Reproduce “Drift, Then Spread: Learning One-Step Generative Policies for Online Reinforcement Learning” 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

Chunyu Luo, Hongyu Zhao, Guangyi Ni, Yue Song, Lingkai Kong

DriftSpread trains one generator for continuous-action control by shifting actions toward higher critic values, then spreading their distribution to represent entropy. It needs no policy-density or score estimates and generates each candidate in one pass. Across five MuJoCo locomotion tasks, it has the highest normalized return among evaluated methods; its fixed-critic theorem does not prove convergence of practical online training.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.