This is potentially a great news for robot learning:

Odyssey-3 (a foundation world model) just dropped and its attacking robotics' biggest bottleneck, task-specific data.

Odyssey's approach is to start with a pretrained world model, then attach an action decoder trained on tens of hours of robot demonstrations.

Odyssey-3 drove real Indian streets after its driving policy trained on only 20 hours of simulation.

A lot of physical AI progress still comes from throwing more demonstrations at one narrowly defined system. Odyssey is arguing that broad world-model pretraining can absorb much of the knowledge that every downstream policy would otherwise need to relearn.

Their reported numbers are early but unusually small: tens of hours for robot arms, tens of hours of humanoid teleoperation, 20 hours of simulated driving, and tens of hours of simulated drone data.

The adaptation layer then learns how to turn Odyssey-3's representations into the actions of that specific system.

This is basically a data-allocation bet. Spend the huge pretraining budget once on a general model of the world, then spend far less experiential data teaching each new body its controls.