FAITH: Feasibility-Aware Safety-Filtered RL for High-Dimensional Systems
FAITH separates task learning from safety filtering, then handles even states where no safe action exists.
The paper presents a model-free safety filter that learns a state-action safety value and uses a feedforward network to approximate minimal intervention. The task policy is trained through the filtered dynamics, avoiding a safety penalty competing inside the policy objective. In infeasible states, the filter moves toward the action with the lowest predicted peak harm rather than failing on an empty safe set. Reported results include strong safety-return tradeoffs in Safety Gym and humanoid tasks, plus demonstrations on a Unitree G1 humanoid. ArXiv · AI/CL/LG's note
The paper presents a model-free safety filter that learns a state-action safety value and uses a feedforward network to approximate minimal intervention. The task policy is trained through the filtered dynamics, avoiding a safety penalty competing inside the policy objective. In infeasible states, the filter moves toward the action with the lowest predicted peak harm rather than failing on an empty safe set. Reported results include strong safety-return tradeoffs in Safety Gym and humanoid tasks, plus demonstrations on a Unitree G1 humanoid. ArXiv · AI/CL/LG's note
score 4