Preprint finds robot policies can improve across tasks during deployment

Researchers report that generalist robot policies can improve across diverse tasks through reinforcement learning during deployment, including gains on tasks excluded from that training. The preprint extends an approach previously explored largely for improving a single deployment task. The researchers developed structured evaluations and compared choices in value estimation, policy extraction and resource allocation. They found that the training task distribution and value-based reinforcement learning were critical to whether improvement generalized to other tasks.