Reproduce “Interactive and Hybrid Imitation Learning: Provably Beating Behavior Cloning” 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

Yichen Li, Chicheng Zhang

STAGGER queries an expert for actions on learner-visited states; WARM-STAGGER adds these corrections to offline demonstrations. Under deterministic realizability and recoverability, STAGGER has a better annotation-cost guarantee than behavior cloning. In MuJoCo HalfCheetah, STAGGER matched or exceeded behavior cloning with half as many annotations; gains were smaller on Ant and Hopper, and the experiments do not establish the theory for continuous control.

Summary by alphaXiv.

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