Odyssey-3: One Pretrained World Model Adapts to Different Robot Arms With Hours of Demos

ChongZzZhang · x · 2026-09-20

Odyssey-3 demonstrates a single pretrained world model that learns the world, conditions it for different embodiments, and acts. With only tens of hours of robot demonstrations it adapts to different robotic arms and tasks — generalizing rather than memorizing trajectories — and can generate environments for agents to inhabit and learn in. Commenters flagged visual glitches at the 10-second mark, but the approach is drawing wide attention.

Original post →

More from Embodied

Embodied channel →