Learning Process Rewards via Success Visitation Matching for Efficient RL

arXiv·medium signal

Tsao, Wagenmaker, and Sergey Levine address the sparse-reward problem in RL by deriving dense process rewards from success-visitation matching rather than hand-designed reward shaping. The method targets settings where the natural task reward is inherently sparse, improving sample efficiency. Relevant to teams doing RL fine-tuning of agents or robotics policies where reward signal is scarce.