Reproduce “RAD: Retrieval High-quality Demonstrations to Enhance Decision-making” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lu Guo, Yixiang Shan, Zhengbang Zhu, Qifan Liang, Lichang Song, Ting Long, Weinan Zhang, Yi Chang
RAD is an offline reinforcement learning method that retrieves a promising nearby waypoint from recorded expert trajectories and generates a plan toward it. Its target-selection and step-estimation modules guide conditional trajectory generation without new training interactions. On AntMaze-Medium-Play with 20% regional coverage, DL-RAD achieves an endpoint success rate of 52.2 ± 2.2, versus 15.6 ± 8.0 for DRC; at 0% coverage on AntMaze-Large-Play, RAD scores zero.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.