DexGPT: a GPT-written real2sim pipeline turns one monocular hand GIF into 44-DOF dexterous sim, open-sourced

animesh_garg · x · 2026-09-10

DexGPT shows GPT writing an entire real2sim pipeline from a single monocular GIF of a human hand: hand tracking, IK retargeting to two 22-DOF Sharpa Wave hands, and grasp refinement, with zero human code. Open-sourced, it achieves passive hinge RMSE of 5.18° (<10° target), max penetration 5.62mm (<5mm target), and 99.5% hand contact frames over 203 replayed states in MuJoCo — though the author flags it as experimental and not yet physically validated.

Original post →

More from Embodied

Embodied channel →