Research
EgoX: Egocentric Cross-Embodiment Manipulation with Embodiment Dreaming
EgoX is an egocentric cross-embodiment learning framework designed to combine demonstrations from humans, real robots, and simulated agents despite differences in morphology, sensing, and control. It pretrains a Modularized Cross-Embodiment Transformer with Embodiment Dreaming (MXT+) using modality-specific tokenizers and action experts over a shared Transformer trunk, then fine-tunes on target-embodiment demonstrations. A training-only dreaming objective predicts latent future visual and proprioceptive states. Across 13 task-embodiment pairs spanning five tasks and six embodiments, cross-embodiment pretraining raised average normalized task progress from 0.63 to 0.80, and from 0.53 to 0.77 under out-of-distribution conditions.
EgoXcross-embodiment learningrobot manipulation