ETH Zurich humanoid walks any terrain with one policy, trained on just 13.2 minutes of motion data
ChongZzZhang · x · 2026-10-23
ETH Zurich's Robotic Systems Lab (with NUS, Apple, and NVIDIA) presents "Humanoid Mobility via Attention-based Contacts": just 13.2 minutes of human motion data teach the robot where its hands, feet, knees, elbows, and pelvis should make contact with the world; reinforcement learning fills in the rest.
Key deployment design:
- No reference motions, no skill-switching mechanism, no manual data augmentation — one policy for every skill
- Two networks run onboard at 50 Hz: a contact planner predicts contact locations for nine body parts within a 1.5 s horizon from LiDAR height maps, proprioception, and a goal; a whole-body controller outputs actions for 23 joints
- Click a goal and the robot traverses diverse challenging terrains
The team's prior AME work appeared in Science Robotics and IEEE T-RO.
Related event: ETH Zurich Trains Humanoid Locomotion with Just 13 Minutes of Human Data(2 posts)→
More from Embodied
- NVIDIA launches Isaac Lab 3.0 with modular architecture, Newton physics and RL post-training — npew · 2026-10-23
- First walk with a Hypershell exoskeleton: most helpful on stairs — DominiqueCAPaul · 2026-10-23
- Robot arm Podrick learns to load utensils by distilling lessons from its own failed runs — ycombinator · 2026-10-23
- Kazakhstan scales AI compute: 1,000+ GPUs next year plus Vera Rubin deployments — NVIDIA Blog · 2026-10-23
- Robotics data is far from saturated: labs are hedge funds hunting for data alpha — paigeinsf · 2026-10-23
- How to evaluate robotics data startups: quality annotation and real-world evals win — paigeinsf · 2026-10-23