DexAgent learns bimanual dexterous manipulation from one human video, hitting 63.6% success
YunzhuLiYZ · x · 2026-09-30
Researchers introduce DexAgent, an agentic Human2Sim2Robot framework that learns bimanual dexterous manipulation from a single human video, built around a self-evolving skill library.
- Evaluated on 11 real-world long-horizon tasks (rigid, articulated, deformable objects): 63.6% policy rollout success rate vs. 18.2% for the strongest baseline.
- Average inference time of 2.1 hours with the tool library, down from 3.3 hours for the baseline.
- The library currently holds 103 skills and 188 verifiers and keeps growing as new tasks and objects appear.
The paper and project page are public, and the team welcomes community contributions to the tool library.
More from Embodied
- VLANeXt Family: 500+ controlled experiments distill 12 practical recipes for VLA models — ccloy · 2026-09-30
- Simple-WAM: One forward pass replaces video denoising, keeping world action models' generalization — Tsinghua-LeapLab · 2026-09-30
- Open-source humanoid Roboparty shows up at IROS, alongside a Gundam you can touch — 4310sy · 2026-09-30
- Scoble says Google Glass is coming back: 'We were early, not wrong' — Scobleizer · 2026-09-30
- DexTaG uses human tactile demos as RL reward guidance to teach robots human-like tool use — gan_chuang · 2026-09-30
- Wayve's supervised robotaxi ride in London via Uber impresses, closing in on Waymo — Sethwinterroth · 2026-09-30