DexRoam turns egocentric human demos into robot skills, replacing 50% of teleop data
机器之心 · wechat · 2026-11-06
A team from HKUST, BAAI and Peking University presents DexRoam (accepted to CoRL 2026, code open-sourced): a full pipeline from egocentric human demonstration capture to whole-body alignment and VLA policy training for mobile bimanual dexterous manipulation.
Method: capture needs only a Meta Quest 3 headset plus ZED Mini stereo camera — no external mocap — recording egocentric video, whole-body joints and hand keypoints. Alignment runs in three stages: body alignment (pelvic-reference mapping), action-semantic alignment (relative action representations removing height/arm-length differences), and temporal alignment (resampling by task progress).
Results: MMD/SWD distribution gaps drop 41.7%/31.1% on average, and aligned human trajectories replay successfully on hardware. Across five tasks with GR00T N1.7 and π₀.₅, "human pretraining + joint finetuning" works best: GR00T N1.7 success 29%→56%, π₀.₅ 32%→57%. Halving robot demo data still matches the full baseline — aligned human data substitutes 50% of costly teleop demos, and even introduces grasping strategies absent from robot data.
More from Embodied
- Galaxea Co-founder Hang Zhao's CoRL Keynote: Split Robot Intelligence into System 2/1/0 by Timescale — YuXiang_IRVL · 2026-11-13
- Apparent OpenAI hardware unboxing photo leaks: robot arm plus handheld console — yungcontent · 2026-11-13
- FINGR dexterous hand solves a 2x2 Rubik's Cube with continuous finger tricks, wins Best Paper at CoRL 2026 workshop — Scobleizer · 2026-11-13
- Fei-Fei Li amplifies new Atlas demo: 'It is very hard to explain how cool Atlas is' — drfeifei · 2026-11-13
- GXO to deploy up to 1,000 Apptronik Apollo humanoid robots in warehouses over four years — CyberRobooo · 2026-11-13
- Waymo robotaxi spotted quietly obeying a 'no right turn on red' sign humans often miss — reed · 2026-11-13