Unitree G1 learns self-rescue via ROOF RL strategy using geometric rewards, no maps
Scobleizer · x · 2026-11-03
Chinese researchers developed ROOF, a reinforcement learning strategy that teaches the Unitree G1 humanoid to rescue itself when trapped under obstacles.
- No cameras or pre-built 3D maps: a geometric reward guides training — less overlap with obstacles and greater distance created means more reward
- The policy isn't told how or which direction to move; G1 must figure that out itself
- The robot learns to shift its body, use limbs for stable support, push movable obstacles aside, and stand up
Robert Scoble notes that Physical AI demos keep getting better.
More from Embodied
- China replicates robot box-folding demo 3 months later with much cheaper hardware — Scobleizer · 2026-11-03
- Same policy can fail outside its training engine: Manda tests sim-to-sim transfer across MuJoCo, PhysX, Newton and Genesis — facontidavide · 2026-11-03
- SO-101 robot arm learns ping pong, best rally hits 23 bounces — Scobleizer · 2026-11-03
- Investor predicts a $2-3k clothes-folding home robot within two years — lukas_m_ziegler · 2026-11-03
- NVIDIA shows frontier AI agents driving Omniverse simulation workflows — CurieuxExplorer · 2026-11-03
- BiGym 2.0: Claude-written policies hit 51.9% on humanoid household tasks, nearing VLA — teortaxesTex · 2026-11-03