RAD: Retrieval High-quality Demonstrations to Enhance Decision-making

arXiv cs.AI · 6 hours ago

Researchers propose RAD, a method for offline reinforcement learning that retrieves high-return states from existing datasets and uses a generative model to create trajectories toward these targets for improved planning. The approach was tested across multiple benchmarks and achieved competitive or superior performance compared to existing methods. This enables policies trained on fixed datasets to generalize better to unseen scenarios by leveraging demonstrated high-return states as planning targets.