ETH Zurich robot learns humanoid mobility from 13 minutes of human motion, no reference at deployment
ChongZzZhang · x · 2026-10-23
Researchers from ETH Zurich's RSL, with Apple and NVIDIA collaborators, unveiled ACP (Humanoid Mobility via Attention-based Contacts), pushing humanoid mobility further.
- Just 13.2 minutes of human motion data teach the robot where its hands, feet, knees, elbows and pelvis should contact the world; RL fills in the rest of the body.
- At deployment no reference motion is needed: two onboard networks run at 50 Hz — a contact planner predicts contacts for 9 body parts over a 1.5s horizon from LiDAR height maps and proprioception, and a whole-body controller drives 23 joints. One policy covers every skill.
- Authors include Junzhe He, Chong Zhang, Marco Hutter and Vladlen Koltun; preprint is out, code and demos to follow.
Related event: ETH Zurich Trains Humanoid Locomotion with Just 13 Minutes of Human Data(2 posts)→
More from Embodied
- Physical Agent Capability Index: Astra leads Argon 1349 to 1324 Elo, both fail 4-bottle task — qinzytech · 2026-10-23
- Beijing's service robot output surges 170% in first three quarters — yogthos · 2026-10-23
- NVIDIA investigates stuttering issues reported on Surface Laptop Ultra — tomwarren · 2026-10-23
- Ex-OpenAI science head Kevin Weil: US must use AI and robotics to win back pharma before it's too late — beffjezos · 2026-10-23
- 54-second SLAM kitchen scan, Gaussian splats processed in under 9 minutes — jonstephens85 · 2026-10-23
- Builder wires StackChan desktop robot to Nous Research's Hermes-Agent, plans open-source release — alexcovo_eth · 2026-10-23