DexWM
Meta FAIR / New York University
An action-conditioned model that predicts future latent representations for dexterous hand–object interaction.
What you can explore
Inspect human-video and robot-data training, hand-consistency objectives and the released rollout evaluation.
Before you use it
The current repository documents RoboCasa simulation evaluation. Treat paper-reported robot results separately from the code available to reproduce them; a new hand needs its own mapping and validation.
Original contributors
Raktim Gautam Goswami, Amir Bar, David Fan and collaborators
License / access: CC BY-NC 4.0, per repository; check data and dependencies separately
Official implementation & setup
Editorial reference · checked 2026-09-22