Reproduce “An Informational Curse of Horizon in Goal-Conditioned Policy 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
John L. Zhou, Yuxuan Dong, Jonathan C. Kao
The authors study how distant hindsight goals affect offline goal-conditioned learning. They show that behavioral-cloning policies lose nearby-goal competence as training goals become more distant, while reinforcement-learning policies are more robust, depending on objective and policy class. Their Sobolev bootstrapping method matches a teacher’s actions and state Jacobian; on cube-triple-play, it raises success from 4 ± 2% to 30 ± 5%, with added training computation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.