Reproduce “ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using 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

Jayjun Lee, Jessica Yin, Asif Rana, Nicholas Blauch, Octi Zhang, Sam Mady, Mohak Bhardwaj, Nima Fazeli and 3 more

ADEPT trains reusable reposing skills in simulation, then adapts them to contact-rich manipulation tasks and distills the policy for real-robot use. Its post-training combines behavior-cloning distillation, critic warm-up, and conservative PPO. Kuka-Allegro peg insertion ablations found that a 10⁻⁵ actor learning rate reached ADR 50 in five seeds, while 10⁻³ failed in five; this finding is limited to tested conditions.

Summary by alphaXiv.

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