ETH RSL's ACP humanoid: reference-free contact planning runs onboard at 50Hz
ChongZzZhang · x · 2026-11-06
Researchers from ETH Zurich RSL (with Apple and NVIDIA collaborators) validated ACP (Humanoid Mobility via Attention-based Contacts) on real humanoids — the author calls it the strongest reference-free pipeline to date.
- Just 13.2 minutes of human motion data teach the robot where 9 body parts (hands, feet, knees, elbows, pelvis) should contact the world; RL fills in the rest
- At deployment no reference motion is used: two networks run onboard at 50Hz, planning contacts from LiDAR height maps within a 1.5s horizon and controlling 23 joints
- Sim results include collision avoidance on sparse stepping stones, foot-only climbing of 70cm obstacles, and 1.2m gap traversal
- Builds on AME-1 (Science Robotics 2025) / AME-2; GitHub update coming soon
More from Embodied
- OpenTac Brings HDR Readout to Open Tactile Sensors, Sensing Light and Heavy Touch at Once — _krishna_murthy · 2026-11-06
- Huawei watch with built-in blood pressure check gets tested in flagship store — EleanorOlcott · 2026-11-06
- 'Software is dead': Rasmic pivots to hardware by building his own AI speaker with AI — Rasmic · 2026-11-06
- Aichi Institute of Tech's humanoid robot finishes last in ekiden relay but crosses the line — 4310sy · 2026-11-06
- IEEE T-Haptics calls papers on touch in contact-rich dexterous manipulation, due March 2027 — HaozhiQ · 2026-11-06
- NVIDIA USDCode MCP Server brings expert Isaac Sim and USD scripting help to any MCP client — modelcontextprotocol · 2026-11-06