Gal Chechik

Metatron from @NVIDIAAI is one frozen whole-body policy for the Unitree G1 that takes partial goals instead of a full reference motion: hand or foot targets with arrival times, a velocity, a heading. Policy • Built on MaskedMimic: RL trains dense motion trackers, then on-policy distillation turns them into one generative policy that accepts masked, approximate goals • On the G1 it distills four trackers: BONES and AMASS+LAFAN motion capture, synthetic recovery, and scene interaction (SAMP) Skills • A task is ordinary code that sets goals at each control step; the policy fills in the rest of the body • Coding agents write and refine that code in simulation; the policy is never retrained Results • Whole-body reaching, 54 targets in MuJoCo: 83% reached vs 39% for a diffusion generator (ARDY) feeding the SONIC tracker • Pushes while walking at 1 m/s: 75% stay upright at 800 and 1000 N, both SONIC pipelines 0% • One checkpoint on the real G1 for walking, recovery, teleoperated pickup and sitting on furniture The surprising bit • A better gait came from better goals, not retraining: Claude Code fitted a foot-trajectory rule from motion-capture clips, and its timed foot targets gave more decisive steps on the real G1 (judged qualitatively) By @ChenTessler, Yifeng Jiang, @AlecSaysThings, @YiShi_333, @UmarIqb, @GalChechik and @xbpeng4. Project: research.nvidia.com/labs/par…