What it is
A cross-embodiment egocentric world model for loco-manipulation, trained on 10k human and 3k robot hours with unified camera-pose and skeleton conditioning. Instead of embodiment-specific action interfaces, it conditions video prediction on a shared visual kinematic interface: camera poses for locomotion plus projected skeletons for manipulation, letting one model transfer across humans, grippers, and humanoids.
Key results
- SOTA across DreamDojo, Ψ0, ω-HOME, Xp-Teleop, and AgiBot
- Human scaling: camera ATE drops 2.21→1.79 cm from 1k→10k h pretraining
- Robot scaling on Ψ0: LPIPS improves 0.427→0.365 from 1k→10k h
Large-scale human egocentric pretraining transfers to robot world modeling when locomotion and manipulation are expressed in a shared kinematic format.