A rollout can look right for the wrong reason.
Action-conditioned world models can generate visually plausible rollouts while ignoring the supplied action trajectory, reverting to task-typical behavior, moving the wrong object, or hallucinating unsupported physical interactions.
WorldSimProbe evaluates the causal chain from an action intervention to the robot motion and environment response realized in the generated video. Its five diagnostic tasks cover local action calibration, global trajectory coverage, action-source behavior preservation, interaction grounding, and interaction dynamics across RoboTwin, ManiSkill, and LIBERO.
Five probes reveal where the action-to-rollout causal chain fails.
Explore interactive probes, aligned rollouts, and failure cases on the project page.
2026-08-20: 🤗 Released the WorldSimProbe validation set, including reference videos for internal evaluator testing, on Hugging Face.2026-08-10: 🤗 Released the public RoboTwin,Maniskill and LIBERO test sets on Hugging Face.2026-08-06: 🔥 Released the complete Task 1-5 evaluation code and submission toolkit.
- Release the Task 1-5 evaluation code and submission toolkit.
- Release the public RoboTwin and LIBERO test sets.
- Release the validation set with reference videos for internal evaluator testing.
- Launch the public leaderboard (ETA: 1 week).
- Release the generation pipeline (ETA: 1-2 months).
Join the WorldSimProbe WeChat group for benchmark updates, submission questions, and community discussion.
For benchmark, submission, or evaluation questions, email worldsimprobe@outlook.com.
WeChat invitation QR codes are time-limited; this image will be refreshed when a new code is issued.
| Task | Diagnostic question | Evaluation |
|---|---|---|
| 1. Local Action Calibration | Does increasing an action perturbation produce the expected change in the rollout? | Simulator-calibrated response ratio |
| 2. Global Trajectory Coverage | Can the model realize a donor action outside the receiver task's typical behavior? | RobotSeg-masked reference-flow similarity |
| 3. Action-Source Behavior Preservation | Does the model preserve behavior from expert, policy, and human control sources? | RobotSeg-masked flow with source-level reporting |
| 4. Interaction Grounding | Does the rollout avoid hallucinating object interaction when the commanded action should not produce contact? | TAPNext++ object tracking with robot-motion verification |
| 5. Interaction Dynamics | Does the rollout realize the intended physical interaction primitive? | Frozen VLM primitive classification |
See tasks.md and metrics.md for the complete task definitions and scoring protocols.
WorldSimProbe requires Python 3.10 or newer. For participants who only need to validate and package submissions:
git clone https://github.com/pxxq25/WorldSimProbe.git
cd WorldSimProbe
python -m venv .venv
source .venv/bin/activate
python -m pip install -e .Leaderboard workers can install the additional Python dependencies with:
python -m pip install -e ".[flow,task4,task5]"This command installs the packaged Python dependencies; it does not download
the evaluator models. Tasks 2-4 additionally require an NVIDIA CUDA worker,
the pinned RobotSeg and TAPNext++ sources, and their checkpoints. Task 5 uses
Qwen/Qwen3-VL-8B-Instruct. Follow
evaluator_setup.md for the complete worker setup,
asset layout, and smoke test.
Install the operator-console dependencies only when collecting real Task 3 human-teleoperation traces:
python -m pip install -e ".[teleoperation]"Validate a completed submission before upload:
worldsimprobe validate-submission \
--manifest submission/submission.jsonl \
--root submission \
--decodePackage the validated submission:
worldsimprobe package-submission \
--manifest submission/submission.jsonl \
--root submission \
--output worldsimprobe_submission.zipDownload the public inference inputs from the WorldSimProbe dataset on Hugging Face. Each released backend package contains its public manifest, initial-context images, action trajectories, schema, and generation contract. Run the model on every sample and prepare:
submission/
├── submission.jsonl
└── videos/
├── <sample_id>.mp4
└── ...
Each sample contains exactly one prediction. Task 1 requires original,
small, and large videos; Tasks 2-5 require one candidate video. Every
video must cover the requested physical-time horizon, subject only to
one-frame timestamp rounding.
Submitted videos use evaluator-owned timing, which defaults to 10 FPS.
Participant manifests cannot override decoded video timing. See
submission.md and video_format.md
for the complete JSONL and video contracts. Files under examples/ are format
examples only, not public evaluation samples.
After validation, package the submission as worldsimprobe_submission.zip,
upload it to Google Drive, and enable viewer access for anyone with the link.
Email the shareable Google Drive link to
worldsimprobe@outlook.com.
The evaluation maintainers validate each submission, join it with a private
reference manifest, and run the task-specific evaluator under
worldsimprobe.evaluation.
The public implementation exposes the metric logic; hidden rows provide the
simulator references, actions, object metadata, task labels, and opaque sample
identities required for scoring.
Task 5 uses a frozen VLM prompt and a shared 12-frame physical-time sampling protocol. Model scores do not use an additional GT-oracle filter.
The Task 3 operator console, RoboTwin adapter, and trace-integrity gate are documented in task3_teleoperation.md. Synthetic human-like action profiles are not accepted as human teleoperation.
WorldSimProbe currently evaluates videos derived from:
These upstream projects are linked for compatible training-data and simulator development. WorldSimProbe submissions contain generated videos, not simulator installations. Additional setup notes are available in simulator_resources.md.
🔐 Public Release and Hidden References
This public repository contains:
- the Task 1-5 evaluation implementations;
- the video submission validator and packaging tools;
- small synthetic examples demonstrating the public formats;
- the Task 3 interface for collecting real human-control traces;
- documentation for reproducing the public evaluation protocol.
Initial context frames, instructions, action trajectories, public timing, and prediction horizons are distributed separately with the public evaluation inputs. The corresponding reference rollouts and evaluator-only metadata remain hidden.
For local evaluator and integration testing, we separately release the WorldSimProbe validation set, which includes its reference videos. These validation references are distinct from the hidden references used to score the public test set.
It intentionally excludes:
- hidden reference videos and future ground-truth rollout frames;
- evaluator-only simulator states, contacts, object annotations, and task labels;
- internal provenance and non-public identifiers;
- model checkpoints and generated benchmark predictions;
- absolute paths from internal machines.
Submissions are joined with hidden references only on the official evaluator.
scripts/check_public_release.py enforces these release constraints.
WorldSimProbe/
├── configs/evaluation/ # Frozen task protocols
├── src/worldsimprobe/
│ ├── evaluation/ # Task 1-5 evaluators
│ ├── submission/ # Validation and packaging
│ └── common/ # Shared timing and metric utilities
├── scripts/ # Evaluator and utility entry points
├── schemas/ # Submission and result schemas
├── examples/ # Synthetic format examples
├── docs/ # Detailed protocols
└── tests/ # Contract and metric tests
If you use WorldSimProbe, please cite the accompanying paper available from the project page. Citation metadata is provided in CITATION.cff.
WorldSimProbe is released under the MIT License.