DISCO transfers robot planning from simulated search into a deployable policy

Researchers report that DISCO, a robot manipulation method described in an arXiv preprint, transfers plans generated in physics simulation into a learned controller that can operate without simulator-based search at deployment. They report better performance than behavior cloning, reinforcement learning and search-based baselines across simulated and real-world object rearrangement tasks. The method searches possible sequences of continuous actions, using a learned policy to propose movements and a value function to assess progress toward completion. GPU-parallelized simulation periodically feeds successful plans back into that model. The reported results concern object rearrangement, rather than general-purpose robot operation.