Reproduce “Monotone Viability Learning for Offline Safe 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
Ziqian Wang, Zhen Zhang
MoViL learns viable states from offline transitions by recursively labeling states whose recorded successors remain safe. At deployment, it screens proposed actions and uses a cost-based fallback when certification fails. On a Dubins car, its viable-set overlap is 99.34 ± 0.25% against an independent Hamilton–Jacobi reference, an empirical result that does not guarantee deployment safety.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.