Reproduce “MAC-v2: Long-horizon model-based value expansion from pixels” 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
Kwanyoung Park, Sergey Levine
MAC-v2 uses latent prediction and action chunks for offline visual reinforcement learning, training its representation for prediction and behavior cloning. It selects candidate chunks from a behavior-cloning policy using a learned Q-value. On OGBench’s visual puzzle-3x3 family, it averages 95 ± 4% success, versus 22 ± 1% for FQL. These simulated results do not establish real-world deployment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.