Reproduce “GSBR: Geometric Support Behavior Regularization for Online RL” 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
Charles Xu, Jacob Berg, Liyiming Ke, Sergey Levine
GSBR guides online robot learning by penalizing actions only outside neighborhoods around sampled actions from a pretrained controller. This lets the actor choose among supported behaviors while permitting nearby improvements at a distance-based cost. Across 19 tasks in four benchmarks, GSBR averages 95% success and 0.95 normalized throughput, versus RLT’s 0.77; on LIBERO, RLT has higher success.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.