Reproduce “Mitigating Covariate Shift in Behavioral Cloning via Robust Stationary Distribution Correction” 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
Seokin Seo, Byung-Jun Lee, Jongmin Lee, HyeongJoo Hwang, Hongseok Yang, Kee-Eung Kim
DrilDICE trains behavioral cloning policies for offline imitation when expert-labeled states are sampled unevenly. It weights examples using a DICE-style correction, choosing high-loss distributions constrained by Bellman flow and proximity to the dataset. In Four Rooms, Room 3’s data share rose from 8.9% to 40%, and worst-quarter normalized score rose from 55.28 ± 4.74 for BC to 80.16 ± 3.44; tests were constructed shifts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.