Generative simulation promises to automate the construction of robot training environments, but a physically valid environment is not necessarily a useful one. According to the abstract for CoDA, the central challenge is identifying environments that address a policy's failures rather than simply passing a physics check.

CoDA's proposed answer is to co-evolve the data engine and the action policy. Instead of treating environment generation as a separate, one-off step, the framework ties it to the policy's current weaknesses, so each new environment is selected to expose and correct specific shortcomings.

The result is a more targeted training loop: as the policy improves, the environments adapt, and as the environments get harder, the policy must follow. This stands in contrast to static or purely validity-driven simulation, which can produce plenty of scenes that teach the robot little.