Reproduce “FAITH: Feasibility-Aware Safety-Filtered RL for High-Dimensional Systems” 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
Songyuan Zhang, Baljeet Singh, Sarthak Ranjeet Kaingade, Chuchu Fan, Bryan Trinh
FAITH learns a safety filter for robot control that corrects a task policy’s proposed actions using an estimate of future peak harm, while training the policy through filtered actions with task reward alone. In humanoid simulations, it achieves 99.95% safety in Walking-Avoid and the highest measured safety rate, 85.45%, in Push-Avoid, where lower return reflects a safety-performance tradeoff.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.