Physical world AI struggles with unstructured data

Physical AI is getting smaller, smarter, and more accessible. Meet @Microduck a tiny $399 open-source robot you can teach new tricks with reinforcement learning. 25 cm tall, 15 degrees of freedom, with vision, LiDAR, IMUs, and an articulated beak. It walks, grasps objects, recovers from falls, and learns new behaviors. The democratization of physical AI is here: https://lnkd.in/eKnPjkQ8

Hugging Face unveiled the Microduck, a twenty five centimeter tall duck robot that can waddle, pick things up with its beak, and get back up when it falls. It even roller skates. Here's the part that matters if you build with AI: it is not a black box. The robot perceives the world with a camera and lidar, and you train its behaviors in simulation, then deploy them straight onto the robot. The SDK, the simulation and the full reinforcement learning stack are on GitHub. Clem Delangue put it plainly: this is the era of open source, affordable robots, built to democratize physical AI and world models. The same open-source wave that made models ownable is reaching hardware. You can audit it and retrain it yourself. Source: https://lnkd.in/eSnurBSH

Everyone talks about robot training data. Far fewer people talk about what happens the moment a machine ships. A deployed machine throws off exhaust constantly: joint states, video, lidar, odometry, and the record of what it tried and how it failed. Training data is scarce and has to be manufactured. Operational data is the opposite problem, a firehose that never stops. The learning loop everyone is counting on runs entirely on it. Collect the exhaust, replay the failures, annotate them, feed the lessons back into the next model. Here is where it breaks. The live view that flags a robot misbehaving now, the replay you scrub after it fails, and the training set you assemble months later all want the same telemetry. Most stacks keep three copies fighting to stay in sync. That is how the annotation explaining a failure so often never reaches the next model's training set. Most teams have not hit this yet because most are not running large fleets. The ones treating telemetry as core infrastructure early are the ones not rebuilding it under load later. https://tsdb.co/c3po-li #Robotics #AI #DataInfrastructure #TimeSeries

One Sensor Sees. Two Sensors Understand. 👁️📡 A camera provides rich visual semantics. LiDAR provides accurate depth and 3D geometry. When these signals are aligned and fused, Computer Vision systems can build a more complete understanding of the scene. 📷 Camera + LiDAR ↓ 🔗 Feature Alignment & Fusion ↓ 🧠 Multimodal Model ↓ 🎯 Richer 3D Perception This combination can help identify objects while also understanding their position, shape, and spatial relationships. At Visual Grab, we work on Computer Vision and Multimodal AI solutions that transform sensor data into reliable visual intelligence. Visual Grab — Turning Vision Into Impact. #VisualGrab #ComputerVision #CameraLiDAR #SensorFusion #MultimodalAI #ComputerVisionAI #3DVision #AI #AutonomousDriving

Hugging Face has introduced the Microduck, a $399 open-source duck robot. The Microduck can perform various actions like waddling, picking up objects, and roller skating, and is designed to be trained with reinforcement learning. Clem Delangue 🤗, CEO of Hugging Face, said the robot aims to democratize physical AI and world models. The robot uses a camera, lidar sensors, and IMUs for perception, and its behaviors can be trained in simulation and deployed on the robot. Read more: https://lnkd.in/gcUF94Cw 📰 Subscribe to the weekly Robotics Brief Newsletter: https://lnkd.in/eGiCpMzH #ai #artificialintelligence #ainews

#Logiicdev is transforming geospatial intelligence with Physical AI! 🌍🚀 The future of autonomous systems, surveying, and smart cities demands more than just 'smart' AI. It requires AI that is #Repeatable, #Deterministic, and #Provable. For industries relying on geospatial data, confidence in results is non-negotiable. Whether it's precision infrastructure mapping, certified drone navigation, or safety-critical robotics, the "black box" approach isn't enough. We are excited to join the global geospatial community at INTERGEO Expo and Conference (dvw.de/intergeo/de) to discuss how #Logiicdev bridges the gap between deep-tech innovation and field reliability. Talk to us and see live demonstrations of our Physical AI solutions, including: i.Verifiable Real-Time #LiDAR Processing: Dynamic mapping with mathematical guarantees. ii.Edge-Intelligence Sensor Fusion: Combining visual, inertial, and spatial data with provable outcomes. iii.Certifiable Decision Logic: Deterministic AI cores on low-power, high-performance hardware (FPGA/RISC-V). Our expert team is ready to discuss how Physical AI can solve your critical reliability challenges and unlock the full potential of your geospatial applications. Let's make AI trustworthy together! #INTERGEO2026 #PhysicalAI #Geospatial #Logiicdev #EdgeAI #FPGA #DeterministicAI #AISafety #CertifiableAI #Robotics #EUTech

Tiny drone-vs-bird video detection is a real, recognized research problem, and I've been building a classical, track-centric pipeline for it. This is the intro post for a series where I'll keep sharing engineering deep-dives as I go: speed/accuracy trade-offs, false-positive reduction, tooling, and whatever else comes up. The core challenge: separating drones, birds, and outright false positives in EO video where targets are often just 3-10 pixels. This is exactly the problem tackled by recurring benchmarks like the Drone-vs-Bird Detection Challenge (WOSDETC) and Anti-UAV. Rather than jump straight to deep learning, my architecture is deliberately classical: MOG2 background subtraction → morphology → blob extraction → Kalman-style tracking → engineered motion/temporal/geometry features → XGBoost classification. The literature backs this up: classical multi-feature methods like SuBSENSE and PAWCS remain the strongest unsupervised, cross-scene performers, ahead of scene-supervised CNNs that don't generalize well to cameras they weren't trained on. Key takeaways: • The task is grounded in recognized benchmarks: the Drone-vs-Bird Detection Challenge (WOSDETC) and Anti-UAV. • Each pipeline stage is validated against the published literature. • Near-term roadmap: GSOC as the easiest drop-in background-subtraction upgrade, with heavier options (SuBSENSE/PAWCS) held in reserve if needed. One preview of what the series digs into: tightening the candidate/track limits in this pipeline cut worst-case frame time from 25-36 seconds down to 0.081 seconds while holding recall at 98.4%. I'll cover the trade-off behind that number in an upcoming post. I'll be linking back to this post as the series grows. How have you approached tiny, high-FP detection tasks with limited data and domain shift? #MachineLearning #ComputerVision #ObjectDetection #EdgeAI #DroneDetection

this humanoid sees the monkey bars, jumps onto them, swings across, and lands. “[30 Aug 2026] Learning Agile Perceptive Traversal of Sparse 3D Structures for Humanoids” the hard part isn’t just learning to swing. the robot first has to see extremely thin bars above its head, decide where they are while moving, jump high enough to catch one, keep tracking the structure while swinging, then safely jump back down. ETH’s system does this directly from a head-mounted solid-state lidar. the policy uses memory because those thin bars only produce sparse sensor returns, and the robot uses passive hooks on its hands to make the physical catches more reliable. on real hardware, it completes the full sequence in 14 of 15 trials across 3 different bar setups, reaching brachiation speeds up to 0.5 m/s. the same perception system can also duck beneath overhead obstacles only 2 × 2 cm thick. humanoid locomotion is starting to become more than walking around obstacles. the robot is learning to use the 3D world itself as part of the route. https://lnkd.in/dZwfk6PP

BEV Annotation vs Key Point Annotation — Part 2 🚗📍 In ADAS/AV data annotation, two techniques drive very different capabilities in perception models. Here's how they compare: 🔹 BEV (Bird's Eye View) Annotation Top-down mapping of vehicles, pedestrians, cyclists, lanes, road boundaries & drivable area. Built on multi-sensor fusion (camera + LiDAR + radar), calibration, and ground-plane projection. Powers models like BEVFormer, BEVDet, CenterPoint, MapTR & TransFusion. Strength: global spatial understanding — critical for trajectory planning, HD maps & 360° scene understanding. 🔹 Key Point Annotation Precise placement of human/object key points — head, shoulder, elbow, hip, knee, ankle — from image/video. Powers pose models like HRNet, OpenPose, MediaPipe & DeepPose. Strength: structure & movement — critical for pose estimation, gesture recognition, driver monitoring & action recognition. ⚙️ Both follow the same quality backbone: taxonomy & SOP adherence, occlusion/truncation handling, peer review, QA against golden datasets, and consistency across frames. 📊 The bigger picture: neither exists in isolation. Data collection → cleaning → annotation → QA/gold sets → train/val/test split → fine-tuning (augmentation, class balancing, domain adaptation, hard-example mining) → evaluation → retraining → deployment. 🤔 Which to choose? BEV → global spatial context, localization, multi-sensor fusion Key Point → precise structure, pose & behavior analysis Both (recommended) → richer perception, better safety & performance 💡 Key takeaway: BEV gives you the "WHERE" (spatial context). Key Point gives you the "WHAT" (structure & movement). Together, they build a more complete understanding of the driving environment. Better Annotation → Better Data → Better Models → Better Perception → Safer Roads 🛣️ #ADAS #AutonomousVehicles #DataAnnotation #ComputerVision #BEV #PoseEstimation #AI #MachineLearning #GIS #LiDAR

what if the “ground truth” used to train your perception model is already wrong? this Omnigrid project tried to build 3D perception from 6 fisheye cameras + LiDAR on a drone. the plan sounded reasonable: project LiDAR points into each camera → use them as depth ground truth → train the perception model except the projected points didn’t line up properly. they tried PyTorch3D. then OpenCV’s fisheye model. then implemented the omnidirectional projection model themselves. still wrong. so they looked at the projection Jacobian. most camera parameters had fairly modest sensitivity. some distortion parameters were orders of magnitude more sensitive. tiny calibration errors could move projected LiDAR points dramatically across the image. and that became the real bottleneck. not the neural network. not the loss function. the calibration used to create the labels. they eventually couldn’t trust the LiDAR projection enough to use it as clean depth ground truth. I love posts like this because they show what perception engineering actually looks like. before debugging the model trained on your “ground truth,” debug how the ground truth was created. https://lnkd.in/dQk-dCn6