How To Evaluate Robot Policies In Simulation At Scale

Eval time goes to policy inference, sim stepping, rendering and resets. Vectorize environments, shard across GPUs, batch inference and serve the policy separately.