strategies (without code)
Using course materials and recommended references

Not Allowed

Sharing or copying code solutions
Using external implementations without permission
Collaborating on actual implementation

Good luck with your assignment!

Affordance, Functionality, and HOIs

Wed, 20 Aug 2025 00:00:00 +0000

Reading list

survey/review/perspective paper book GitHub

Required - Affordance

Dark, Beyond Deep: A Paradigm Shift to Cognitive AI with Humanlike Common Sense (Section 5), Engineering 2020
Inferring Forces and Learning Human Utilities From Videos, CVPR 2016

Required - HOI/HSI

Required - Functionality

Scene Parsing by Integrating Function, Geometry and Appearance Models, CVPR 2013
Make it Home: Automatic Optimization of Furniture Arrangement, SIGGRAPH 2011

Required - Humanoid Robotics

DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills, SIGGRAPH 2018
SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control, Science Robotics 2026

Optional - Affordance in General

Visual Affordance and Function Understanding: A Survey, ACM Computing Surveys 2021
The Ecological Approach to Visual Perception, Boston: Houghton Mifflin (1979)
Understanding Context: Environment, Language, and Information Architecture (Chapter 4), O’Reilly Media, Inc. (2014)

Optional - Scene Affordance

Holistic 3D Scene Parsing and Reconstruction from a Single RGB Image, ECCV 2018
A Multi-Scale CNN for Affordance Segmentation in RGB Images, ECCV 2016
Putting Humans in a Scene: Learning Affordance in 3D Indoor Environments, CVPR 2019

Optional - Object Affordance

Reasoning about Object Affordances in a Knowledge Base Representation, ECCV 2014
O2O-Afford: Annotation-Free Large-Scale Object-Object Affordance Learning, CoRL 2022
Hallucinated humans: Learning latent factors to model 3D environments, Diss. Cornell University, 2015

Optional - HOI/HSI

Reconstructing Hand-Object Interactions in the Wild, CVPR 2021
Synthesizing Diverse and Physically Stable Grasps With Arbitrary Hand Structures Using Differentiable Force Closure Estimator, RA-L 2021
GenDexGrasp: General Dexterous Grasping, ICRA 2023
Modeling 4D Human-Object Interactions for Event and Object Recognition, CVPR 2013
Full-Body Articulated Human-Object Interaction, ICCV 2023
HUMANISE: Language-conditioned Human Motion Generation in 3D Scenes, NIPS 2022
Diffusion-based Generation, Optimization, and Planning in 3D Scenes, CVPR 2023

Optional - Functionality

Human-centric Indoor Scene Synthesis Using Stochastic Grammar, CVPR 2018
Recognition of natural scenes from global properties: Seeing the forest without representing the trees, Cognitive Psychology 2009
Shape2Pose: Human-Centric Shape Analysis, SIGGRAPH 2014
Modeling the Shape of the Scene: A Holistic Representation of the Spatial Envelope, IJCV 2001
Understanding Bayesian rooms using composite 3D object models, CVPR 2013

Optional - Humanoid Robotics

CLONE: Closed-Loop Whole-Body Humanoid Teleoperation for Long-Horizon Tasks, CoRL 2025
OmniXtreme: Breaking the Generality Barrier in High-Dynamic Humanoid Control, RSS 2026
Ψ₀: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation, RSS 2026
ZEST: Zero-shot Embodied Skill Transfer for Athletic Robot Control, Science Robotics 2026
GROVE: A Generalized Reward for Learning Open-Vocabulary Physical Skill, CVPR 2025
CLoSD: Closing the Loop between Simulation and Diffusion for multi-task character control, ICLR 2025
MaskedMimic: Unified Physics-Based Character Control Through Masked Motion Inpainting, SIGGRAPH Asia 2024
AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control, SIGGRAPH 2021
PARC: Physics-based Augmentation with Reinforcement Learning for Character Controllers, SIGGRAPH 2025
MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives, SIGGRAPH 2026

The Essence of AI

Sat, 10 Dec 2022 00:00:00 +0000

Intro to AI

MIT 6.034: AI [Fall 2010]
Berkeley CS188: Intro to AI [Fall 2022] [Fall 2021] [Fall 2020]
Harvard CS50: Intro to AI with Python [edX]
Stanford CS221: AI: Principles and Techniques [Autumn 2019] [Autumn 2021]

Perspectives on AI by Prof. Song-Chun Zhu

Intro to ML

Prof. Ying Nian Wu at UCLA has organized some fantastic machine learning course material and videos for undergrads. You can view on his homepage. See also A Note on Machine Learning Methods.
西瓜书
Stanford CS229: ML [Autumn 2018]
Stanford CS230: Deep Learning [Autumn 2018]
Intro to RL with David Silver
Deep RL by OpenAI Spinning Up

Advanced ML

Pattern Recognition and Machine Learning
The Elements of Statistical Learning
Machine Learning: a Probabilistic Perspective
Probabilistic Machine Learning: An Introduction
Probabilistic Machine Learning: Advanced Topics
Stanford CS224W: ML with Graphs [Fall 2021]
Stanford CS234: RL [Winter 2019]
Stanford CS330: Deep Multi-Task and Meta Learning [Autumn 2019] [Autumn 2020] [Autumn 2021]
Neurosymbolic Reading Group

Intro to CV

Siyuan Huang and I are offering a CV course at PKU
Intro to Computer Vision by Noah Snavely
Computer Vision by Justin Johnson and David Fouhey
Deep Learning for Computer Vision by Justin Johnson
Advances in Computer Vision by Bill Freeman
MATLAB Functions for Multiple View Geometry by Andrew Zisserman

Advanced CV

Stanford CS231n: CNN for Visual Recognition [Spring 2017]
Stanford CS25: Transformers United [Autumn 2021]
Computer Vision: Statistical Models for Marr’s Paradigm
Computer Vision: Stochastic Grammars for Parsing Objects, Scenes, and Events
Cognitive Models for Visual Commonsense Reasoning
Monte Carlo Methods

Intro to NLP

Speech and Language Processing
Course and Models by Hugging Face

Advanced NLP

Stanford CS224N: NLP with Deep Learning [Winter 2019] [Winter 2021]
Stanford CS224U: NLU [Spring 2019] [Spring 2021]
NYU Computational Linguistics and Cognitive Science [Spring 2023]

Intro to Psychology

改变心理学的40项研究
对伪心理学说不
心理学与生活
津巴多普通心理学
认知心理学
感觉与知觉
MIT 9.00SC: Intro to Psychology [Fall 2004] [Fall 2011]
MIT 9.13: The Human Brain [Spring 2019]

Intro to CoRe

Advanced CoRe

MIT RES.9-003: Brains, Minds And Machines Summer Course [Summer 2015]
MIT 6.868J: The Society of Mind [Fall 2011]
Causal Cognition by Tobi Gerstenberg
Intro to Causal Inference by Brady Neal
Language and Thought by Noah D. Goodman
Probabilistic Models of Cognition
Communication, Intentionality, and the Origins of Language by Michael C. Frank and Fei Xu
A minimalist guide to program synthesis by Evan Pu
Nancy’s Brain Talks by Nancy Kanwisher
Categories and Concepts by Brenden Lake
Github Awesome AGI and CoCoSci maintained by Yu-Zhe Shi

Intro to Robotics

Isaac Sim
ROS
Intro to Robotics by Marc Toussaint
Stanford CS223A: Intro to Robotics
MIT 6.4210/6.4212: Robotic Manipulation [Fall 2022]
Northwestern Modern Robotics
学用Isaac Sim快速入门机器人仿真开发

Advanced Robotics

MIT 6.832: Underactuated Robotics [Spring 2019] [Spring 2022]
MIT 16.412J: Cognitive Robotics [Spring 2016]

Intro to MAS

Course Project Discussion

Mon, 31 Aug 2026 00:00:00 +0000

Two project discussions will be held during the semester. Each discussion will take place during the last 30 minutes of class. Each group will be assigned two specific time slots, one for each discussion.

Discussion 1: Project Planning

The first discussion will focus on:

Project background and objectives
Environment setup and required tools
Initial approach and project plan
Task allocation and timeline
Potential difficulties

Before the discussion, each group should review its selected topic, conduct some preliminary research, and prepare specific questions for the instructor.

Discussion 2: Midterm Project Review

The second discussion will focus on:

Current progress
Preliminary or intermediate results
Technical challenges
Feedback and improvements
Plans for the remaining work

Each group should prepare a brief progress update and some intermediate results, such as a prototype, screenshots, figures, tables, or experimental results. Groups may also prepare a few simple slides to present their progress and intermediate results.

Expectations

All group members should attend and participate.
Keep your introduction brief and leave enough time for discussion.
Prepare specific questions to make effective use of your assigned time.
Record the feedback and update your project plan after each discussion.

Bonus

Thu, 20 Aug 2026 00:00:00 +0000

实验：积木塔物理特征排序实验

任务简介

本实验研究人们如何利用物理特征来进行物理推理。每一回合，你会看到两个描述积木塔或其中积木的特征，你的任务是选择哪个特征对于预测每块积木在每个时刻的精确物理状态更有用。物理状态包括积木块的位置、旋转角度，以及对应的速度、角速度等。实验总时长约为40-60分钟。

被试要求

视力要求：视力或矫正（佩戴眼镜）视力正常，非色盲色弱；
其他要求：无精神或神经系统疾病史。

报名方式

填写问卷，并添加问卷中显示的主试微信。报名人数不限，多多益善。

问卷填写截止时间：9月30日
实验完成截止时间：10月11日

Bonus

完成实验的学生+1分

Lab 3: Block Tower Stability Predictor

Wed, 17 Sep 2025 00:00:00 +0000

Background

Predicting structural stability—such as whether a block tower will collapse—is not only an intellectual exercise but also a critical challenge in architecture, robotics, and disaster prevention. Humans, guided by visual intuition and experience, can swiftly judge stability and even anticipate collapse dynamics, a skill rooted in implicit physical knowledge and causal reasoning.

Current AI models struggle with this task: physics engine simulations are accurate but computationally costly and narrow in scope, while deep learning excels at visual recognition but lacks physical inference. Bridging perception, physics, and reasoning to achieve “intuitive physical perception” is thus essential for progress toward general intelligence. The ShapeStacks dataset advances this goal by providing synthetic block towers with stability annotations, enabling the training of models that approximate human-like intuitive reasoning.

Learning Objective

By completing this assignment, you will:

Understand the concept of visual stability prediction in block towers
Implement an InceptionV4-based model for binary classification (stable vs. unstable)
Train and evaluate the model using the ShapeStacks dataset

Environment Setup

Prerequisites

Python 3.6
Linux environment (recommended)
NVIDIA GPU with CUDA support

Installation

In order to run the intuitive physics models efficiently on GPU, the NVIDIA drivers, CUDA and cuDNN frameworks which are compatible with Tensorflow, and all other required packages should be installed.

conda create -n stability_predictor python=3.6
conda install cudatoolkit=10.0
conda install cudnn==7.6.5
pip install tensorflow-gpu==1.15.0

Assignment Tasks

Task 1: Data Provider

Objective:
The goal of this task is to build a robust data loading pipeline for the ShapeStacks dataset in shapestacks_provider.py.

Description: Complete the data loading pipeline in shapestacks_provider.py:

Implement _get_filenames_with_labels() to assign stability labels
Complete _create_dataset() to create TensorFlow datasets
Implement _parse_record() to read and preprocess images
Finish the main shapestacks_input_fn() function

You can download data here (pickup code: pq3L). The data structure is like:

${DATASET}/
|__ meta/
 |__ blacklist_stable.txt
 |__ blacklist_unstable.txt
|__ mjcf/
 |__ meshes/
 |__ textures/
 |__ assets.xml
 |__ env_blocks-easy-h=2-vcom=0-vpsf=0-v=1.xml
 |__ ...
 |__ env_ccs-hard-h=6-vcom=5-vpsf=0-v=120.xml
 |__ world_blocks-easy-h=2-vcom=0-vpsf=0-v=1.xml
 |__ ...
 |__ world_ccs-hard-h=6-vcom=5-vpsf=0-v=120.xml
|__ recordings/
 |__ env_blocks-easy-h=2-vcom=0-vpsf=0-v=1/
 |__ ...
 |__ env_ccs-hard-h=6-vcom=5-vpsf=0-v=120/
|__ splits/
 |__ blocks_all/
 |__ ...
 |__ ccs_all/
 |__ eval.txt
 |__ test.txt
 |__ train.txt
 |__ eval_bgr_mean.npy
 |__ test_bgr_mean.npy
 |__ train_bgr_mean.npy
 |__ default/
 |__ ...

Task 2: Model Implementation

Objective:
The aim of this task is to develop and complete the model architecture in inception_model.py, enabling both multi-class and binary classification for block tower stability prediction.

Description: Complete the model definition in inception_model.py:

Implement inception_v4_model_fn() for multi-class classification
Implement inception_v4_logregr_model_fn() for binary classification
Add proper loss functions and evaluation metrics

Task 3: Training Loop

Objective:
The purpose of this task is to implement an effective training workflow in train_inception_v4_shapestacks.py.

Description: The script train_inception_v4_shapestacks.py can be used to train a visual stability predictor on the ShapeStacks dataset. The main parameters are:

--data_dir which needs to point to the dataset location SHAPESTACKS_DATASET
--model_dir which defines a MODEL_DIR where all the tensorflow output and snapshots will be stored during training
--split_name which defines which split of data to use: ccs_all / blocks_all

Set your lab root path to SHAPESTACKS_CODE_HOME before training and evaluation:

export SHAPESTACKS_CODE_HOME=./

An example run of the training script looks like this:

(stability_predictor) $ python intuitive_physics/stability_predictor/train_inception_v4_shapestacks.py --data_dir lab3_dataset/shapestacks --model_dir ./output --split_name ccs_all

You can track the training progress by pointing a tensorboard to the model’s root directory:

(stability_predictor) $ tensorboard --logdir=stability_predictor:${MODEL_DIR}

The most recent model checkpoints during training are kept in the models’s root directory. If the training script finds existing checkpoints in MODEL_DIR, it will automatically load the most recent one of them and resume training from there.

During training, the checkpoints which perform best on the validation set are also saved to the snapshots/ subdirectory. The amount of best checkpoints to keep can be set via --n_best_eval.

Complete the training script in train_inception_v4_shapestacks.py:

Set up the Estimator with proper configuration
Implement the training loop with evaluation

Task 4: Testing and Evaluation

Objective:
The goal of this task is to implement a comprehensive testing and evaluation workflow in test_inception_v4_shapestacks.py.

Description: After a stability predictor has been trained, the latest checkpoint or a particular snapshot can be loaded back into a tf.estimator.Estimator You can also set the model_dir parameter of tf.estimator.Estimator to MODEL_DIR/snapshots/<snapshot_name> to load the weights of a particular snapshot.

Complete the testing script in test_inception_v4_shapestacks.py:

Load and evaluate trained models
Generate result files for different splits

Evaluation and Grading

Your implementation will be evaluated based on the accuracy achieved on the test sets. After completing your implementation, you should run the script test_inception_v4_shapestacks.py twice, specifying the --split_name as ccs_all and blocks_all, respectively.

This will generate two result files:

results_ccs_all.py

results_blocks_all.py

Each result will be graded separately, with 50 points assigned to each split:

ccs_all dataset accuracy (50 points)
blocks_all dataset accuracy (50 points)

The final score will be the sum of the two.

Implementation Guidelines

Only modify code within the TODO sections
Do not change the filenames of any files required for submission; submit them exactly as specified.
You may need to adjust environment setup steps (such as Python version, CUDA, or cuDNN) according to your own system configuration.

Hand-in Requirements

Submit the following files as a .zip archive to GradeScore:

data_provider/shapestacks_provider.py
intuitive_physics/stability_predictor/train_inception_v4_shapestacks.py
intuitive_physics/stability_predictor/test_inception_v4_shapestacks.py
tf_models/inception/inception_model.py
results_ccs_all.py
results_blocks_all.py

Academic Integrity

This assignment must be completed individually. You may talk about general ideas and concepts with your classmates, but all code you submit must be written by you alone. Any form of plagiarism will result in failing the course.

Permitted:

Discussing theoretical concepts
Exchanging debugging approaches (without sharing code)
Referring to course materials and recommended resources

Prohibited:

Sharing or duplicating code solutions
Using external code implementations without explicit permission
Working together on the actual coding tasks

Good luck with your assignment!

Intuitive Physics

Wed, 20 Aug 2025 00:00:00 +0000

Reading list

survey/review/perspective paper book chapter

Required

Intuitive Physics: Current Research and Controversies, TiCS 2019
Building machines that learn and think like people (Section 4.1.1), BBS 2016
Dark, Beyond Deep: A Paradigm Shift to Cognitive AI with Humanlike Common Sense (Section 4), Engineering 2020
Intuitive physics as probabilistic inference
Simulation as an engine of physical scene understanding, PNAS 2013

Optional - Perceiving Physics

Visual Perception of Intuitive Physics and Its Computational Foundations, Annual Review of Vision Science 2026
Infants Physical Reasoning and the Cognitive Architecture that Supports It, The Cambridge Handbook of Cognitive Development (2022)

Optional - Intuitive Physics

Rapid trial-and-error learning with simulation supports flexible tool use and physical reasoning, PNAS 2020
Core knowledge, Developmental Science 2007
Physical symbol systems, Cognitive Science 1980
Probabilistic Simulation Predicts Human Performance on Viscous Fluid-Pouring Problem, CogSci 2016
Mind Games: Game Engines as an Architecture for Intuitive Physics, TiCS 2017
The fine structure of surprise in intuitive physics: when, why, and how much?, CogSci 2020
Intuitive physics learning in a deep-learning model inspired by developmental psychology, Nature Human Behavior 2022
On the Learning Mechanisms in Physical Reasoning, NeurIPS 2022
Sources of uncertainty in intuitive physics, TiCS 2013
Different physical intuitions exist between tasks, not domains, Computational Brain & Behavior 2018
Partial mental simulation explains fallacies in physical reasoning, Cognitive Neuropsychology 2022
“Just in Time” representations for mental simulation in intuitive physics, CogSci 2023
A Survey on Machine Learning Approaches for Modelling Intuitive Physics, arXiv 2202.06481

Optional - Controversies of Intuitive Physics

How Robust Are Probabilistic Models of Higher-Level Cognition?, Psychological Science 2013
Relevant and Robust: A Response to Marcus and Davis (2013), Psychological Science 2013
Limits on simulation approaches in intuitive physics, Cognitive Psychology 2021

Optional - Intuitive Physics for Computer Vision

Beyond point clouds: Scene understanding by reasoning geometry and physics, CVPR 2013
Scene Understanding by Reasoning Stability and Safety, IJCV 2015
Galileo: Perceiving Physical Object Properties by Integrating a Physics Engine with Deep Learning, NeurIPS 2015
Learning to See Physics via Visual De-animation, NeurIPS 2017
Physion: Evaluating Physical Prediction from Vision in Humans and Machines, NeurIPS 2021

Optional - Containment

Evaluating Human Cognition of Containing Relations with Physical Simulation, CogSci 2015
Tracking Occluded Objects and Recovering Incomplete Trajectories by Reasoning about Containment Relations and Human Actions, AAAI 2018
Commonsense reasoning about containers using radically incomplete information, Artificial Intelligence 2017

Into Research

Sat, 10 Dec 2022 00:00:00 +0000

What is Research?

History of Science

科学的历程

Scientific Writing

Style: Lessons in Clarity and Grace

Interested in CoRe? TL;DR

Examples of good presentations (YouTube)

Lab 4: Peter-Clark (PC) Algorithm

Mon, 15 Sep 2025 00:00:00 +0000

Background

In many scientific fields, moving beyond simple correlation to understand cause-and-effect relationships is a primary goal. Causal discovery is a discipline that aims to infer causal structures directly from observational data. Instead of just knowing that X and Y are related, we want to determine if X causes Y, Y causes X, or if a hidden common cause influences both.

The PC algorithm, named after its creators Peter Spirtes and Clark Glymour, is a foundational constraint-based method for causal discovery. It operates on the principle of conditional independence. The core idea is that if two variables X and Y are causally related (e.g., X -> Y), they will be statistically dependent. However, this dependence might vanish when we condition on other variables. The PC algorithm systematically uses statistical tests for conditional independence to prune a graph, revealing the underlying causal skeleton and orienting as many edges as possible.

This assignment will guide you through implementing the three main stages of the PC algorithm:

Skeleton Discovery: Identifying which variables are connected, regardless of direction.
V-Structure Identification: Finding and orienting “colliders” (X -> Z <- Y), which form the initial set of directed edges.
Edge Orientation: Propagating the directional information from V-structures throughout the graph using a set of logical rules.

Learning Objectives

By completing this assignment, you will:

Understand the principles of constraint-based causal discovery.
Implement the key stages of the PC algorithm: skeleton discovery, V-structure identification, and edge orientation.
Learn how to apply conditional independence tests to learn graphical structures from data.
Gain practical experience in translating an algorithm from theory to code.
Evaluate the performance of a causal discovery algorithm against a ground truth.

Environment Setup

Language: Python 3.9+
Required packages:

conda create -n causality python=3.10

conda activate causality

pip install numpy networkx pandas scipy scikit-learn statsmodels pydot matplotlib graphviz causal-learn

Provided Files:
- framework.py – Contains the scaffold for the PC algorithm. You will fill in the ## TODO sections.
- data_linear_10.txt – The observational dataset your algorithm will run on.

Note: The causal graph is represented by an adjacency matrix where:

graph[i, j] = graph[j, i] = -1 represents an undirected edge i -- j.
graph[i, j] = 1 and graph[j, i] = 0 represents a directed edge i -> j.
graph[i, j] = 0 and graph[j, i] = 0 represents no edge between i and j.

Assignment Tasks

Task 1: Skeleton Discovery (40 pts)

Objective: Implement the first phase of the PC algorithm to find the undirected skeleton of the causal graph.

Description: You will complete the skeleton_discovery() function in framework.py. This function starts with a fully connected undirected graph. Your task is to iteratively remove edges between pairs of nodes (i, j) that are found to be conditionally independent. You will test for independence conditioning on sets of neighbors of increasing size (d = 0, 1, 2, ...).

For each pair of adjacent nodes (i, j), and for each possible conditioning set S of size d drawn from the neighbors of i (or j), you will perform a conditional independence test using the provided fisherz function.
If the test’s p-value is greater than the significance level alpha, the nodes i and j are conditionally independent. The edge between them should be removed.
You must also record the separating set S that made i and j independent in the sepsets dictionary. This dictionary is crucial for the next task.
The process stops when no more edges can be removed.

Task 2: V-Structure Identification (30 pts)

Objective: Implement the second phase of the PC algorithm to identify and orient V-structures (colliders).

Description: You will complete the identify_v_structures() function in framework.py. A V-structure is a pattern of the form X -> Z <- Y, where X and Y are not themselves adjacent. This structure is identifiable because X and Y are independent, but become dependent when conditioned on Z.

Your implementation should iterate through all triples of nodes (i, k, j) that form a “v” shape: i is adjacent to k, j is adjacent to k, but i and j are not adjacent.
For each such triple, you will check the sepsets dictionary from Task 1. If the middle node k is not in the separating set for (i, j), then the structure i-k-j must be a V-structure.
You should then orient the edges as i -> k <- j in the adjacency matrix.

Task 3: Edge Orientation (30 pts)

Objective: Apply a set of logical orientation rules (Meek’s rules) to orient as many of the remaining undirected edges as possible.

Description: You will complete the orient_edges() function in framework.py. After identifying V-structures, some directed edges are known. These directions can be propagated through the graph to orient other edges, based on two main principles: avoiding the creation of new V-structures and avoiding the creation of cycles.

You will implement a loop that repeatedly applies the orientation rules to the graph until no more edges can be oriented.
Rule 1 (Avoid New V-Structures): If you have a structure X -> Y - Z and X and Z are not adjacent, you must orient the edge as Y -> Z. Orienting it as Y <- Z would create a new V-structure (X -> Y <- Z), which would have been detected in Task 2.
Rule 2 (Avoid Cycles): If you have a chain X -> Y -> Z and also an undirected edge between X and Z, you must orient it as X -> Z. Orienting it as X <- Z would create a cycle (X -> Y -> Z <- X).
Rule 3: Orient x - y into x -> y if there are two paths x-k->y and x-l->y, with k and l non-adjacent.
The hints in the framework.py file provide a starting point for implementing these rules.

Task 4: Analysis & Evaluation (Ungraded)

Objective: Run your completed PC algorithm on the provided data and analyze how the choice of the significance level alpha affects the output.

Description: The analyze_causal_graph function is already written for you. Once your implementation is complete, you can run framework.py directly. The script will execute your PC algorithm with three different alpha values (0.01, 0.05, 0.2). For each value, it will:

Calculate the Structural Hamming Distance (SHD) between your algorithm’s output and the true graph. SHD is a score that counts the number of edge additions, deletions, or reversals needed to match the true graph (lower is better).
Visualize the learned causal graph.

Observe the output and think about the following questions:

How does a stricter alpha (e.g., 0.01) affect the number of edges in the learned skeleton?
Which alpha value gives the best SHD score?
What does this tell you about the trade-off between finding true causal links and falsely identifying links that aren’t there?

Implementation Guidelines

Only modify code within the ## TODO sections in framework.py.
Do not change the function signatures or import statements.
The helper functions for conditional independence testing, graph plotting, and SHD calculation are already provided for you.

Hand-in Requirements

Submission Format

Submit only one file: framework.py.

File Requirements

Your framework.py must contain all implemented functions.
Do not modify the original function signatures or provided helper functions.

Causality

Wed, 20 Aug 2025 00:00:00 +0000

Reading list

survey/review/perspective paper book Encyclopedia

Required

Dark, Beyond Deep: A Paradigm Shift to Cognitive AI with Humanlike Common Sense (Section 3), Engineering 2020
Estimating causal effects of treatments in randomized and nonrandomized studies, Journal of Educational Psychology 1974
Theory-based Causal Induction, PsychReview 2009

Optional - Causal Learning

Causal Models
A Theory of Causal Learning in Children: Causal Maps and Bayes Nets, PsychReview 2004
Where science starts: Spontaneous experiments in preschoolers’ exploratory play, Cognition 2011

Optional - Causal Perception

Perceptual causality and animacy, TiCS 2001
Causation From Perception, Perspectives on Psychological Science 2011

Optional - Causal Judgment

The Causal Asymmetry, PsychReview 2006
Bayesian Generic Priors for Causal Learning, PsychReview 2008
A counterfactual simulation model of causal judgments for physical events, PsychReview 2021

Optional - Perspective

Causality in Thought, Annual Review of Psychology 2015
What is wrong with Bayes net?, The monist (2021)
The Seven Tools of Causal Inference, with Reflections on Machine Learning, Communications of the ACM 2019
Causality: Models, Reasoning, and Inference, Cambridge University Press (2009)

Optional - Recent Advances

Cladder: Assessing causal reasoning in language models, NeurIPS 2023
Causality for large language models, arXiv 2024
Robust agents learn causal world models, ICLR 2024

Handouts

Mon, 14 Sep 2026 00:00:00 +0000

Slides

Google Drive

Project Topics

Topic List (PDF)

Gradescope Assignments

Google Drive

Lab 5: Behavior Cloning for Robotic Manipulation

Tue, 09 Sep 2025 00:00:00 +0000

Background

This assignment focuses on Behavior Cloning (BC), a classical imitation learning technique where an agent learns to mimic expert demonstrations via supervised learning. BC is widely used in robotics, manipulation, and reinforcement learning as a simple yet powerful baseline.

You will work with the ManiSkill platform and PyTorch to complete a full imitation learning pipeline, including dataset handling, network design, training, and evaluation.

Learning Objectives

By completing this assignment, you will:

Understand the principles and workflow of Behavior Cloning.
Gain hands-on experience with data loading, training, and evaluation in imitation learning.
Implement and design a neural network policy using PyTorch.
Train and evaluate a behavior cloning policy on manipulation tasks.

Environment Setup

Install Dependencies

Recommended: Linux, Conda, CUDA GPU (optional)

conda create -n bc python=3.10
conda activate bc
pip install -r requirements.txt

Download Datasets and Generate Demonstration Trajectories

Download the demonstration data (default to ~/.maniskill/demos/):

python -m mani_skill.utils.download_demo "PushCube-v1"

We need to first generate data for IL training, via replaying the demonstrations and saving the trajectories based on end-effector pose:

python -m mani_skill.trajectory.replay_trajectory \
 --traj-path ~/.maniskill/demos/PushCube-v1/motionplanning/trajectory.h5 \
 --use-first-env-state -c pd_ee_delta_pos -o state \
 --save-traj --num-envs 1 -b physx_cpu --vis

This should open a window, show, and save the demonstrations. Let it play for a while to generate all 1000 trajectories. You can also use --count to specify how many trajectories you want to save (e.g., --count 500).

The data will be dumped to the same directory of the trajectory.h5 file, e.g., ~/.maniskill/demos/PushCube-v1/motionplanning/trajectory.state.pd_ee_delta_pos.physx_cpu.h5.

Complete the Assignment & Run Tests

Fill in all required functions in bc_utils.py (# TODO sections only). Do not change function signatures or import statements.

Subtask 1: Implement Behavior Cloning Loss Function (30 pts)

Implement the MSE loss function for BC. Given the ground truth actions $\mathbf{a}_i$ and the predictions $\hat{\mathbf{a}}_i$, for samples $i = 1, 2, \dots, n$, the loss function is:

$$ \mathcal{L} = \frac{1}{n} \sum_{i=1}^n \Vert\hat{\mathbf{a}}_i - \mathbf{a}_i \Vert^2 $$

Complete the compute_bc_loss(...) function according to the comments after it.

Subtask 2: Implement Observation Normalization (30 pts)

Observation normalization is very helpful for efficiently learning the policy. While we cannot know the true mean and standard deviation of the observation space, we can estimate them using the statistics of the given batch (assuming the batch is large enough).

Given a batch of observations (each as a vector), $\mathbf{o}_i, i = 1, 2, \dots, B$, with $B$ being the batch size, we estimate the mean and standard deviation for each dimension $j$:

$$ \tilde{\mu}_j = \frac{1}{B} \sum_{i=1}^n \mathbf{o}_{ij}, \quad \tilde{\sigma}_j = \sqrt{\frac{1}{B} \sum_{i=1}^B (\mathbf{o}_{ij} - \tilde{\mu}_j)^2} $$

Then, normalization is computed for each dimension $j$:

$$ \mathbf{o}_{ij}^{\text{norm}} = \frac{\mathbf{o}_{ij}- \tilde{\mu}_j}{\tilde{\sigma}_j} $$

Implement the normalize_observations(...) function according to the comments after it. Note that you should actively avoid division by zero appropriately.

Subtask 3: Implement MLP Policy (20 pts)

Implement the function to create an MLP (multi-layer perceptron) as a subclass of torch.nn.Module, according to the dimensions of the state and action spaces. The network should be properly designed in size and depth to ensure sufficient expressiveness without overfitting.

Implement the create_actor_network(...) function according to the comments after it.

Subtask 4: Train the Policy (20 pts)

After filling in the required functions, you should be able to train your BC policy on the provided tasks:

# PushCube task
python run.py --env-id "PushCube-v1" \
 --demo-path ~/.maniskill/demos/PushCube-v1/motionplanning/trajectory.state.pd_ee_delta_pos.physx_cpu.h5 \
 --control-mode "pd_ee_delta_pos" --sim-backend "cpu" \
 --max-episode-steps 100 --total-iters 10000 --batch-size 64

The demo-path specified here should match the generated data above. During training, the results (including checkpoints and videos) will be saved under the runs/ directory. In submission, copy the checkpoint as results/checkpoint.pt (under the homework directory), and the auto grader will check it against your implemented MLP network. Also, pick two videos that show successful manipulation (pushing the cube into the target area), and name them to results/success_1.mp4, results/success_2.mp4.

Note that its evaluation results (manipulation success rate) will not affect your score on this lab. However, you should make sure that the successful results can be cherry-picked from the training results. If there are none, consider checking your implementations, running the training code for more iterations or re-running it with a different seed.

Submission Checklist

Please submit the following files in a single zip file:

Your completed Python code, with all required functions implemented and necessary comments.
Your trained model checkpoint and successful manipulation videos under results/.

Autograder Testing

Your submission will be tested against multiple test cases
Partial credit will be awarded for partially correct implementations
Timeout limit: 10 minutes in total

Mirroring, Imitation, and Tool

Wed, 20 Aug 2025 00:00:00 +0000

Reading list

survey/review/perspective paper book

Required - Mirroring and Imitation

The Mirror-Neuron System, Annual Review of Neuroscience 2004
The anthropomorphic brain: The mirror neuron system responds to human and robotic actions, NeuroImage 2007
A survey of robot learning from demonstration, RAS 2009

Required - Tool

The neural bases of complex tool use in humans, TiSC 2004
The cognitive bases of human tool use, BBS 2012
Rapid trial-and-error learning with simulation supports flexible tool use and physical reasoning, PNAS 2020

Optional - Mirroring and Imitation: Observational Studies

Imitative learning of artificial fruit processing in children (Homo sapiens) and chimpanzees (Pan troglodytes), Jounal of Comparative Psychology 1996
Rational imitation in preverbal infants, Science 2002
Imitation in Animals and Artifacts, MIT Press (2002)
Causal knowledge and imitation/emulation switching in chimpanzees (Pan troglodytes) and children (Homo sapiens), Animal Cognition 2005
Extending the mirror neuron system model, I: Audible actions and invisible grasps, Biological Cybernetics 2006
Copying actions and copying outcomes: Social learning through the second year, Developmental Psychology 2006
The hidden structure of overimitation, PNAS 2007
Imitation of causally opaque versus causally transparent tool use by 3-and 5-year-old children, Animal Cognition 2007
EEG evidence for mirror neuron activity during the observation of human and robot actions: Toward an analysis of the human qualities of interactive robots, Neurocomputing 2007
Emulation, imitation, over-imitation and the scope of culture for child and chimpanzee, Philos. Trans. R. Soc. B 2009
When pliers become fingers in the monkey motor system, PNAS 2018

Optional - Mirroring and Imitation: Computational Models

Policy Search for Motor Primitives in Robotics, NeurIPS 2008
Robot learning from demonstration by constructing skill trees, IJRR 2011
Theories and computational models of affordance and mirror systems: An integrative review, Neuroscience & Biobehavioral Reviews 2013
Acquiring and Generalizing the Embodiment Mapping From Human Observations to Robot Skills, RA-L 2016
Learning Social Affordance Grammar from Videos: Transferring Human Interactions to Human-Robot Interactions, ICRA 2017
Feeling the Force: Integrating Force and Pose for Fluent Discovery through Imitation Learning to Open Medicine Bottles, IROS 2017
An Algorithmic Perspective on Imitation Learning, Foundations and Trends in Robotics 2018
Mirroring without Overimitation: Learning Functionally Equivalent Manipulation Actions, AAAI 2019

Optional - Tool: Observational Studies

Observational learning of tool-use by young chimpanzees, Human Evolution 1987
Shaping of hooks in new caledonian crows, Science 2002
Neural correlates of Early Stone Age toolmaking: technology, language and cognition in human evolution, Philos. Trans. R. Soc. B 2008
Cognitive insights from tool use in nonhuman animals, APA Handbook of Comparative Psychology: Volume 2: Perception, Learning, and Cognition (2017)
Stone toolmaking and the evolution of human culture and cognition, Philos. Trans. R. Soc. B 2011
Tool Use and Causal Cognition, Oxford University Press (2011)
Experimental evidence for the co-evolution of hominin tool-making teaching and language, Nature Communication 2015
Is there a link between the crafting of tools and the evolution of cognition?, Cognitive Science 2014

Optional - Tool: Computational Models

Understanding Tools: Task-Oriented Object Modeling, Learning, and Recognition, CVPR 2015
Affordance Detection of Tool Parts from Geometric Features, ICRA 2015
A Model-Based Approach to Finding Substitute Tools in 3D Vision Data, ICRA 2016
Self-supervised learning of tool affordances from 3D tool representation through parallel SOM mapping, ICRA 2017
Differentiable Physics and Stable Modes for Tool-Use and Manipulation Planning, RSS 2018
Improvisation through Physical Understanding: Using Novel Objects as Tools with Visual Foresight, RSS 2019
Force-and-Motion Constrained Planning for Tool Use, IROS 2019
Autonomous Tool Construction Using Part Shape and Attachment Prediction, RSS 2019
Tool Macgyvering: Tool Construction Using Geometric Reasoning, ICRA 2019
KETO: Learning Keypoint Representations for Tool Manipulation, ICRA 2020
How to Select and Use Tools?: Active Perception of Target Objects Using Multimodal Deep Learning, RA-L 2021
Understanding Physical Effects for Effective Tool-use, RA-L 2022
Robot tool use: A survey, Frontiers in Robotics and AI 2023
Learning to Design and Use Tools for Robotic Manipulation, CoRL 2023
RobotSmith: Generative Robotic Tool Design for Acquisition of Complex Manipulation Skills, NeurIPS 2026

Lab 6: Sampling-based Path Planning

Tue, 09 Sep 2025 00:00:00 +0000

Background

In robotics, a central challenge is motion planning—finding a collision-free path from a start position to a goal while avoiding obstacles. Sampling-based path planning tackles this problem by randomly sampling points in free space and incrementally building a tree or graph. This approach avoids explicitly computing the entire configuration space and scales more effectively to higher dimensions.

A widely used method in this family is the Rapidly-Exploring Random Tree (RRT), which grows a tree by steering from the nearest node toward random samples. Over time, many variants of RRT have been developed to improve efficiency and solution quality.

In this assignment, you will implement RRT and its bidirectional variant RRT-Connect in a 2D environment with obstacles, and learn how to generate collision-free paths using these algorithms.

Learning Objectives

By completing this assignment, you will:

Understand the fundamentals of sampling-based path planning.
Implement the RRT algorithm and extend it to RRT-Connect.
Compare single-tree and bidirectional tree planning approaches.

Environment Setup

Language: Python 3.9+
Required packages:
```
pip install numpy matplotlib
```
rrt.py – contains the scaffold for RRT.
rrt_connect.py – contains the scaffold for RRT-Connect.
replay.py – visualization utilities for replaying the planning process.
framework.py - include to compenstate code

Note: Obstacles are represented as circles (cx, cy, r) in a bounded 2D map (default size: [0,10] × [0,10]).

Assignment Tasks

Task 1: Collision Detection (40 pts)

Objective: Implement a function to determine whether a sampled node collides with any obstacle.

Description: You will complete the is_collision() function in framework.py. Given a 2D position and a list of circular obstacles (x, y, radius), your function should compute Euclidean distances, inflate each radius by a small tolerance, and return a boolean indicating if the position lies inside or on any inflated obstacle boundary.

Task 2: Extend Tree Node (40 pts)

Objective: Implement a function to extend from a current node toward a target by at most a specified step size.

Description: You will complete the cal_next_node_value() function in framework.py. Given a current node, a target/sample, and max_step, your function should compute the direction from the current node to the target and return the next node along that line. If the target is closer than max_step, extend directly to it; otherwise, move exactly max_step toward it.

Task 3: Bi-Tree Growth (20 pts)

Objective: Implement a helper to alternate growth between two trees for bidirectional planning.

Description: You will complete the swap_trees() function in framework.py. Each time it is called, the function should toggle a boolean swap_flag and return the two trees in reversed order. This enables the planner to extend alternately from the start tree and the goal tree across iterations.

Implementation Guidelines

Code Structure

Only modify code within ## TODO sections in framework.py
Do not change function signatures or import statements
Maintain consistent tensor shapes and data types

Debugging Tips

Use breakpoint() to check intermediate results
Visualize intermediate results when possible

Hand-in Requirements

Submission Format

Submit only one file: framework.py
Platform: Gradescope autograder

File Requirements

Your framework.py must contain all implemented functions
Do not modify function signatures or imports
Include necessary comments explaining your implementation approach

Autograder Testing

Your submission will be tested against multiple test cases
Partial credit will be awarded for partially correct implementations
Timeout limit: 10 minutes in total

Additional Resources

Documentation

Communication and Language

Wed, 20 Aug 2025 00:00:00 +0000

Reading list

survey/review/perspective paper

Required

Origins of Human Communication (Chapter 1), MIT Press (2010)
Iconicity: From sign to system in human communication and language, Pragmatics & Cognition 2015
Multi-Agent Cooperation and the Emergence of (Natural) Language, ICLR 2017
Predicting Pragmatic Reasoning in Language Games, Science 2012
Grounded language acquisition through the eyes and ears of a single child, Science 2024

Optional - Communication

From communication to language-a psychological perspective, Cognition 1974
Twelve-month-olds communicate helpfully and appropriately for knowledgeable and ignorant partners, Cognition 2008
Communication about absent entities in great apes and human infants, Cognition 2015
The interaction engine: cuteness selection and the evolution of the interactional base for language, Philos. Trans. R. Soc. B 2022
Animal Communication in Linguistic and Cognitive Perspective, Annual Review of Linguistics 2023
Social Learning in Animals: The Roots of Culture, Academic Press (1996)

Optional - Language and Thought

The best game in town: The reemergence of the language-of-thought hypothesis across the cognitive sciences, BBS 2022
The myth of language universals: Language diversity and its importance for cognitive science, BBS 2009
From Word Models to World Models: Translating from Natural Language to the Probabilistic Language of Thought, arXiv 2023
Dissociating language and thought in large language models, Trends in Cognitive Sciences 2024
Language is primarily a tool for communication rather than thought, Nature 2024
Can language restructure cognition? The case for space, Trends in Cognitive Science 2004
Russian blues reveal effects of language on color discrimination, PNAS 2007
Does language guide event perception? Evidence from eye movements, Cognition 2008
Combined evidence from artificial neural networks and human brain-lesion models reveals that language modulates vision in human perception, Nature Human Behaviour 2026

Optional - Pragmatics

Word Meaning in Minds and Machines, PsychReview 2021
Meaning in Language: An Introduction to Semantics and Pragmatics, Oxford Handbook in Linguistics (2010)
The Rational Speech Act Framework, Annual Review of Linguistics 2023
Pragmatics and Discourse, Speech, language, and communication (1995)
Pragmatic Language Interpretation as Probabilistic Inference, TiCS 2016
Colors in Context: A Pragmatic Neural Model for Grounded Language Understanding, TACL 2017

Optional - Implicature, Metaphpor, and Humor

Implicature
Formalizing the Pragmatics of Metaphor Understanding, CogSci 2014
The Funny Thing About Incongruity: A Computational Model of Humor in Puns, CogSci 2013
Knowledge and Implicature: Modeling Language Understanding as Social Cognition, TopiCS 2013

Optional - Gaze and Gesture

The eyes have it: the neuroethology, function and evolution of social gaze, Neuroscience and Biobehavioral Review 2000
Gaze in conversation, The Handbook of Conversation Analysis, 2012
Social Eye Gaze in Human-robot Interaction: A Review, JHRI 2017
The origin of pointing: Evidence for the touch hypothesis, Science Advances 2019
Perspective determines the production and interpretation of pointing gestures, Psychonomic Bulletin & Review 2021
A New Look at Infant Pointing, Child Development 2007
Prelinguistic Infants, but Not Chimpanzees, Communicate About Absent Entities, Psychological Science 2009

Optional - Models for Nonverbal Communication

Pragmatic Inference and Visual Abstraction Enable Contextual Flexibility During Visual Communication, Computational Brain & Behavior 2020
Visual resemblance and interaction history jointly constrain pictorial meaning, Nature Communications 2023
Individual vs. Joint Perception: a Pragmatic Model of Pointing as Communicative Smithian Helping, CogSci 2021
What Is the Point? A Theory of Mind Model of Relevance, CogSci 2022
Learning Triadic Belief Dynamics in Nonverbal Communication from Videos, CVPR 2021
YouRefIt: Embodied Reference Understanding with Language and Gesture, ICCV 2021
Emergent Graphical Conventions in a Visual Communication Game, NeurIPS 2022

Optional - Evolution

Culture: Copying, Compression, and Conventionality, Cognitive Science 2015
Why Are There Different Languages? The Role of Adaptation in Linguistic Diversity, Trends in Cognitive Science 2016
Typology in the 21st century: Major current developments, Linguistic Typology 2007
Gesture - Visible Action as Utterance, Cambridge University Press 2015
Primate Vocalization, Gesture, and the Evolution of Human Language, Current Anthropology 2008
How Language Began: Gesture and Speech in Human Evolution, Cambridge University Press 2015
From mouth to hand: Gesture, speech, and the evolution of right-handedness, PloS One 2003
Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language, PNAS 2008
Human sound systems are shaped by post-Neolithic changes in bite configuration, Science 2019

Optional - Emergent Languages

Iterated Learning: A Framework for the Emergence of Language, Artificial Life 2003
Emergent Multi-Agent Communication in the Deep Learning Era, arXiv preprint arXiv:2006.02419
Emergence of Grounded Compositional Language in Multi-Agent Populations, AAAI 2018
Emergent Graphical Conventions in a Visual Communication Game, NeurIPS 2022
A framework for the emergence and analysis of language in social learning agents, Nature Communications 2024

Optional - Pantomime

How Pantomime Works: Implications for Theories of Language Origin, Frontiers in Communication 2019
Pantomime and imitation in great apes, Interaction Studies 2018
In support of the role of pantomime in language evolution, Journal of Language Evolution 2018
Pantomime as a foundation for ritual and language, Studia Liturgica 2018
Defining Pantomime for Language Evolution Research, Topoi 2018
Demonstration and pantomime in the evolution of teaching and communication, Language & Communication 2021

Optional - Pedagogy

Natural pedagogy, TiCS 2009
Natural pedagogy as evolutionary adaptation, Philos. Trans. R. Soc. B 2011
Preschoolers use pedagogical cues to guide radical reorganization of category knowledge, Cognition 2014
The effects of information utility and teachers’ knowledge on evaluations of under-informative pedagogy across development, Cognition 2022

Intentionality

Wed, 20 Aug 2025 00:00:00 +0000

Reading list

survey/review/perspective paper book tutorial

Required

Does the Chimpanzee Have a Theory of Mind?, Behavioral and Brain Sciences 1978
Dark, Beyond Deep: A Paradigm Shift to Cognitive AI with Humanlike Common Sense (Section 6), Engineering 2020
Discerning intentions in dynamic human action, TiCS 2001
‘Obsessed with goals’: Functions and mechanisms of teleological interpretation of actions in humans, Acta Psychologica 2007
Action understanding as inverse planning, Cognition 2009
Plan Activity and Intent Recognition Tutorial, AAAI 2019

Optional - Observational Studies

First-person action experience reveals sensitivity to action efficiency in prereaching infants, PNAS 2013
Infants selectively encode the goal object of an actor’s reach, Cognition 1998
Goal-Directed Action Representation in Autism, Journal of Autism and Developmental Disorder 2006
Rational quantitative attribution of beliefs, desires and percepts in human mentalizing, Nature Human Behaviour 2009
Unwilling Versus Unable: Infants’ Understanding of Intentional Action, Developmental Psychology 2005
Twelve-month-olds’ comprehension and production of pointing, British Journal of Developmental Psychology 2012
Understanding Attention: 12- and 18-Month-Olds Know What Is New for Other Persons, Developmental Psychology 2003
Fourteen- through 18-month-old infants differentially imitate intentional and accidental actions, Infant Behavior and Development 1998
Do apes ape?, Social Learning in Animals: The Root of Culture (1996)
Taking the intentional stance at 12 months of age, Cognition 1995
Teleological reasoning in infancy: the naı̈ve theory of rational action, TiCS 2003

Optional - Goal Attribution

Inferring “Dark Matter” and “Dark Energy” from Videos, ICCV 2013
Help or Hinder: Bayesian Models of Social Goal Inference, NeurIPS 2009
A survey of inverse reinforcement learning: Challenges, methods and progress, Artificial Intelligence 2021
First-Person Activity Forecasting with Online Inverse Reinforcement Learning, ICCV 2017
Maximum Entropy Inverse Reinforcement Learning, AAAI 2018
A Survey of Preferenced-Based Reinforcement Learning Methods, JMLR 2017

Optional - Action Prediction

Predicting Human Activities Using Stochastic Grammar, ICCV 2017
In the Eye of Beholder: Joint Learning of Gaze and Actions in First Person Video, ECCV 2018
Human Activity Recognition and Prediction: A Survey, arXiv preprint arXiv:1806.11230
Social LSTM: Human Trajectory Prediction in Crowded Spaces, CVPR 2016
Social GAN: Socially Acceptable Trajectories with Generative Adversarial Networks, CVPR 2018
A Generalized Earley Parser for Human Activity Parsing and Prediction, PAMI 2020

Animacy and ToM

Wed, 20 Aug 2025 00:00:00 +0000

Reading list

survey/review/perspective paper Encyclopedia book chapter

Required - Animacy

Perceptual causality and animacy, TiCS 2001
An Experimental Study of Apparent Behavior, The American Journal of Psychology 1944
The psychophysics of chasing: A case study in the perception of animacy, Cognitive Psychology 2009
A unified psychological space for human perception of physical and social events, Cognitive Psychology 2021

Required - Theory of Mind (ToM)

Humans Have Evolved Specialized Skills of Social Cognition: The Cultural Intelligence Hypothesis, Science 2007
How Sophisticated Is Infants’ Theory of Mind?, The Cambridge Handbook of Cognitive Development (2022)
Machine Theory of Mind, ICML 2018
Action understanding as inverse planning, Cognition 2009
Modeling Other Minds: A Computational Account of Social Cognition and Its Development, Annual Review of Developmental Psychology 2025

Optional - Machine ToM

A Bayesian Theory of Mind Approach to Nonverbal Communication, HRI 2019
Cognitive Science as a Source of Forward and Inverse Models of Human Decisions for Robotics and Control, Annual Review of Control, Robotics, and Autonomous Systems 2021
An Embodied Cognition Approach to Mindreading Skills for Socially Intelligent Robots, IJRR 2009
Joint Inference of States, Robot Knowledge, and Human (False-)Beliefs, ICRA 2020
Implementing Theory of Mind on a Robot Using Dynamic Epistemic Logic, IJCAI 2020
The Hanabi challenge: A new frontier for AI research, Artificial Intelligence 2020
Recursive Mentalizing and Common Knowledge in the Bystander Effect, JEPG 2016
Communication in Action: Planning and Interpreting Communicative Demonstrations, JEPG 2021
Action Understanding as Inverse Planning, Cognition 2009
Machine Theory of Mind, ICML 2018

Optional - Psychological and Phylosophical Studies on Animacy

Animal Consciousness
Infants’ concept of animacy, Cognitive Development 1996
The Wolfpack Effect: Perception of Animacy Irresistibly Influences Interactive Behavior, Psychological Science 2010
Perception of animacy from the motion of a single object, Perception 2000
An Experimental Study of Apparent Behavior, American Journal of Psychology 1944
Perceiving Animacy and Intentionality: Visual Processing or Higher-Level Judgment?, Social Perception (MIT Press) 2013
The Perception of Causality, 1946/1963

Optional - Animacy Modeling

Optional - Philosophical Studies on ToM

The Computational Theory of Mind
Modularity of Mind
Folk Psychology as Mental Simulation
The Mind/Brain Identity Theory
Evolution in Mind: Evolutionary Dynamics, Cognitive Processes, and Bayesian Inference, TiCS 2007
Does the Chimpanzee Have a Theory of Mind?, Behavioral and Brain Sciences 1978

Optional - False Belief

Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children’s understanding of deception, Cognition 1983
Do 15-Month-Old Infants Understand False Beliefs?, Science 2005
Action Anticipation Through Attribution of False Belief by 2-Year-Olds, Psychological Science 2007
Eighteen-month-old infants show false belief understanding in an active helping paradigm, Cognition 2009
From infants’ to children’s appreciation of belief, TiCS 2012
Three-year-olds’ difficulty with false belief: The case for a conceptual deficit, British Journal of Developmental Psychology 1987
Making sense of early false-belief understanding, TiCS 2014
Identity: Key to children’s understanding of belief, Science 2011

Optional - Joint Attention

The coordination of attention and action in great apes and humans, Philos. Trans. R. Soc. B 2021
Joint attention and early language, Child Development 1986
Infants’ contribution to the achievement of joint reference, Child Development 1991
Early referential understanding: Infants’ ability to recognize referential acts for what they are, Developmental Psychology 1993
The social-pragmatic theory of word learning, Pragmatics 2000
Two-year-olds exclude novel objects as potential referents of novel words based on pragmatics, Cognition 2009
Embodied attention and word learning by toddlers, Cognition 2012
Perceiving intentions and learning words in the second year of life, Language Acquisition and Conceptual Development (2001)

Optional - Joint Goals and Commitments

Chimpanzee helping in collaborative and noncollaborative contexts, Animal Behaviour 2010
Social cognition and metacognition in great apes: a theory, Animal Cognition 2022
Children’s Developing Commitments to Joint Goals, Child Development 2011
Three-Year-Olds’ Understanding of the Consequences of Joint Commitments, PLoS One 2013
3- and 5-Year-Old Children’s Adherence to Explicit and Implicit Joint Commitments, Developmental Psychology 2018
Every product needs a process: unpacking joint commitment as a process across species, Philos. Trans. R. Soc. B 2021

Abstract Reasoning

Wed, 20 Aug 2025 00:00:00 +0000

Reading list

survey/review/perspective paper book

Required

Dark, Beyond Deep: A Paradigm Shift to Cognitive AI with Humanlike Common Sense (Section 8), Engineering 2020
Mental models and human reasoning, PNAS 2010
How to Grow a Mind: Statistics, Structure, and Abstraction, Science 2011
Human-level concept learning through probabilistic program induction, Science 2015
Abstraction and analogy-making in artificial intelligence, Annals of the New York Academy of Sciences 2021

Optional

The structure-mapping engine: Algorithm and examples, Artificial Intelligence 1989
Psychological Reasoning in Infancy, Annual Review of Psychology 2016
The Perception of Relations, TiCS 2021
On the Measure of Intelligence, arXiv preprint arXiv:1911.01547
Rule Learning by Seven-Month-Old Infants, Science 1999
The Origin of Concepts, Oxford University Press (2009)
Bayesian Models of Conceptual Development: Learning as Building Models of the World, Annual Review of Developmental Psychology 2020
Scientific Thinking in Young Children: Theoretical Advances, Empirical Research, and Policy Implications, Science 2012
Very young infants learn abstract rules in the visual modality, PLOS One 2017
Emergent analogical reasoning in large language models, Nature Human Behaviour 2023
Human-level few-shot concept induction through minimax entropy learning, Science Advances 2024
Interactive Visual Reasoning under Uncertainty, NeurIPS 2023
Attention over Learned Object Embeddings Enables Complex Visual Reasoning, NeurIPS 2021
Large Language Models Are Not Strong Abstract Reasoners, IJCAI 2024

Optional - Raven’s Progressive Matrices (Spatial-Temporal)

What One Intelligence Test Measures: A Theoretical Account of the Processing in the Raven Progressive Matrices Tes, PsychReview 1990
RAVEN: A Dataset for Relational and Analogical Visual Reasoning, CVPR 2019
Measuring abstract reasoning in neural networks, ICML 2018
Learning Perceptual Inference by Contrasting, NeurIPS 2019
Abstract Diagrammatic Reasoning with Multiplex Graph Networks, ICLR 2020
Abstract Spatial-Temporal Reasoning via Probabilistic Abduction and Execution, CVPR 2021
Learning Algebraic Representation for Systematic Generalization in Abstract Reasoning, ECCV 2022

Optional - Abstraction and Concept Learning

DreamCoder: Bootstrapping Inductive Program Synthesis with Wake-Sleep Library Learning, PLDI 2021
The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision, ICLR 2019
Human-like systematic generalization through a meta-learning neural network, Nature 2023
Human-like Few-Shot Learning via Bayesian Reasoning over Natural Language, NeurIPS 2023
Learning Algebraic Representation for Systematic Generalization in Abstract Reasoning, ECCV 2022
HALMA: Humanlike Abstraction Learning Meets Affordance in Rapid Problem Solving, arXiv preprint arXiv:2102.11344

Optional - Blicket Detection (Causal)

Detecting Blickets: How Young Children Use Information about Novel Causal Powers in Categorization and Induction, Child Development 2000
When children are better (or at least more open-minded) learners than adults: Developmental differences in learning the forms of causal relationships, Cognition 2014
ACRE: Abstract Causal Reasoning Beyond Covariation, CVPR 2021

Optional - Mathematical Reasoning

Core systems of number, TiCS 2004
Abstract number and arithmetic in preschool children, PNAS 2005
Newborn infants perceive abstract numbers, PNAS 2009
Number without a language model, PNAS 2010
Representations of space, time, and number in neonates, PNAS 2014
Machine Number Sense: A Dataset of Visual Arithmetic Problems for Abstract and Relational Reasoning, AAAI 2020
Verbal counting and the timing of number acquisition in an indigenous Amazonian group, PLOS One 2022
Solving olympiad geometry without human demonstrations, Nature 2024
Proposing and solving olympiad geometry with guided tree search, arXiv preprint arXiv:2412.10673
Peano: learning formal mathematical reasoning, Phil. Trans. R. Soc. A 2023

Optional - Neuro-symbolic Reasoning

Optional - Differential Optimization

Differentiable Convex Optimization Layers, NeurIPS 2019
OptNet: Differentiable Optimization as a Layer in Neural Networks, ICML 2017
On Differentiating Parameterized Argmin and Argmax Problems with Application to Bi-level Optimization, arXiv preprint arXiv:1607.05447

Course Introduction & Logistics

Tue, 22 Sep 2026 00:00:00 +0000

Syllabus

Week 01 (09.09): Introduction, Topic Selection, Paper Reading, and Academic Integrity
Week 02 (09.16): Group Presentation Round 1, Part I
Week 03 (09.23): Group Presentation Round 1, Part II
Week 04 (09.30): Group Presentation Round 2, Part I
Week 05 (10.07): National Day Holiday, No Class
Week 06 (10.14): Group Presentation Round 2, Part II
Week 07 (10.21): Midterm Project Presentation
Week 08 (10.28): NVIDIA EBC Beijing Site Visit
Week 09 (11.04): Midterm Feedback Discussion and Q&A
Week 10 (11.11): Group Presentation Round 3, Part I
Week 11 (11.18): Group Presentation Round 3, Part II
Week 12 (11.25): Experiment Execution and Preliminary Results
Week 13 (12.02): Extended Experiments and Cross-Model Comparison
Week 14 (12.09): Writing Workshop I
Week 15 (12.16): Writing Workshop II
Week 16 (12.23): Capstone Project Presentation and Submission Preparation

Grading

Participation and attendance: 10%
Group presentations: 20%
Midterm assessment: 20%
Capstone project: 50%

Group Presentations

Round 1: cognitive dimension definition and literature review.
Round 2: existing benchmarks, open-source datasets, and replication plan.
Round 3: original test-suite design.
Format: group presentation, preferably in English for technical content.

Midterm Project Review

Content: research question, cognitive motivation, evaluation plan, literature review progress, and paper outline.
Paper outline: 1-2 pages in English.
Format: 15-minute group presentation.

Capstone Project

Team: 1-2 students.
Topic: choose one listed on the Research Topics page.
Duration: the entire semester.
Goal: develop the project into a research paper for submission to CogSci.
Template: Cognitive_Boundaries_Template.
Submission: course submissions will use PKU Netdisk.

Course Logistics, Introduction to Research, Academic Writing, and Ethics

Wed, 26 Aug 2026 00:00:00 +0000

Overview

In the first week, we will lay the foundation for the entire course by introducing key concepts in research, academic writing, and research ethics. This module is crucial for developing the skills and understanding necessary for successful engagement in AI research.

Instructor

Yixin Zhu, PKU

Topics Covered

Introduction to scientific research in AI
Principles of academic writing
LaTeX and Overleaf for document preparation
Research ethics and academic integrity
Effective research collaboration techniques
Overview of common research tools in AI

Assignments

There is no formal graded assignment this week. However, students are expected to:

Set up a LaTeX environment (preferably Overleaf)
Complete a simple LaTeX document to demonstrate basic proficiency
Familiarize themselves with the course structure and requirements

Additional Resources

Notes

All reports for future assignments in this course must be submitted in LaTeX format via Gradescope.
Students are encouraged to start thinking about potential research areas they would like to explore in the later weeks of the course.

Utility

Wed, 20 Aug 2025 00:00:00 +0000

Reading list

survey/review/perspective paper book Encyclopedia

Required

Dark, Beyond Deep: A Paradigm Shift to Cognitive AI with Humanlike Common Sense (Section 7), Engineering 2020
A Survey of Preference-Based Reinforcement Learning Methods, JMLR 2017
Optimal social choice functions: A utilitarian view, Artificial Intelligence 2015
Normative theories of rational choice: Expected utility
Ten-month-old infants infer the value of goals from the costs of actions Science 2017
Inferring Forces and Learning Human utilities from videos, CVPR 2016

Optional - Theories and Observational Studies

Chimpanzees Recruit the Best Collaborators, Science 2006
Chimpanzees Are Rational Maximizers in an Ultimatum Game, Science 2007
Chimpanzees play the ultimatum game, PNAS 2013
Children Discard a Resource to Avoid Inequity, JEPG 2011
Five-year-olds understand fair as equal in a mini-ultimatum game, Journal of Experimental Child Psychology 2013
Bayesian Psychology and Human Rationality, Rationality, Academic Press (2017)
A Theory of Fairness, Competition, and Cooperation, The Quarterly Journal of Economics 1999

Optional - Computational Models

Apprenticeship Learning via Inverse Reinforcement Learning, ICML 2004
The Naïve Utility Calculus: Computational Principles Underlying Commonsense Psychology, TiCS 2016
Reward is enough, Artificial Intelligence 2021
Deep Reinforcement Learning from Human Preferences, NeurIPS 2017
On the Expressivity of Markov Reward, NeurIPS 2021
Active Preference-Based Learning of Reward Functions, RSS 2017
Reward-rational (implicit) choice: A unifying formalism for reward learning, NeurIPS 2020
Replacing Rewards with Examples: Example-Based Policy Search via Recursive Classification, NeurIPS 2021
Preference-based Reinforcement Learning with Finite-Time Guarantees, NeurIPS 2020

Understanding and sharing intentions: The origins of cultural cognition, BBS 2005
Chimpanzees are indifferent to the welfare of unrelated group members, Nature 2005
Altruistic Helping in Human Infants and Young Chimpanzees, Science 2008
Varieties of altruism in children and chimpanzees, TiCS 2009
Toddlers’ Prosocial Behavior: From Instrumental to Empathic to Altruistic Helping, Child Development 2010
Young Children Selectively Avoid Helping People With Harmful Intentions, Child Development 2010
Examining the Diversity of Prosocial Behavior: Helping, Sharing, and Comforting in Infancy, Infancy 2011

Optional - Cooperation

Foundations of cooperation in young children, Cognition 2008
Young Children Share the Spoils After Collaboration, Psychological Science 2010
Collaboration encourages equal sharing in children but not in chimpanzees, Nature 2011
Children, but Not Chimpanzees, Prefer to Collaborate, Current Biology 2011
When and how do non-human great apes communicate to support cooperation?, Philos. Trans. R. Soc. B 2021
How 2- and 4-year-old children coordinate social interactions with peers, Philos. Trans. R. Soc. B 2021

Optional - Fairness

Fairness Expectations and Altruistic Sharing in 15-Month-Old Human Infants, PloS One 2011
The developmental origins of fairness: the knowledge-behavior gap, TiCS 2014
Development of in-group favoritism in children’s third-party punishment of selfishness, PNAS 2014
The ontogeny of fairness in seven societies, Nature 2015
The developmental foundations of human fairness, Nature Human Behavior 2017
The developmental roots of fairness: infants’ reactions to equal and unequal distributions of resources, Developmental Science 2011
Prospect Theory: An Analysis of Decision under Risk, Econometrica 1979

Optional - Ownership

Ownership Matters: People Possess a Naïve Theory of Ownership, TiCS 2019
Determining who owns what: Do children infer ownership from first possession?, Cognition 2008
First possession: An assumption guiding inferences about who owns what, Psychonomic Bulletin & Review 2008
That’s Mine! Learning Ownership Relations and Norms for Robots, AAAI 2019
Ownership and Value in Childhood, Annual Review of Developmental Psychology 2022

Optional - Morality

A mutualistic approach to morality: The evolution of fairness by partner choice, BBS 2013
The moral psychology of obligation, BBS 2020
Human Morality Is Based on an Early-Emerging Moral Core, Annual Review of Developmental Psychology 2022

XAI and Teaming

Wed, 20 Aug 2025 00:00:00 +0000

Reading list

survey/review/perspective paper book/book chapter Encyclopedia

Required - Explainable Artificial Intelligence (XAI)

The Nature of Explanation, Cambridge University Press (1943)
“Why Should I Trust You?”: Explaining the Predictions of Any Classifier, ACL 2016
Performance vs. competence in human-machine comparisons, PNAS 2020
A tale of two explanations: Enhancing human trust by explaining robot behavior, Science Robotics 2019

Required - Teaming

Trust in Automation: Designing for Appropriate Reliance, Human Factors 2004
In situ bidirectional human-robot value alignment, Science Robotics 2022

Optional - Debate over LLM

On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜, ACM FAccT 2021
Language Models Represent Space and Time, arXiv 2310.02207

Optional - Explainable Artificial Intelligence (XAI)

Explanation in artificial intelligence: Insights from the social sciences, Artificial Intelligence 2019
Interpretable CNNs for Object Classification, PAMI 2021
Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI, Information Fusion 2020
The Emerging Landscape of Explainable Automated Planning & Decision Making, IJCAI 2020
Explanation Perspectives from the Cognitive Sciences-A Survey, IJCAI 2020

Handouts

Wed, 26 Aug 2026 00:00:00 +0000

Slides

Google Drive

Module 0: Programming Best Practices

Wed, 26 Aug 2026 00:00:00 +0000

Overview

In the second week, we focus on essential programming best practices that are important for AI research and development. This module introduces tools and techniques designed to improve coding efficiency, code quality, and collaboration skills.

Instructor

Chi Zhang, PKU

Topics Covered

Linux and server basics
Python programming and Conda environment management
Version control with Git
Introduction to Hugging Face
Code standards and documentation
Debugging techniques for AI models
Basics of Continuous Integration/Continuous Deployment (CI/CD)

Assignments

Full Assignment: This assignment involves Git operations such as creating branches, committing changes, resolving conflicts, and pushing to a remote repository, as well as an introduction to Hugging Face. Please follow the instructions and document your work in a report. Your report should clearly describe the steps you completed, the commands or tools you used, any challenges or conflicts you encountered, and how you resolved them.

Short Assignment: There is no short assignment for Module 0.

Additional Resources

Notes

Students are encouraged to start using these tools and practices in their daily coding routines.
The skills learned this week will be essential for all future modules and assignments.
Remember to submit your assignment via Gradescope.

Module 1: Machine Learning

Wed, 26 Aug 2026 00:00:00 +0000

Overview

This week focuses on Machine Learning fundamentals and practices. We will cover essential concepts and tools used in modern ML workflows, with a particular emphasis on PyTorch and its ecosystem.

Instructor

Muhan Zhang, PKU

Topics Covered

PyTorch basics
- Installation (Linux-only)
- Core libraries and functions
CNN/Transformer training
- Training in PyTorch
- Hyperparameter tuning
Introduction to Weights & Biases (WandB)
Reproducibility in ML experiments

Assignments

Full Assignment: Using PyTorch, complete a simple classification task. Use WandB to record the entire experiment process, ensuring reproducibility.

Short Assignment: Write a one-page report (written in LaTeX) outlining the implementation process for Part 1 of the full assignment. Explain the required steps and specify which PyTorch functions you used.

Additional Resources

Notes

Ensure you have a Linux environment set up for PyTorch installation. If you are using a different OS, consider using a virtual machine or WSL (Windows Subsystem for Linux).
Familiarize yourself with WandB before starting the assignment; it will be crucial for tracking your experiments.
Submit your assignments via Gradescope.

Research Topics

Wed, 26 Aug 2026 00:00:00 +0000

Topics

What Unlocks a Blocked Word? Partial Cues and Lexical Search in Chinese Visual-Phonetic Puzzles

Question: When a solver is stuck on a Chinese visual-phonetic puzzle, what kind of partial information gets them moving again?

Possible methods: large-scale puzzle log analysis, hint-state reconstruction, Chinese lexical constraint modeling, cue-effect regression, human-VLM comparison.

References:

Goldblum, N., & Frost, R. (1988). The crossword puzzle paradigm: The effectiveness of different word fragments as cues for the retrieval of words. Memory & Cognition, 16(2), 158-166. https://doi.org/10.3758/BF03213485
Knoblich, G., Ohlsson, S., Haider, H., & Rhenius, D. (1999). Constraint relaxation and chunk decomposition in insight problem solving. Journal of Experimental Psychology: Learning, Memory, and Cognition, 25(6), 1534-1555. https://doi.org/10.1037/0278-7393.25.6.1534
Liang, J., Kabbara, A., Liu, J., Luo, R., Kim, K., & Guerzhoy, M. (2025). Semantic, orthographic, and phonological biases in humans’ Wordle gameplay. Findings of the Association for Computational Linguistics: IJCNLP-AACL 2025, 1128-1135. https://doi.org/10.18653/v1/2025.findings-ijcnlp.67
MacGregor, J. N., & Cunningham, J. B. (2008). Rebus puzzles as insight problems. Behavior Research Methods, 40(1), 263-268. https://doi.org/10.3758/BRM.40.1.263
Perfetti, C. A., Liu, Y., & Tan, L. H. (2005). The lexical constituency model: Some implications of research on Chinese for general theories of reading. Psychological Review, 112(1), 43-59. https://doi.org/10.1037/0033-295X.112.1.43
Shen, W., Hyona, J., Wang, Y., Hou, M., & Zhao, J. (2021). The role of tonal information during spoken-word recognition in Chinese: Evidence from a printed-word eye-tracking study. Memory & Cognition, 49(1), 181-192. https://doi.org/10.3758/s13421-020-01070-0

Physical-to-Animate Boundaries Across Humans and AI

Question: How do humans and AI systems infer animacy, goals, and intentions from physical motion?

Possible methods: parameterized stimuli, psychophysical measurement, human-AI behavioral comparison, computational modeling of physical-animate transitions.

References:

Griffiths, T. L., Chater, N., & Tenenbaum, J. B. (2024). Bayesian Models of Cognition: Reverse Engineering the Mind. MIT Press.
Heider, F., & Simmel, M. (1944). An experimental study of apparent behavior. The American Journal of Psychology, 57(2), 243-259.
Ji-An, L., Benna, M. K., & Mattar, M. G. (2025). Discovering cognitive strategies with tiny recurrent neural networks. Nature, 644(8078), 993-1001.
Shu, T., Peng, Y., Zhu, S.-C., & Lu, H. (2021). A unified psychological space for human perception of physical and social events. Cognitive Psychology, 128, Article 101398.
Zhou, C., Han, M., Liang, Q., Hu, Y.-F., & Kuai, S.-G. (2019). A social interaction field model accurately identifies static and dynamic social groupings. Nature Human Behaviour, 3(8), 847-855.

The Whorfian Judge: Grammar-Driven Bias in LLM Moral Judgment

Question: Do grammatical cues in different languages change LLM moral judgments in ways similar to human linguistic-relativity effects?

Possible methods: parallel accidental-event vignettes, agentive and non-agentive phrasing, cross-lingual prompts, moral blame and liability ratings.

References:

Fausey, C. M., & Boroditsky, L. (2010). Subtle linguistic cues influence perceived blame and financial liability. Psychonomic Bulletin & Review, 17(5), 644-650.
Wang, C. et al. (2025). Under the Shadow of Babel: How Language Shapes Reasoning in LLMs. arXiv:2506.16151.
Mihaylov, V., & Shtedritski, A. (2024). What an elegant bridge: Multilingual LLMs are biased similarly in different languages. Proceedings of the Fourth Workshop on Multilingual Representation Learning.
Lamia, L. Z., Hossain, M. F. B., & Khan, M. M. (2025). Who Holds the Pen? Caricature and Perspective in LLM Retellings of History. Proceedings of EMNLP 2025.
Keshmirian, A., Baltaji, R., Hemmatian, B., Asghari, H., & Varshney, L. R. (2025). Many LLMs Are More Utilitarian Than One. arXiv:2507.00814.

Event Representations of Verbs Across Semantic Levels

Question: Do human-like LLM judgments on verb-event similarity imply human-like representational geometry?

Possible methods: odd-one-out triplet judgments, human-model representational similarity analysis, embedding baselines, MDS or clustering.

References:

Binz, M. et al. (2025). A foundation model to predict and capture human cognition. Nature, 644, 1002-1009.
Borghi, A. M., Mazzuca, C., & Tummolini, L. (2025). The role of social interaction in the formation and use of abstract concepts. Nature Reviews Psychology, 4, 470-483.
Du, C. et al. (2025). Human-like object concept representations emerge naturally in multimodal large language models. Nature Machine Intelligence, 7, 860-875.
Gao, C. et al. (2025). Increasing alignment of large language models with language processing in the human brain. Nature Computational Science, 5, 1080-1090.
Griffiths, T. L., Lake, B. M., McCoy, R. T., Pavlick, E., & Webb, T. W. (2026). Whither symbols in the era of advanced neural networks? Trends in Cognitive Sciences.
Hebart, M. N., Zheng, C. Y., Pereira, F., et al. (2020). Revealing the multidimensional mental representations of natural objects underlying human similarity judgements. Nature Human Behaviour, 4(11), 1173-1185.

From Frames to Events: Human-Like Memory Organization in VLMs

Question: Do vision-language models organize continuous videos into event-level memory structures, or do they rely mainly on frame-level surface cues?

Possible methods: controlled synthetic videos, event boundary manipulations, prediction-error probes, memory questions about item binding and temporal order.

References:

Baldassano, C., Chen, J., Zadbood, A., Pillow, J. W., Hasson, U., & Norman, K. A. (2017). Discovering event structure in continuous narrative perception and memory. Neuron, 95(3), 709-721.
Zheng, J., Schjetnan, A. G., Yebra, M., Gomes, B. A., Mosher, C. P., Kalia, S. K., et al. (2022). Neurons detect cognitive boundaries to structure episodic memories in humans. Nature Neuroscience, 25(3), 358-368.
Sinclair, A. H., Manalili, G. M., Brunec, I. K., Adcock, R. A., & Barense, M. D. (2021). Prediction errors disrupt hippocampal representations and update episodic memories. Proceedings of the National Academy of Sciences, 118(51), e2117625118.
Liu, Y., Ma, Z., Qi, Z., Wu, Y., Shan, Y., & Chen, C. W. (2024). Et bench: Towards open-ended event-level video-language understanding. Advances in Neural Information Processing Systems, 37, 32076-32110.
Fountas, Z., Benfeghoul, M. A., Oomerjee, A., Christopoulou, F., Lampouras, G., Bou Ammar, H., & Wang, J. (2025). Human-inspired episodic memory for infinite context LLMs. International Conference on Learning Representations.

LLM Auto-Curriculum and Intrinsic-Motivation-Driven Agents

Question: Can an AI teacher generate adaptive curricula that improve learning in interactive agents, and how should progress be measured?

Possible methods: intrinsic motivation, curiosity-driven reward, active learning, teacher-student agent setup, open-world or multi-task environments.

References:

Schmidhuber, J. (2010). Formal theory of creativity, fun, and intrinsic motivation.
Oudeyer, P.-Y., Kaplan, F., & Hafner, V. V. (2007). Intrinsic motivation systems for autonomous mental development.
Pathak, D. et al. (2017). Curiosity-driven exploration by self-supervised prediction. ICML.
Burda, Y. et al. (2019). Large-scale study of curiosity-driven learning. ICLR.
Recent Progress: Intrinsic Motivation for Artificial Agents / Situated Self-guided Learning (2025).

AI-Assisted Writing and Creativity Perception

Question: How does AI-assisted writing change people’s perception of their own creativity, and is this effect mediated by cognitive load?

Possible methods: within-subject writing experiment, NASA-TLX, creativity self-attribution scale, repeated-measures ANOVA, mediation and moderation analysis.

References:

Amabile, T. M. (1983). The social psychology of creativity: A componential conceptualization. Journal of Personality and Social Psychology, 45(2), 357-376.
Bem, D. J. (1972). Self-perception theory. Advances in Experimental Social Psychology, 6, 1-62.
Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12(2), 257-285.
Norton, M. I., Mochon, D., & Ariely, D. (2012). The IKEA effect: When labor leads to love. Journal of Consumer Psychology, 22(3), 453-460.
Hart, S. G., & Staveland, L. E. (1988). Development of NASA-TLX. Advances in Psychology, 52, 139-183.
Epstein, R. et al. (2023). Who’s the author? Attribution of AI-generated text. Computers in Human Behavior, 144, 107690.
Deci, E. L., & Ryan, R. M. (2000). The what and why of goal pursuits: Self-determination theory. Psychological Inquiry, 11(4), 227-268.
Stanford HAI. (2024). AI Index Report 2024.
Bandura, A. (1997). Self-efficacy: The Exercise of Control. W. H. Freeman.
Gero, K. I. et al. (2022). Sparks: Inspiration for science writing using language models. Proceedings of CHI 2022.

Visual-Language Boundaries in Human Ventral Temporal Cortex

Question: How much do visual and visual-language AI models explain category-selective representations in human ventral temporal cortex, especially the emergence of word-selective responses?

Possible methods: CLIP/ResNet feature extraction, representational similarity analysis, VTC-focused or whole-brain searchlight RSA, infant connectivity analysis, SVR-based prediction of later functional maps.

References:

Kubota, E., Grill-Spector, K., & Nordt, M. (2023). Rethinking cortical recycling in ventral temporal cortex. Trends in Cognitive Sciences. https://doi.org/10.1016/j.tics.2023.09.006
Saygin, Z. M., Osher, D. E., Norton, E. S., Youssoufian, D. A., Beach, S. D., Feather, J., Gaab, N., Gabrieli, J. D. E., & Kanwisher, N. (2016). Connectivity precedes function in the development of the visual word form area. Nature Neuroscience, 19(9), 1250-1255. https://doi.org/10.1038/nn.4354
Chen, H., Liu, B., Wang, S., Wang, X., Han, W., Wang, X., Zhu, Y., & Bi, Y. (2026). Combined evidence from artificial neural networks and human brain-lesion models reveals that language modulates vision in human perception. Nature Human Behaviour, 10(3), 615-631. https://doi.org/10.1038/s41562-025-02357-5
Wang, A. Y., Kay, K., Naselaris, T., Tarr, M. J., & Wehbe, L. (2023). Better models of human high-level visual cortex emerge from natural language supervision with a large and diverse dataset. Nature Machine Intelligence, 5(12), 1415-1426. https://doi.org/10.1038/s42256-023-00753-y
Khosla, M., & Wehbe, L. (2026). Higher visual areas act like domain-general filters with strong selectivity and functional specialization. Nature Communications. https://doi.org/10.1038/s41467-026-73938-9

Linguistic Prediction and Gaze-Guided Neural Updating in Children

Question: During naturalistic movie viewing, do children’s linguistic predictions guide where they look, and does gaze selection shape neural updating when new words or events arrive?

Possible methods: GPT-2 surprisal, word embeddings, target-word event construction, eye-tracking features, gaze prediction, EEG encoding/decoding around word onset, individual-difference analysis.

References:

Broderick, M. P., Anderson, A. J., Di Liberto, G. M., Crosse, M. J., & Lalor, E. C. (2018). Electrophysiological correlates of semantic dissimilarity reflect the comprehension of natural, narrative speech. Current Biology, 28(5), 803-809.
Heilbron, M., Armeni, K., Schoffelen, J.-M., Hagoort, P., & de Lange, F. P. (2022). A hierarchy of linguistic predictions during natural language comprehension. Proceedings of the National Academy of Sciences, 119(32), e2201968119.
Haskins, A. J., Packard, K. O., & Robertson, C. E. (2026). Conceptual priorities shape individual gaze patterns during naturalistic visual attention. Proceedings of the National Academy of Sciences, 123(24), e2604369123.
Nau, M., Greene, A., Tarder-Stoll, H., Lossio-Ventura, J. A., Pereira, F., Chen, J., Baldassano, C., & Baker, C. I. (2025). Neural and behavioral reinstatement jointly reflect retrieval of narrative events. Nature Communications, 16, 7865.
Santos-Mayo, A., Gilbert, F., Mirifar, A., Tebbe, A.-L., Fang, R., Ding, M., & Keil, A. (2026). Concept2Brain: An AI model for predicting neurophysiological responses to text and pictures. Nature Communications.

Multimodal Computational Representations of Children’s Continuous EEG

Question: Which computational representations of naturalistic audiovisual experience best explain children’s continuous EEG during movie viewing?

Possible methods: acoustic, linguistic, visual, object-level, gaze-contingent, and multimodal feature extraction; encoding models; nested cross-validation; permutation tests; age or language-ability subgroup analysis.

References:

Broderick, M. P., Anderson, A. J., Di Liberto, G. M., Crosse, M. J., & Lalor, E. C. (2018). Electrophysiological correlates of semantic dissimilarity reflect the comprehension of natural, narrative speech. Current Biology, 28(5), 803-809.
Santos-Mayo, A., Gilbert, F., Mirifar, A., Tebbe, A.-L., Fang, R., Ding, M., & Keil, A. (2026). Concept2Brain: An AI model for predicting neurophysiological responses to text and pictures. Nature Communications.
Heilbron, M., Armeni, K., Schoffelen, J.-M., Hagoort, P., & de Lange, F. P. (2022). A hierarchy of linguistic predictions during natural language comprehension. Proceedings of the National Academy of Sciences, 119(32), e2201968119.
Nau, M., Greene, A., Tarder-Stoll, H., Lossio-Ventura, J. A., Pereira, F., Chen, J., Baldassano, C., & Baker, C. I. (2025). Neural and behavioral reinstatement jointly reflect retrieval of narrative events. Nature Communications, 16, 7865.
Haskins, A. J., Packard, K. O., & Robertson, C. E. (2026). Conceptual priorities shape individual gaze patterns during naturalistic visual attention. Proceedings of the National Academy of Sciences, 123(24), e2604369123.

Module 2: Computer Vision

Wed, 26 Aug 2026 00:00:00 +0000

Overview

This week focuses on Computer Vision, a key area in AI that deals with how computers can gain high-level understanding from digital images or videos. We will explore essential concepts, techniques, and frameworks used in modern computer vision tasks.

Instructor

Siyuan Huang, BIGAI

Topics Covered

Introduction to Computer Vision
Overview of torchvision
Application of mainstream frameworks (e.g., ViT) in vision tasks
Single-view depth estimation
Embodied Vision tasks

Assignments

Full Assignment:

Complete the following two tasks:

Task1: Use the pretrained models provided by Omnidata to estimate depth and normals.
Task2: Given multi-view RGB-D data (images and depth), complete the code in hw.ipynb to reconstruct the object’s mesh.

Short Assignment:

Complete the following task:

Task1: Use the pretrained models provided by Omnidata to estimate depth and normals.

Additional Resources

Notes

This module builds on the PyTorch skills you have developed in previous weeks. Make sure you are comfortable with basic PyTorch operations before diving into vision-specific tasks.
For the full assignment, consider how these vision models can be used in embodied AI contexts.
As always, document your code thoroughly and use version control (Git) for your project.
Submit your assignments via Gradescope.

NVIDIA EBC Beijing Visit

Wed, 26 Aug 2026 00:00:00 +0000

NVIDIA EBC Beijing Visit

The NVIDIA EBC Beijing visit is planned for Week 8.

Visit Information

Scheduled week: Week 8.
Date: October 28, 2026.
Tentative time: 14:00–17:00.
Venue: NVIDIA EBC Beijing, Fortune Financial Center Phase III, 41F.
Transportation: will be arranged collectively by the course.

Attendance and Safety Requirements

Students who confirm participation should reserve the full time for the visit.
Attendance will be checked before departure and after arrival at the venue.
Students should bring a valid ID document for visitor registration.
Students should follow the announced transportation plan and should not leave the group without informing the TA.
Students who cannot attend after confirming participation should contact the TA in advance.
Photography, recording, and sharing visit materials should follow NVIDIA visitor and confidentiality requirements.
Students should follow all on-site instructions from the instructor, TA, and NVIDIA staff.

Presentation Round 1: How to Give a Literature Review Presentation

Tue, 22 Sep 2026 00:00:00 +0000

Overview

A literature review presentation is not a sequence of paper summaries. Your goal is to help the audience understand a body of research: what has been studied, what the evidence collectively shows, where important disagreements remain, and what we still do not understand. The presentation should tell a coherent scientific story rather than reproduce everything you have read.

These guidelines apply to the Round 1 group presentation listed in the course syllabus; each presentation round runs over two class sessions. They are adapted in spirit from presentation materials used in MIT Brain and Cognitive Sciences courses, especially the guidelines for paper presentations and literature-review presentations in MIT 9.71 and 9.916.

Goals of Round 1

By the end of your presentation, the audience should understand:

What cognitive phenomenon or ability you are studying.
How cognitive science has studied it.
What the most important findings are.
How findings from different studies fit together.
Where the literature agrees or disagrees.
What important boundaries or open problems remain.
Why these findings matter for evaluating cognitive abilities in AI systems.

Preparing this presentation should also help you develop a deeper understanding of your project. By comparing and synthesizing the literature, you should begin to see how your topic is structured, which distinctions are theoretically important, what evidence is already available, and where meaningful open problems remain. This understanding will provide the conceptual foundation for the later stages of the project: evaluating existing benchmarks, identifying their limitations, designing your own tests, and eventually building a clear motivation and literature narrative for your final paper.

The goal is synthesis, not coverage. You are not expected to include every paper you have read.

Start with the Bigger Picture

Begin by helping the audience understand what broader phenomenon, problem, or debate the literature is trying to explain, and then move toward the specific issue your review addresses. Keep general background concise and include only what the audience needs to understand the rest of your presentation.

A useful introduction establishes three things in order: what is the phenomenon, why does it matter, and what needs to be explained. The opening should give the audience enough context to understand why the literature matters and what they should pay attention to.

For example, instead of beginning with

First, we will introduce Paper A, then Paper B, and then Paper C.

a stronger opening would first clarify the broader issue that connects those studies, giving the audience a clear frame for interpreting the evidence that follows.

For this course, that broader issue will often be a cognitive phenomenon or ability, such as:

How do humans distinguish animate behavior from ordinary physical motion?
How are continuous events organized into meaningful units in memory?

These examples are meant to organize the literature review. They do not need to be your final Capstone research question.

Clarify the Key Concepts and Scope

Before reviewing the evidence, clarify the key concepts that your presentation will rely on and define the scope of the literature you are reviewing. Do not assume that the audience uses important terms in exactly the same way you do; if a term has multiple meanings in the literature, briefly explain which meaning is relevant here.

It can also be useful to clarify:

What concepts or phenomena are central to the review?
How are closely related concepts distinguished?
What kinds of studies are included or excluded?
What level of explanation is the review concerned with?

The goal is to give the audience a clear conceptual frame before you begin comparing evidence across studies.

For this course, the central concept will often be a cognitive ability or phenomenon. Briefly explain how the literature defines and studies it, rather than relying on a dictionary definition.

Organize the Literature Around Ideas, Not Individual Papers

A literature review should help the audience understand how a body of research fits together. Avoid organizing the presentation as a sequence of independent paper summaries: if every paper becomes its own section, the audience may learn what each study did but still not understand what the literature collectively tells us.

Less Effective: One Section per Paper

Paper 1: background, methods, results.
Paper 2: background, methods, results.
Paper 3: background, methods, results.

This structure follows the order of the papers rather than the logic of the literature.

More Effective: One Section per Idea

Organize studies around larger ideas, such as:

a scientific issue;
a theoretical claim;
an empirical distinction;
a disagreement between explanations;
a boundary condition;
a methodological difference that changes the interpretation.

Then use individual studies as evidence for or against those ideas. A useful structure is idea or issue → evidence from multiple studies → what we learn. This allows the audience to see not only what has been studied, but also where findings converge, where they differ, and how the field has developed.

For this course, a review of animacy perception might be organized as:

Theme 1 — Motion cues contribute to animacy perception: evidence from Studies A and B.
Theme 2 — Motion alone may be insufficient: evidence from Studies C and D.
Theme 3 — Contingency and goal structure provide additional information: evidence from Studies E and F.

The papers support the structure of the argument; the papers themselves are not the structure.

Example to Explore: Organizing a Talk Around Debates

Lecture 7: Category Selectivity, Controversies, and MVPA from MIT 9.13 The Human Brain is a useful example of organizing a scientific talk around questions, debates, and evidence rather than individual papers. Notice how multiple studies are connected into a coherent argument and how conflicting findings are synthesized.

Present Only the Essential Evidence

A major part of preparing a literature review presentation is deciding what not to show. For each important study, include only what the audience needs to understand its contribution to your argument:

What did the study test?
What was the critical comparison or manipulation?
What was the key finding?
Why does this finding matter here?

A useful compression is problem → critical comparison → key finding → implication. Whenever possible, explain the experimental logic visually, using stimuli, condition diagrams, task timelines, or one key figure. You do not need to reproduce every methodological detail, analysis, or result from the original paper.

Example to Explore: Compressing a Single Experiment

Face Recognition: The Face Inversion Effect from Nancy’s Brain Talks shows how a single experiment is reduced to its essential logic: the critical comparison, the key finding, and what that finding tells us.

Synthesize and Evaluate the Literature

After presenting several studies, tell the audience what they mean together. A strong literature review should make clear what we know, where findings disagree, and what remains unclear.

Do not treat published conclusions as automatically correct. Consider whether:

different studies actually test the same phenomenon;
alternative explanations remain;
the design isolates the intended effect;
conflicting findings may result from different paradigms or assumptions.

The goal is to move from individual findings to a more informed view of the literature.

Example to Explore: Bringing Findings Back Together

Recap: What We Have Learned About Face Recognition from Behavioral Experiments from Nancy’s Brain Talks shows how evidence from several experiments is brought back together to explain what has been learned about face recognition as a whole.

Connect the Review to This Course

End by asking what the literature implies for understanding and evaluating AI:

What distinction appears important in human cognition?
What might this distinction imply for evaluating AI systems?

At this stage, you do not need to design the final experiment. Your goal is to identify the conceptual boundaries and open problems that will motivate the later stages of the project.

A Short Worked Example

Consider the topic Physical-to-Animate Boundaries Across Humans and AI. Instead of walking through Heider & Simmel → Paper B → Paper C → Paper D, organize the review around ideas:

Motion contributes to animacy perception: supporting evidence.
Motion alone may be insufficient: contrasting evidence.
Contingency and goal structure may also matter: additional evidence.
Synthesis: what information appears important for human animacy perception?
Open boundary for AI: does successful AI classification reflect the same distinctions?

The literature should form an argument, not a reading list.

Practical Presentation Tips

Tell one coherent story. Organize the talk around ideas, not a sequence of papers.
Be selective. Include only the background, methods, and results needed for your argument.
Make experimental logic clear. Use stimuli, diagrams, or key figures whenever possible.
State the take-home message. Do not expect the audience to interpret a figure on their own.
Keep clear transitions. Explain why each theme or study follows from the previous one.
Cite the studies you discuss.
Practice the full presentation and make sure it fits the assigned time.

Before You Present: Checklist

Is the broader problem clear?
Are the key concepts and scope clearly explained?
Is the literature organized around ideas rather than individual papers?
Have you selected only the evidence needed for your argument?
Can the audience understand the logic of the key studies?
Have you explained what the findings collectively show, where they disagree, and what remains unclear?
Have you connected the literature to the broader goals of this course?
Are your figures readable and sources properly cited?
Have you rehearsed the full presentation within the time limit?

What You Should Have After Round 1

By the end of Round 1, your project should have:

A clearly defined cognitive phenomenon.
A structured view of the relevant human literature.
Major empirical findings and theoretical distinctions.
Consensus, conflicts, and open boundaries.

These will become the foundation for Round 2, where you will examine how existing computational methods, datasets, and benchmarks operationalize and evaluate the same cognitive ability.

References

Module 3: Cognitive Reasoning

Wed, 26 Aug 2026 00:00:00 +0000

Overview

This week explores the fascinating field of Cognitive Reasoning in AI, focusing on how machines can mimic human-like reasoning processes. We will delve into various aspects of cognitive reasoning, including physical reasoning, social reasoning, communication, and abstract learning.

Instructor

Yixin Zhu, PKU

Topics Covered

Physical reasoning in AI systems
Social reasoning and its applications
Concept learning
Communication in AI
The importance of reading and thinking in AI

Assignments

In this assignment, you will read the recommended materials and write a paper. This assignment is designed to encourage you to engage deeply with foundational and cutting-edge concepts in AI and cognitive science. It encourages you to think critically about the evolution of AI theories and their practical applications.

Full Assignment: Your paper should be at least 4 pages (not including references).

Short Assignment: Your paper should be at least 2 pages (not including references).

Additional Resources

Bayesian Reasoning and Machine Learning by David Barber
Artificial Intelligence: A Modern Approach by Stuart Russell and Peter Norvig
The Society of Mind by Marvin Minsky
How the Mind Works by Steven Pinker

Notes

This module builds on concepts from previous weeks.
For the full assignment, focus on the reasoning process rather than complex implementations.
In your short assignment, consider both the current state of cognitive reasoning in AI and potential future developments.
Submit your assignments via Gradescope.

AI Token API Guide

Mon, 24 Aug 2026 00:00:00 +0000

本文档提供 AI Token API 平台使用指南，帮助学生完成课程中的相关任务。

硅基流动 SiliconFlow

注册流程

网址：https://cloud.siliconflow.cn
这里可以填写手机号，并且填写邀请码

用户A使用用户B的邀请码注册，用户A、B都可以得一张16元的代金券（且无上限）
学生可以互相填邀请码
TA的邀请码：BbcPNQda
- 或者直接通过右侧网址注册，自动成为被TA邀请的用户：https://cloud.siliconflow.cn/i/BbcPNQda

填写邀请码、注册成功后进去进行实名认证

需要学生满18周岁；可以用于实名认证的证件类型有

需要支付宝辅助完成刷脸认证：

认证完领取代金券：

⚠️一种失败的方式

网址：https://www.siliconflow.com/（不要使用这个网址，需要使用上方注册流程中的中文站网址）

会出现两种登录方式：

跳转至主要界面：

没找到实名认证的地方！

所以需要使用中文站的网址，通过手机号注册，完成实名认证

Notes

同一实名个体无法通过注册多个号领取多次代金券

可通过邀请获得无上限的额外代金券
- 截止日期：2026年12月31日，需尽快完成注册邀请！

主页下方【活动中心】-【推荐官计划】中，点击【成为推荐官】，即可获得自己的邀请码

代金券是有有效期的

使用方式

不充钱也可以安全使用！

新建API密钥，妥善保管

有详细的API使用手册

https://api-docs.siliconflow.cn/docs/api/chat-completions-post

可以看到账户没钱也可以调用

智谱｜BigModel

注册流程

网址：https://open.bigmodel.cn/
- 登录/注册方式：手机号or微信扫码

注册之后，即可获得注册福利2000万tokens
- 添加技术顾问可以额外领2000万 Tokens

Notes

不用实名认证，多个手机号可以领取多次
2000万tokens不全是通用额度，部分会限定使用的模型
赠送的tokens有3个月的使用期限，需要尽快使用防止过期

使用方式

【控制台】-【API Key】，新建项目专属API，或者直接使用平台已有的默认项目，没太所谓
有详细的API文档

https://docs.bigmodel.cn/cn/api/introduction

只能调用智谱自己的模型GLM系列
也有免费模型可以调用
- GLM-4.7-Flash
- GLM-4-Flash-250414
- GLM-Z1-Flash

阿里云百炼

注册流程

有多种登录/注册方式

需要访问阿里云百炼-华北2（北京）地域，阅读并同意协议后，系统将自动开通阿里云百炼并发放免费推理额度。

如果未弹出服务协议，表示已经开通过阿里云百炼且获得免费额度

账号中有免费额度
- 有15个各100万tokens的模型可以调用，总共1.5亿tokens的免费额度

Notes

一定要开启免费额度用完即停：仅消耗平台赠送的免费额度，免费额度用尽后平台将自动停止服务

每个model的免费额度有效日期不同，需要关注过期时间
当某个模型额度用完后，系统不会自动切换到其他有额度的模型。如需使用其他模型的免费额度，需在代码或工具配置中手动修改 model 参数。
免费额度剩余量为输入 Token 和输出 Token 共用的总额度，不单独区分输入或输出计算。调用时产生的输入 Token 和输出 Token 会共同扣减该总额度，控制台显示的剩余量即为扣除后的总剩余 Token 数。

使用方式

新建API key
- 需要记得保存好key，只显示一次

有API使用文档

https://bailian.console.aliyun.com/cn-beijing?tab=doc#/doc/?type=model&url=2840915

Presentation Round 2: From Literature to a Research Plan

Tue, 22 Sep 2026 00:00:00 +0000

Overview

Round 2 is where your project moves from understanding a research area to designing a study. In Round 1, your goal was to understand a cognitive phenomenon: how it has been defined, what evidence exists, where findings agree or disagree, and what important questions remain open. In Round 2, your goal is to turn one of those open questions into a testable first version of your research project.

This is not another literature review. Prior work now serves a different purpose: it helps you determine what has already been tested, how it has been operationalized, what remains missing, and how your own study can make a meaningful next step.

These guidelines apply to the Round 2 group presentation listed in the course syllabus; each presentation round runs over two class sessions.

Goals of Round 2

By the end of your presentation, the audience should understand:

What is the question?
Why does it matter?
How has prior work approached it?
How will you test it?
What results would be informative?
Where does the design still need refinement?

From Round 1 to a Research Gap

Briefly remind the audience of the one or two conclusions from Round 1 that motivate your next step. A useful transition is what do we already know → what remains unresolved → what should we test next. Your Round 2 question should be narrower and more testable than the broad topic reviewed in Round 1.

For example, a project on animacy perception might move from Round 1 to Round 2 as follows:

Human observers use both motion and goal information when judging whether behavior is animate.
Existing AI systems can often classify animate behavior correctly.
But do humans and AI rely on the same information when making these judgments?

Find Your Nearest-Neighbor Studies

Identify approximately 2–4 studies that are closest to the study you now want to conduct. These are not necessarily the most famous papers in your topic. They are the studies whose research question, stimuli, task, dataset, computational approach, or evaluation method most closely resemble what you plan to do.

For each study, focus on question → operationalization → critical comparison → finding → limitation. Do not give another complete paper presentation. The goal is to understand the existing design space and locate the smallest meaningful step beyond it.

Example to Explore: Finding the Nearest Prior Work

The Stanford CS197C project proposal asks students to identify the work most similar to their proposed project and then articulate the meaningful change that distinguishes the new study. The useful question is not simply “Has anyone done this exact project?” but: what is the closest existing study, and what scientifically meaningful thing are we changing?

Identify the Meaningful Difference

A good project usually changes one important assumption, variable, population, modality, task, or comparison.

For example:

Existing work: models perform well on a task in which two potentially important cues always appear together.
Open question: which cue is actually responsible for the successful behavior?
Meaningful change: manipulate the two cues independently while keeping the rest of the task constant.

This allows the new study to distinguish between two explanations that the original design could not separate.

Operationalize the Research Question

A research question becomes experimentally useful only when its key concepts can be connected to observable data. A useful structure is cognitive construct → stimulus / data → task → response → measure. For your project, clarify two things.

What Is the Construct?

What cognitive ability, representation, process, or distinction are you trying to understand?

What Is the Evidence?

What will participants or models observe, and what response will they produce? Consider:

Stimulus / data: words, images, videos, narratives, interactive environments, or existing datasets.
Task: classify, choose, predict, compare, recall, interact, or generate.
Response: choice, text, probability, action, trajectory, or representation.
Measure: accuracy, reaction time, choice probability, similarity structure, error pattern, or another metric.

Most importantly, ask why this measure should tell us something about the construct we care about. A task producing a numerical score does not automatically mean that the score measures the intended cognitive ability.

Example to Explore: From an Idea to an Operational Plan

The Stanford Undergraduate Research project proposal guidelines show how a research proposal connects what the project asks, why it matters, and how the study will answer it. When reviewing your own design, ask whether the chain research question → method → evidence is clear.

For AI Evaluation: Make the Interface Explicit

When your project evaluates AI, operationalization requires an additional step: specify exactly what information the model receives, how the task is presented, and how its response becomes evidence. A useful structure is input → prompt / adaptation → model → output → scoring. Clarify:

Input: what information does the model receive? Text, image, video, demonstrations, interaction history, or other context?
Prompt / adaptation: how is the task presented? What instructions, examples, or response constraints are provided?
Output: what counts as the model’s response? A multiple-choice answer, probability, generated text, action, trajectory, or representation?
Scoring: how is that output converted into a measurement? Accuracy, similarity, likelihood, human rating, behavioral agreement, or another metric?
Model conditions: what must remain controlled? Model / version, decoding settings, number of repeated runs, and other settings that may affect the result.

For this course, one additional question is especially important: are the human and AI tasks informationally comparable?

Giving humans and AI the “same task” does not necessarily mean presenting the same surface stimulus. Make clear what information is available to humans, what information is available to the model, what differs between them, and why the comparison is still informative about the cognitive ability of interest.

For example, if humans watch a video but an LLM receives only a textual description, a performance difference may reflect input modality rather than the cognitive ability itself. Likewise, if the model receives detailed demonstrations or instructions that human participants do not receive, the two systems may effectively be solving different tasks.

The full inferential chain should therefore be explicit: cognitive question → human / AI input → task interface → response → measure → inference.

Examples to Explore: Standardized AI Evaluation

Stanford CRFM — Holistic Evaluation of Language Models (HELM) emphasizes standardized scenarios, adaptation procedures, prompts, and metrics, reminding us that model performance cannot be interpreted independently of how the task is presented.
The NIST Generative AI Evaluation Plan explicitly specifies task definitions, task conditions, input/output formats, metrics, and evaluation protocols. An AI benchmark is not just a dataset and a score; the full evaluation interface must be defined.

Develop a Preliminary Experimental Design

Round 2 does not require a final, perfect design. It requires a first version that is specific enough to test, critique, and improve. Your preliminary design should make clear:

the research question and prediction;
the conditions or comparison groups;
the key controls or baselines;
a basic analysis or evaluation plan.

You do not need every parameter finalized. What matters is that someone else can understand what will be compared and why that comparison answers your question.

Make the Critical Comparison Explicit

Every useful study has a comparison that does the main inferential work. Ask: what comparison would distinguish the explanation we care about from an alternative explanation? Examples include:

Condition A vs. Condition B;
humans vs. AI on the same task;
a model with information X vs. a model without information X;
observed performance vs. a baseline or null model.

You should also be able to state the expected qualitative pattern:

If H1 is supported: A > B.
If H1 is not supported: A ≈ B.

If the critical comparison is unclear, the design probably needs further refinement.

Example to Explore: What a Computational Project Proposal Should Specify

The Stanford CS224S course project guidelines connect the task, approach, and evaluation so that the reader can see what will actually be tested.

Check Feasibility and Scope

Before moving forward, ask:

Can the required data or stimuli be obtained?
Can the study be completed within the semester?
Are participant recruitment, ethics, model access, and computational cost manageable?
What is the minimum viable study that can still answer the question?

A focused study that cleanly answers one question is better than an ambitious design with no clear inference.

Show and Structure Your Study

A strong Round 2 presentation should make the research logic easy to follow and concrete enough to evaluate. Whenever possible, include one slide that shows the study as a pipeline, for example stimuli / data → task → human / AI responses → measures → comparison, or Condition A vs. Condition B → critical difference → supports / challenges hypothesis. Someone unfamiliar with your project should be able to understand what goes in, what happens, what comes out, and what gets compared.

Show Something Concrete

You do not need final results in Round 2. However, when possible, show an early piece of the actual study:

candidate stimuli or an example trial;
a questionnaire, prototype, or interface;
sample model outputs;
a baseline implementation;
a small pilot or preliminary visualization.

The goal is to make the design concrete and expose practical problems early.

Organize the Presentation Around the Research Logic

There is no mandatory slide template, but a strong presentation will often follow question → existing work → gap → design → evidence → open questions:

Research question: what exactly are you trying to test?
Nearest existing work: what are the closest prior studies or benchmarks?
Gap: what important question remains unresolved?
Preliminary design: how will you test it?
Evidence: what comparison, metric, or result pattern would answer the question?
Open questions: what still needs feedback or refinement?

Spend less time repeating the literature from Round 1 and more time showing what you plan to test and why the design can answer the question.

Ask for Targeted Feedback

Round 2 should function partly as a design review. End with 2–4 specific unresolved questions that would help improve the next version of the study. For example:

Are these stimuli isolating the intended variable?
Is the current baseline appropriate?
Should this factor be manipulated or controlled?
Does the proposed measure distinguish the competing hypotheses?

Avoid ending with

Any suggestions?

Ask questions that can lead directly to a design decision.

Example to Explore: How a Project Progresses

The Stanford CS341 project information and checkpoints show how research progresses from problem and related work → methodology and experimental setup → preliminary results and current challenges. A good Round 2 presentation should leave the audience with both a clear understanding of the proposed study and a clear sense of what needs to happen next.

Practical Reminders

Build on the nearest existing work.
Make one meaningful scientific change.
Keep the construct distinct from the measure.
Make the critical comparison explicit.
Use appropriate controls or baselines.
Consider alternative explanations.
Keep the study feasible and pilot early.
Ask for targeted feedback.

Before You Present: Checklist

Question and Gap

Can we state the research question clearly?
Do we know the closest existing studies?
Can we explain what remains unresolved and why it matters?

Design

Is the construct operationalized clearly?
Are the inputs or stimuli, task, responses, and measures explicit?
Is the critical comparison clear?
Are the main controls or baselines included?

Evidence

Do we know what result would support or challenge the hypothesis?
Have we considered at least one alternative explanation?

Feasibility

Can the minimum viable study be completed this semester?
Do we know what participants, data, models, tools, or approvals are required?

Presentation

Can someone understand the study from the experimental diagram?
Have we shown something concrete where possible?
Have we prepared specific questions for feedback?

What You Should Have After Round 2

By the end of Round 2, your project should have:

A focused research question.
The nearest existing work.
A clearly motivated research gap.
A preliminary, testable design.
A critical comparison and evaluation plan.
Expected result patterns and alternative explanations.
Specific unresolved design decisions.

These are not only preparation for the next stage of the project. They should also begin to form the backbone of your final paper:

Research question and gap → Introduction.
Nearest existing work and unresolved issues → Related Work.
Operationalization, experimental design, and evaluation plan → Methods.
Expected result patterns → the analyses and figures you will eventually need in Results.
Alternative explanations and unresolved decisions → Discussion, Limitations, and Future Directions.

Your Round 2 presentation should therefore leave you with a first answer to what are we asking, why is it still unresolved, how will we test it, and what evidence would answer the question. These materials will become the foundation for the next stage, where you will refine the design, implement the task or benchmark, pilot the study, establish baselines, and begin collecting preliminary results.

References

Stanford Undergraduate Research — Detailed Project Proposal Guidelines: a general guide to connecting the research question, significance, approach, and project scope.
Stanford CS197C — Project Proposal: useful for thinking about nearest-neighbor work, meaningful differences, and project contribution.
Stanford CS224S — Course Project: useful for checking whether the task, data, approach, and evaluation are sufficiently explicit.
Stanford CS341 — Project Information and Checkpoints: a useful example of how a project progresses from problem definition and related work to methodology, implementation, and preliminary results.
Stanford CRFM — Holistic Evaluation of Language Models (HELM): useful for AI-focused projects; pay attention to scenarios, prompts and adaptation, model outputs, and evaluation metrics.
NIST — Generative AI Evaluation Plan: useful for thinking carefully about task conditions, input/output formats, metrics, and evaluation protocols.

Module 4: Robotics

Wed, 26 Aug 2026 00:00:00 +0000

Overview

This week delves into the fascinating world of Robotics, exploring how AI and machine learning techniques can be applied to create intelligent robotic systems. We’ll cover fundamental concepts, popular frameworks, and practical applications in the field of robotics.

Instructor

Yixin Zhu, PKU

Topics Covered

Introduction to ROS (Robot Operating System) and Gazebo
Overview of OmniVerse and TongVerse
Fundamentals of motion planning in robotics
Integration of AI techniques in robotic systems

Assignments

Full Assignment: Complete a simple robotic task in a simulation environment. You will use various tools and frameworks to understand the principles of robot motion planning. Refer to lab.md for more details.

Short Assignment: Refer to homework.md and submit a PDF report (written in LaTeX) via Gradescope.

Additional Resources

Notes

This module combines concepts from previous weeks (especially computer vision and machine learning) with robotics-specific knowledge.
For the full assignment, focus on implementing a working solution rather than achieving perfect performance.
Consider the ethical implications of AI in robotics when writing your paper.
As always, document your code thoroughly and use version control (Git) for your project.
Submit your assignments via Gradescope.

Module 5: Natural Language Processing

Mon, 14 Sep 2026 00:00:00 +0000

Overview

This week focuses on Natural Language Processing (NLP), a crucial field in AI that deals with the interaction between computers and humans using natural language. We’ll explore fundamental concepts, modern architectures, and practical applications of NLP.

Instructor

Zilong Zheng, BIGAI

Topics Covered

From PyTorch to Transformers
Tokenization, Embedding, and Attention mechanisms
Encoder-Decoder and Transformer architectures
Introduction to LangChain and LlamaIndex (Optional)

Assignments

Full Assignment:

Implement a task-oriented agent system with a context graph that incrementally extracts structured, interpretable, and auditable graph nodes from the agent’s tool-call results, reasoning traces, and intermediate conclusions. Build on the open-source Belief Context Graph (BCG) project.
Use the constructed graph to identify and improve the weakest link in the agent’s reasoning chain, such as evidence-grounded verification, parallel tool-call planning, contradiction detection, or error attribution.
Evaluate your system on the BrowseComp benchmark, and report the quantitative results (task performance and token cost) in your report.

Short Assignment: For the short assignment, you only need to submit a PDF report (written in LaTeX). The report should include the implementation approach for the full assignment.

Additional Resources

Notes

This module builds upon your PyTorch knowledge and introduces NLP-specific concepts and libraries.
Pay special attention to the attention mechanism and its implementation in Transformers.
For the full assignment, consider experimenting with pre-trained models and fine-tuning them for your specific task.
As always, document your code thoroughly and use version control (Git) for your project.
Submit your assignments via Gradescope.

Module 6: AI Hardware and Acceleration

Wed, 26 Aug 2026 00:00:00 +0000

Overview

This week delves into the world of AI Hardware and Acceleration, exploring how specialized hardware and acceleration techniques can significantly enhance the performance of AI systems. We will cover various aspects of hardware optimization and acceleration methods used in modern AI applications.

Instructors

Xiyuan Tang and Yaoyu Tao, PKU

Topics Covered

Introduction to AI hardware acceleration
Data sensing and processing for edge intelligent
Software-hardware co-design for AI chip architectures
Emerging AI computation acceleration technologies

Assignments

Full Assignment: Compare matrix multiplication performance using NumPy and PyTorch with varying matrix sizes and dimensions. Analyze the impact of GPU acceleration on training time and accuracy for an MNIST model.

Short Assignment: Compare matrix multiplication performance using NumPy and PyTorch with varying matrix sizes and dimensions.

Additional Resources

Efficient Processing of Deep Neural Networks by Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, Joel Emer
AI and Memory Wall by Amir Gholami
Hardware Acceleration for Efficient Deep Learning by Intel
Edge AI: The Future of Artificial Intelligence by NVIDIA
Slurm Workload Manager - Official documentation for Slurm
Slurm Quick Start Tutorial - A beginner’s guide to using Slurm
Using GPUs with Slurm - Guide on how to request and use GPUs in a Slurm environment
Slurm for GPU Scheduling - Detailed guide on submitting GPU jobs with Slurm

Notes

This module integrates concepts from previous weeks, particularly in machine learning and robotics, with a focus on hardware optimization.
For the full assignment, you will need access to a GPU. You are encouraged to utilize the computing resources provided for this course (see Computing Resources). Note that the course does not provide GPU compute or compensation for GPU hours to non-Tong Class students; you will need to arrange GPU access yourself, whether through your own resources, rented GPU hours from a cloud provider, or a free tier such as Google Colab or Kaggle Notebooks. If your cluster uses Slurm for job scheduling, familiarize yourself with basic Slurm commands such as srun, sbatch, and squeue.
For the short assignment, try to connect the hardware advancement you discuss with potential applications or improvements in AI systems.
As always, document your code thoroughly and use version control (Git) for your project.
Submit your assignments via Gradescope.

Module 7: Multi-Agent Systems

Wed, 26 Aug 2026 00:00:00 +0000

Overview

This week focuses on Multi-Agent Systems and Reinforcement Learning, exploring how multiple AI agents can interact, learn, and solve complex problems collectively. We will cover fundamental concepts, algorithms, and applications in this exciting field of AI.

Instructor

Yaodong Yang, PKU

Topics Covered

Introduction to Multi-Agent Systems
Fundamentals of Reinforcement Learning
Ray framework for distributed computing
Multi-agent concepts and applications
Introduction to distributed learning

Assignments

Full Assignment:

Implement the PPO algorithm and test it on the CartPole environment.
Implement the MADDPG algorithm and test it in the Speaker Listener environment.

Short Assignment:

Implement the PPO algorithm and test it on the CartPole environment.

Additional Resources

Multi-Agent Reinforcement Learning by Stefano V. Albrecht, Filippos Christianos, and Lukas Schäfer
Ray Documentation
Reinforcement Learning: An Introduction by Richard S. Sutton and Andrew G. Barto
Multiagent Systems: Algorithmic, Game-Theoretic, and Logical Foundations by Yoav Shoham and Kevin Leyton-Brown

Notes

This module builds on concepts from previous weeks, particularly in machine learning.
For the full assignment, focus on setting up a working multi-agent environment and implementing a basic RL algorithm.
In your short assignment, consider both current applications and potential future developments in Multi-Agent Systems.
As always, document your code thoroughly and use version control (Git) for your project.
Submit your assignments via Gradescope.

Capstone Project

Wed, 26 Aug 2026 00:00:00 +0000

Overview

This extended module, spanning an entire month, is dedicated to devising and developing your own AI project. You will have the opportunity to apply the knowledge and skills you have acquired throughout the course to a real-world problem of your choice. This module emphasizes project planning, implementation, and iterative development. You will also document your work, analyze your results, and prepare a presentation to showcase your project. Writing and presenting your project report is a crucial step in the scientific process, allowing you to communicate your findings effectively and reflect on your work.

Instructors

Yixin Zhu, Chi Zhang, Muhan Zhang, Yaodong Yang, Xiyuan Tang, and Yaoyu Tao, PKU

Topics Covered

Project Development

Project ideation and problem definition
Proposal writing and project planning
Implementation of AI models and systems
Integration of multiple AI technologies (ML, CV, NLP, robotics, cognitive reasoning, MAS, etc.)
Ethical considerations in AI project development
Performance evaluation and optimization
Project documentation and presentation

Report Write-up

Scientific writing principles
Structure of a research paper
Data visualization and presentation
Critical analysis and discussion of results
Ethical considerations and future work
Oral presentation skills

Assignments

Throughout these weeks, you will work on your individual capstone project under the guidance of a randomly assigned advisor. Regular check-ins and in-person progress reports will be required. You will also complete a mid-project review by submitting a two-page abstract that presents the project’s introduction, related work, and methodology. The mid-project assessment will be worth 10% of your final grade. At the end of the module, you must submit the final deliverables listed below; the project report accounts for 20% of your final grade. Your assigned advisor will also provide an overall evaluation of your work, worth a further 5%.

Final Deliverables

A fully implemented AI project with source code and documentation.
A comprehensive project report (at least 4 pages in length, excluding references, formatted in the LaTeX template) detailing your problem, approach, implementation, results, and discussion.
A 10–15 minute recorded presentation showcasing your project.
Submit your code with documentation, report (LaTeX), and presentation video through Gradescope. The presentation video should be shared via a timestamped PKU Disk link.

Additional Resources

Notes

You are encouraged to utilize the computing resources provided for this course (see Computing Resources) in your project implementation, and to consult your project advisor for compute support.
Regular consultations with your assigned advisor will be available throughout these weeks. Make sure to schedule sessions to get feedback and guidance on your project.
Consider the ethical implications of your AI project. Refer to the AI ethics guidelines discussed in previous modules.
Start working on your project early and iterate based on feedback. Do not leave everything to the last minute!
Start writing your report early. Writing often clarifies thinking and may reveal areas where additional work is needed.
Use clear, concise language. Avoid jargon where possible, and explain technical terms when they are first used.
Make your figures and tables self-explanatory. They should be understandable even without the main text.
Practice your presentation multiple times. Time yourself to ensure you stay within the 15-minute limit.
Be prepared to answer questions about your work. Think about potential questions in advance.
Remember to acknowledge any collaborators, advisors, or resources that significantly contributed to your project.
Your capstone project could be a significant part of your portfolio, potentially useful for future academic or career opportunities!

Write Project Report

Sat, 15 Aug 2026 00:00:00 +0000

Overview

This final module focuses on writing and presenting your project report. You will document your work, analyze your results, and prepare a presentation to showcase your project. This is a crucial step in the scientific process, allowing you to communicate your findings effectively and reflect on your work.

Instructors

Yixin Zhu, Chi Zhang, Muhan Zhang, Yaodong Yang, Xiyuan Tang, and Yaoyu Tao, PKU

Topics Covered

Scientific writing principles
Structure of a research paper
Data visualization and presentation
Critical analysis and discussion of results
Ethical considerations and future work
Oral presentation skills

Final Deliverables

A fully implemented AI project with source code and documentation.
A comprehensive project report (at least 4 pages in length, excluding references, formatted in the LaTeX template) detailing your problem, approach, implementation, results, and discussion.
A 10–15 minute recorded presentation showcasing your project.
Submit your code with documentation, report (LaTeX), and presentation video through Gradescope. The presentation video should be shared via a timestamped PKU Disk link.

Additional Resources

Notes

Submit all your deliverables (code, report, and recorded presentation) via Gradescope in the required formats.
Start writing your report early. Writing often clarifies thinking and may reveal areas where additional work is needed.
Use clear, concise language. Avoid jargon where possible, and explain technical terms when they are first used.
Make your figures and tables self-explanatory. They should be understandable even without the main text.
Practice your presentation multiple times. Time yourself to ensure you stay within the 15-minute limit.
Be prepared to answer questions about your work. Think about potential questions in advance.
Remember to acknowledge any collaborators, advisors, or resources that significantly contributed to your project.

[EMNLP26] Inferring the Unspoken: Aligning Embodied Agents with Implicit Preferences

Sat, 29 Aug 2026 00:00:00 +0000

4DStreamCtrl: Interactive Video Generation with Online 4D Control

Sat, 29 Aug 2026 00:00:00 +0000

[TPAMI26] AffordX: Generalizable and Slim Affordance Reasoning for Task-oriented Manipulation

Sat, 01 Aug 2026 00:00:00 +0000

[Science26] A sub–10-millisecond neural dynamical system based on phase-change memristors

Sun, 05 Jul 2026 00:00:00 +0000

[SmartBot26] T-800: An 800 Hz Data Glove for Precise Hand Gesture Tracking

Thu, 11 Jun 2026 00:00:00 +0000

[IJCV26] AlphaChimp: Tracking and Behavior Recognition of Chimpanzees

Tue, 14 Apr 2026 00:00:00 +0000

[CogSci26] Rational Communication Shapes Morphological Composition

Mon, 13 Apr 2026 10:00:00 +0000

[CogSci26] Timing is Everything: Temporal Scaffolding of Semantic Surprise in Humor

Mon, 13 Apr 2026 10:00:00 +0000

[CogSci26] Multi-Level Narrative Evaluation Outperforms Lexical Features for Mental Health

Mon, 13 Apr 2026 09:30:00 +0000

[CogSci26] Overhang Tower: Resource-Rational Adaptation in Sequential Physical Planning

Mon, 13 Apr 2026 04:00:00 +0000

[CogSci26] Grounding Before Generalizing: How AI Differs from Humans in Causal Transfer

Mon, 13 Apr 2026 00:00:00 +0000

[T-ASE26] TacMan-Turbo: Proactive Tactile Control for Robust and Efficient Articulated Object Manipulation

Sat, 11 Apr 2026 00:00:00 +0000

[ACL-Findings26] Code over Words: Overcoming Semantic Inertia via Code-Grounded Reasoning

Tue, 07 Apr 2026 00:00:00 +0000

[CVPR26] MotionMaster: Generalizable Text-Driven Motion Generation and Editing

Sat, 28 Mar 2026 00:00:00 +0000

[CVPR26] Scalable Trajectory Generation for Whole-Body Mobile Manipulation

Fri, 27 Mar 2026 00:00:00 +0000

Visual Distraction Undermines Moral Reasoning in Vision-Language Models

Wed, 18 Mar 2026 00:00:00 +0000

OmniClone: Engineering a Robust, All-Rounder Whole-Body Humanoid Teleoperation System

Tue, 17 Mar 2026 00:00:00 +0000

[RA-L26] Simultaneous Tactile-Visual Perception for Learning Multimodal Robot Manipulation

Wed, 04 Mar 2026 00:00:00 +0000

Exhibited at China Int’l Supply Chain Expo (CISCE) 2026, featured on CCTV

LessMimic: Long-Horizon Humanoid Interaction with Unified Distance Field Representations

Tue, 24 Feb 2026 00:00:00 +0000

[ICRA26] Vi-TacMan: Articulated Object Manipulation via Vision and Touch

Sun, 01 Feb 2026 00:00:00 +0000

[ICLR26] Learning physics-grounded 4D dynamics with neural gaussian force fields

Wed, 28 Jan 2026 00:00:01 +0000

[ICLR26] Neural Force Field: Few-shot Learning of Generalized Physical Reasoning

Wed, 28 Jan 2026 00:00:00 +0000

[CHI26] NarrativeLoom: Enhancing Creative Storytelling through Multi-Persona Collaborative Improvisation

Mon, 26 Jan 2026 00:00:00 +0000

[NatureMachineIntelligence26] Proposing and solving olympiad geometry with guided tree search

Sun, 25 Jan 2026 00:00:00 +0000

Cover Article for Nature Machine Intelligence, January 2026

Featured on People’s Daily Front Page

Featured on Guangming

Featured on China Science Daily

UniAct: Unified Motion Generation and Action Streaming for Humanoid Robots

Thu, 01 Jan 2026 00:00:00 +0000

[NatureHumanBehaviour26] Combined evidence from artificial neural networks and human brain-lesion models reveals that language modulates vision in human perception

Sun, 14 Dec 2025 00:00:00 +0000

Featured on Science and Technology Daily

Featured on China Science Daily

Featured on Peking University

[NeurIPS25] Heterogeneous Adversarial Play in Interactive Environments

Mon, 20 Oct 2025 00:00:00 +0000

[NeurIPS25] DrivAerStar: An Industrial-Grade CFD Dataset for Vehicle Aerodynamic Optimization

Sat, 18 Oct 2025 00:00:00 +0000

Cross-Scenario Unified Modeling of User Interests at Billion Scale

Thu, 16 Oct 2025 00:00:00 +0000

[NeurIPS25] GlobalTomo: A global dataset for physics-ML seismic wavefield modeling and FWI

Wed, 01 Oct 2025 00:00:00 +0000

Exhibited at China Int’l Supply Chain Expo (CISCE) 2026

[NeurIPS25] Taccel: Scaling Up Vision-based Tactile Robotics via High-performance GPU Simulation

Sat, 20 Sep 2025 00:00:01 +0000

Featured on GTC Paris

Featured on NVIDIA Tech Blog about Newton Physics

Featured on Linux Foundation Blog about Newton Physics

[T-RO25] Integration of Robot and Scene Kinematics for Sequential Mobile Manipulation Planning

Sun, 31 Aug 2025 00:00:00 +0000

[RA-L25] B*: Efficient and Optimal Base Placement for Fixed-Base Manipulators

Thu, 21 Aug 2025 00:00:00 +0000

[CoRL25] CLONE: Closed-Loop Whole-Body Humanoid Teleoperation for Long-Horizon Tasks

Tue, 19 Aug 2025 00:00:00 +0000

[IROS25] Ag2x2: Robust Agent-Agnostic Visual Representations for Zero-Shot Bimanual Manipulation

Thu, 31 Jul 2025 00:00:00 +0000

Multivariate Long-term Time Series Forecasting with Fourier Neural Filter

Thu, 12 Jun 2025 00:00:00 +0000

[NatureMachineIntelligence25] Embedding high-resolution touch across robotic hands enables adaptive human-like grasping

Mon, 09 Jun 2025 00:00:00 +0000

Science Robotics Editors’ Choice

Featured on People’s Daily

Featured on Global Times

Featured on Guangming Ribao

Featured on Guangming

Interview with Guangming

Featured on Science and Technology Daily

Featured on China Science Daily

Featured on China Youth Daily

Featured on Macau Daily

Featured on Hubei Daily

Featured on Peking University

[CogSci25] Probing and Inducing Combinational Creativity in Vision-Language Models

Thu, 17 Apr 2025 00:00:00 +0000

[CogSci25] Word Embeddings Track Social Group Changes Across 70 Years in China

Wed, 16 Apr 2025 00:00:00 +0000

[CogSci25] A simulation-heuristics dual-process model for intuitive physics

Wed, 09 Apr 2025 00:00:00 +0000

[CVPR25] GROVE: A Generalized Reward for Learning Open-Vocabulary Physical Skill

Thu, 03 Apr 2025 00:00:00 +0000

Featured on Peking University

[CVPR25] Dynamic Motion Blending for Versatile Motion Editing

Wed, 26 Mar 2025 00:00:00 +0000

[T-RO25] Tac-Man: Tactile-Informed Prior-Free Manipulation of Articulated Objects

Wed, 06 Nov 2024 00:00:00 +0000

[RA-L24] MiniTac: An Ultra-Compact 8 mm Vision-Based Tactile Sensor for Enhanced Palpation in Robot-Assisted Minimally Invasive Surgery

Tue, 29 Oct 2024 00:00:00 +0000

[NeurIPS24] PhyRecon: Physically Plausible Neural Scene Reconstruction

Mon, 28 Oct 2024 00:00:00 +0000

[SIGGRAPHAsia24] Autonomous Character-Scene Interaction Synthesis from Text Instruction

Fri, 04 Oct 2024 00:00:00 +0000

[IROS24] Ag2Manip: Learning Novel Manipulation Skills with Agent-Agnostic Visual and Action Representations

Mon, 09 Sep 2024 00:00:00 +0000

[IROS24] PreAfford: Universal Affordance-Based Pre-Grasping for Diverse Objects and Environments

Wed, 21 Aug 2024 00:00:00 +0000

[ECCV24] Zero-Shot Image Feature Consensus with Deep Functional Maps

Sun, 18 Aug 2024 00:00:00 +0000

[CogSci24] Evaluating and Modeling Social Intelligence: A Comparative Study of Human and AI Capabilities

Sat, 01 Jun 2024 00:00:00 +0000

[ScienceAdvances24] Human-level few-shot concept induction through minimax entropy learning

Fri, 19 Apr 2024 00:00:00 +0000

[CVPR24] Move as You Say, Interact as You Can: Language-guided Human Motion Generation with Scene Affordance

Wed, 27 Mar 2024 00:00:00 +0000

[CVPR24] AnySkill: Learning Open-Vocabulary Physical Skill for Interactive Agents

Thu, 21 Mar 2024 00:00:00 +0000

[CVPR24] Scaling Up Dynamic Human-Scene Interaction Modeling

Wed, 13 Mar 2024 00:00:00 +0000

[RA-L24] Grasp Multiple Objects with One Hand

Fri, 08 Mar 2024 00:00:00 +0000

[3DV24] Single-view 3D Scene Reconstruction with High-fidelity Shape and Texture

Tue, 20 Feb 2024 00:00:00 +0000

[ICLR24] I-PHYRE: Interactive Physical Reasoning

Thu, 01 Feb 2024 00:00:03 +0000

[ICLR24] SparseDFF: Sparse-View Feature Distillation for One-Shot Dexterous Manipulation

Thu, 01 Feb 2024 00:00:01 +0000

[ICLR24] Neural-Symbolic Recursive Machine for Systematic Generalization

Thu, 01 Feb 2024 00:00:00 +0000

[NeurIPS23] ProBio: A Protocol-guided Multimodal Dataset for Molecular Biology Lab

Fri, 27 Oct 2023 00:00:00 +0000

[NeurIPS23] Active Reasoning in an Open-World Environment

Wed, 25 Oct 2023 00:00:00 +0000

[NeurIPS23] ChimpACT: A Longitudinal Dataset for Understanding Chimpanzee Behaviors

Mon, 23 Oct 2023 00:00:00 +0000

[NeurIPS23] Evaluating and Inducing Personality in Pre-trained Language Models

Fri, 06 Oct 2023 00:00:00 +0000

[NeurIPS23] Interactive Visual Reasoning under Uncertainty

Thu, 05 Oct 2023 00:00:00 +0000

[ICCV23] Full-Body Articulated Human-Object Interaction

Fri, 18 Aug 2023 00:00:01 +0000

Featured on PKU News

[ICCV23] X-VoE: Measuring eXplanatory Violation of Expectation in Physical Events

Thu, 17 Aug 2023 00:00:00 +0000

Featured on PKU News

[IROS23] Learning a Causal Transition Model for Object Cutting

Sun, 30 Jul 2023 00:00:00 +0000

[IROS23] Part-level Scene Reconstruction Affords Robot Interaction

Tue, 25 Jul 2023 00:00:00 +0000

[IROS23] Sequential Manipulation Planning for Over-actuated Unmanned Aerial Manipulators

Mon, 10 Jul 2023 00:00:00 +0000

[ACL23Demo] PersLEARN: Research Training through the Lens of Perspective Cultivation

Fri, 30 Jun 2023 00:00:00 +0000

[ICML23] MEWL: Few-shot multimodal word learning with referential uncertainty

Sun, 28 May 2023 00:00:00 +0000

Featured on PKU News

[ICML23] On the Complexity of Bayesian Generalization

Sat, 20 May 2023 00:00:00 +0000

[AIR22] Artificial Social Intelligence: A Comparative and Holistic View

Sat, 01 Apr 2023 00:00:00 +0000

Featured on PKU

[CVPR23] Diffusion-based Generation, Optimization, and Planning in 3D Scenes

Thu, 23 Mar 2023 00:00:00 +0000

[ICLR23] Understanding Embodied Reference with Touch-Line Transformer

Tue, 14 Feb 2023 00:00:01 +0000

[ICLR23] A Minimalist Dataset for Systematic Generalization of Perception, Syntax, and Semantics

Tue, 14 Feb 2023 00:00:00 +0000

[ICRA23] GenDexGrasp: Generalizable Dexterous Grasping

Mon, 13 Feb 2023 00:00:01 +0000

[ICRA23] Rearrange Indoor Scenes for Human-Robot Co-Activity

Mon, 13 Feb 2023 00:00:00 +0000

[Engineering23] A Reconfigurable Data Glove for Reconstructing Physical and Virtual Grasps

Tue, 17 Jan 2023 00:00:00 +0000

[NeurIPS22] On the Learning Mechanisms in Physical Reasoning

Mon, 03 Oct 2022 00:00:02 +0000

[NeurIPS22] Emergent Graphical Conventions in a Visual Communication Game

Mon, 03 Oct 2022 00:00:01 +0000

[NeurIPS22] HUMANISE: Language-conditioned Human Motion Generation in 3D Scenes

Mon, 03 Oct 2022 00:00:00 +0000

[IJCV22] Scene Reconstruction with Functional Objects for Robot Autonomy

Tue, 20 Sep 2022 00:00:00 +0000

Featured on PKU

[ScienceRobotics22] In situ bidirectional human-robot value alignment

Sun, 17 Jul 2022 00:00:00 +0000

Featured on Science

Featured on Science Robotics

[ECCV22] Learning Algebraic Representation for Systematic Generalization in Abstract Reasoning

Fri, 15 Jul 2022 00:00:00 +0000

[ECCV22Workshop] PartAfford: Part-level Affordance Discovery from 3D Objects

Thu, 14 Jul 2022 00:00:00 +0000

[IROS22] Sequential Manipulation Planning on Scene Graph

Fri, 01 Jul 2022 00:00:02 +0000

[IROS22] Downwash-aware Control Allocation for Over-actuated UAV Platforms

Fri, 01 Jul 2022 00:00:01 +0000

[RA-L/IROS22] Understanding Physical Effects for Effective Tool-use

Tue, 28 Jun 2022 00:00:00 +0000

[ICML22] Latent Diffusion Energy-Based Model for Interpretable Text Modeling

Sun, 05 Jun 2022 00:00:00 +0000

[JCB2022] Sharing Rewards Undermines Coordinated Hunting

Wed, 01 Jun 2022 00:00:00 +0000

[CogSci22] What Is the Point? A Theory of Mind Model of Relevance

Sun, 01 May 2022 00:00:00 +0000

[RA-L/ICRA22] Object Gathering with a Tethered Robot Duo

Tue, 04 Jan 2022 00:00:00 +0000

[AAIL21] Patching interpretable And-Or-Graph knowledge representation using augmented reality

Wed, 20 Oct 2021 00:00:00 +0000

[NeurIPS21] Unsupervised Foreground Extraction via Deep Region Competition

Fri, 01 Oct 2021 00:00:00 +0000

[RA-L21] Synthesizing Diverse and Physically Stable Grasps with Arbitrary Hand Structures using Differentiable Force Closure Estimator

Sat, 14 Aug 2021 00:00:00 +0000

[ICCV21] YouRefIt: Embodied Reference Understanding with Language and Gesture

Sun, 01 Aug 2021 00:00:01 +0000

[ICCV21] Spatio-temporal Self-Supervised Representation Learning for 3D Point Clouds

Sun, 01 Aug 2021 00:00:00 +0000

[IROS21] Consolidating Kinematic Models to Promote Coordinated Mobile Manipulations

Thu, 01 Jul 2021 00:00:02 +0000

[IROS21] Efficient Task Planning for Mobile Manipulation: a Virtual Kinematic Chain Perspective

Thu, 01 Jul 2021 00:00:01 +0000

[IROS21] Communicative Learning with Natural Gestures for Embodied Navigation Agents with Human-in-the-Scene

Thu, 01 Jul 2021 00:00:00 +0000

[CogSci21] Individual vs. Joint Perception: a Pragmatic Model of Pointing as Communicative Smithian Helping

Tue, 01 Jun 2021 00:00:00 +0000

[ICLR21Workshop] HALMA: Humanlike Abstraction Learning Meets Affordance in Rapid Problem Solving

Fri, 07 May 2021 00:00:00 +0000

[ACL-Findings21] GRICE: A Grammar-based Dataset for Recovering Implicature and Conversational rEasoning

Sat, 01 May 2021 00:00:00 +0000

[CVPR21] Learning Triadic Belief Dynamics in Nonverbal Communication from Videos

Sun, 14 Mar 2021 00:00:04 +0000

[CVPR21] ACRE: Abstract Causal Reasoning Beyond Covariation

Sun, 14 Mar 2021 00:00:03 +0000

[CVPR21] Abstract Spatial-Temporal Reasoning via Probabilistic Abduction and Execution

Sun, 14 Mar 2021 00:00:02 +0000

[ICRA21] Reconstructing Interactive 3D Scenes by Panoptic Mapping and CAD Model Alignments

Sun, 14 Mar 2021 00:00:01 +0000

[ICRA21] Congestion-aware Multi-agent Trajectory Prediction for Collision Avoidance

Sun, 14 Mar 2021 00:00:00 +0000

[IJNME21] Lagrangian‐Eulerian Multi‐Density Topology Optimization with the Material Point Method

Mon, 08 Mar 2021 00:00:00 +0000

[ECCV20] LEMMA: A Multi-view Dataset for Learning Multi-agent Multi-task Activities

Wed, 15 Jul 2020 00:00:00 +0000

[IROS20] Human-Robot Interaction in a Shared Augmented Reality Workspace

Wed, 01 Jul 2020 00:00:01 +0000

[IROS20] Graph-based Hierarchical Knowledge Representation for Robot Task Transfer from Virtual to Physical World

Wed, 01 Jul 2020 00:00:00 +0000

[Engineering20] Dark, Beyond Deep: A Paradigm Shift to Cognitive AI with Humanlike Common Sense

Fri, 01 May 2020 00:00:02 +0000

[SIGGRAPH20] A Massively Parallel and Scalable Multi-GPU Material Point Method

Fri, 01 May 2020 00:00:01 +0000

[SIGGRAPH20] IQ-MPM: An Interface Quadrature Material Point Method for Non-sticky Strongly Two-Way Coupled Nonlinear Solids and Fluids

Fri, 01 May 2020 00:00:00 +0000

[ICRA20] Joint Inference of States, Robot Knowledge, and Human (False-)Beliefs

Sat, 01 Feb 2020 00:00:01 +0000

[ICRA20] Congestion-aware Evacuation Routing using Augmented Reality Devices

Sat, 01 Feb 2020 00:00:00 +0000

[ScienceRobotics19] A tale of two explanations: Enhancing human trust by explaining robot behavior

Wed, 18 Dec 2019 00:00:00 +0000

[AAAI20] Theory-based Causal Transfer: Integrating Instance-level Induction and Abstract-level Structure Learning

Fri, 01 Nov 2019 00:00:01 +0000

[AAAI20] Machine Number Sense: A Dataset of Visual Arithmetic Problems for Abstract and Relational Reasoning

Fri, 01 Nov 2019 00:00:00 +0000

[NeurIPS19] Learning Perceptual Inference by Contrasting

Sun, 01 Sep 2019 00:00:00 +0000

[NeurIPS19] PerspectiveNet: 3D Object Detection from a Single RGB Image via Perspective Points

Sun, 01 Sep 2019 00:00:00 +0000

[ICCV19] Holistic++ Scene Understanding: Single-view 3D Holistic Scene Parsing and Human Pose Estimation with Human-Object Interaction and Physical Commonsense

Thu, 01 Aug 2019 00:00:00 +0000

[IROS19] Learning Virtual Grasp with Failed Demonstrations via Bayesian Inverse Reinforcement Learning

Sat, 01 Jun 2019 00:00:00 +0000

[CogSci19] Decomposing Human Causal Learning: Bottom-up Associative Learning and Top-down Schema Reasoning

Mon, 01 Apr 2019 00:00:00 +0000

[CVPR19] RAVEN: A Dataset for Relational and Analogical Visual Reasoning

Fri, 01 Mar 2019 00:00:00 +0000

[TURC19] VRGym: A Virtual Testbed for Physical and Interactive AI

Fri, 01 Mar 2019 00:00:00 +0000

[ICRA19] High-Fidelity Grasping in Virtual Reality using a Glove-based System

Fri, 01 Feb 2019 00:00:00 +0000

[ICRA19] Self-Supervised Incremental Learning for Sound Source Localization in Complex Indoor Environment

Fri, 01 Feb 2019 00:00:00 +0000

[AAAI19] MetaStyle: Three-Way Trade-Off Among Speed, Flexibility and Quality in Neural Style Transfer

Thu, 01 Nov 2018 00:00:00 +0000

[AAAI19] Mirroring without Overimitation: Learning Functionally Equivalent Manipulation Actions

Thu, 01 Nov 2018 00:00:00 +0000

[NeurIPS18] Cooperative Holistic Scene Understanding: Unifying 3D Object, Layout, and Camera Pose Estimation

Sat, 01 Sep 2018 00:00:00 +0000

[ECCV18] Holistic 3D Scene Parsing and Reconstruction from a Single RGB Image

Sun, 01 Jul 2018 00:00:00 +0000

[IJCV18] Configurable 3D Scene Synthesis and 2D Image Rendering with Per-Pixel Ground Truth using Stochastic Grammars

Tue, 01 May 2018 00:00:00 +0000

[CogSci18] Human Causal Transfer: Challenges for Deep Reinforcement Learning

Sun, 01 Apr 2018 00:00:00 +0000

[CVPR18] Human-centric Indoor Scene Synthesis Using Stochastic Grammar

Thu, 01 Mar 2018 00:00:00 +0000

[SIGGRAPH18] A Moving Least Squares Material Point Method with Displacement Discontinuity and Two-Way Rigid Body Coupling

Thu, 01 Mar 2018 00:00:00 +0000

[ICRA18] Interactive Robot Knowledge Patching using Augmented Reality

Thu, 01 Feb 2018 00:00:00 +0000

[ICRA18] Unsupervised Learning of Hierarchical Models for Hand-Object Interactions

Thu, 01 Feb 2018 00:00:00 +0000

[AAAI18] Tracking Occluded Objects and Recovering Incomplete Trajectories by Reasoning about Containment Relations and Human Actions

Wed, 01 Nov 2017 00:00:00 +0000

[IROS17] A Glove-based System for Studying Hand-Object Manipulation via Joint Pose and Force Sensing

Sat, 01 Jul 2017 00:00:00 +0000

[IROS17] Feeling the Force: Integrating Force and Pose for Fluent Discovery through Imitation Learning to Open Medicine Bottles

Sat, 01 Jul 2017 00:00:00 +0000

[CogSci17] Consistent Probabilistic Simulation Underlying Human Judgment in Substance Dynamics

Sat, 01 Apr 2017 00:00:00 +0000

[TVCG17] The Martian: Examining Human Physical Judgments Across Virtual Gravity Fields

Thu, 01 Sep 2016 00:00:00 +0000

[SIGGRAPHAsia16Workshop] A Virtual Reality Platform for Dynamic Human-Scene Interaction

Fri, 01 Apr 2016 00:00:00 +0000

[IJCAI16] What is Where: Inferring Containment Relations from Videos

Tue, 01 Mar 2016 00:00:00 +0000

[CVPR16] Inferring Forces and Learning Human Utilities From Videos

Mon, 01 Feb 2016 00:00:00 +0000

[CogSci16] Probabilistic Simulation Predicts Human Performance on Viscous Fluid-Pouring Problem

Fri, 01 Jan 2016 00:00:00 +0000

[CVPR15] Understanding Tools: Task-Oriented Object Modeling, Learning and Recognition

Sun, 01 Feb 2015 00:00:00 +0000

[CogSci15] Evaluating Human Cognition of Containing Relations with Physical Simulation

Thu, 01 Jan 2015 00:00:00 +0000

Dr. Android and Mr. Hide: Fine-grained security policies on unmodified Android

Thu, 01 Dec 2011 00:00:00 +0000