Odyssey announced Odyssey-3, a foundation world model for controlling physical and virtual systems. The same model was used to drive robot arms, humanoids, a car, a drone and video games. Odyssey says each task needed only a few hours of paired observation and action data. In robot experiments it recovered from missed grasps and oddly placed objects not shown in training. Flexion built humanoid control policies on top of it and reports better tolerance of lighting changes. A driving policy trained only in simulation drove on Indian streets in closed loop. Policies trained on Grand Theft Auto V produced movement in two other games without further training. Odyssey also announced an evaluation collaboration with Poke & Wiggle on transfer across robot bodies. Odyssey says world models remain roughly two orders of magnitude smaller than language models. It plans a public release in the coming weeks.
What changed
Robot and vehicle control systems were each trained for one narrow task on large task-specific datasets.
What it unlocks
Training a control policy for a robot arm, humanoid, car, drone or game from only a few hours of task data on one shared pretrained model.
- 20 hours of simulated driving data
- 77% as far between interventions
- ~2 hours of GTA footage for transfer
- tens of hours of robot demonstrations
What you need to act on it
- public release still pending
- API access request
Sources