arXivRobotics
FAITH: Feasibility-Aware Safety-Filtered RL for High-Dimensional Systems
FAITH：兼顧可行性與高維度機器人控制的無模型安全強化學習
FAITH introduces a model-free safety filter that decouples task rewards from safety constraints, minimizing predicted harm when no safe action exists, and demonstrates success on a 29-DoF Unitree G1 humanoid robot.
2 min read