Paper Overview
Field: ML Authors: Songyuan Zhang, Baljeet Singh, Sarthak Ranjeet Kaingade, Chuchu Fan, Bryan Trinh Published: 2026-10-08 arXiv: 2610.12432
Summary
Safe reinforcement learning commonly places safety and task performance in the same policy objective, where they can introduce competing updates. Safety filters separate them at action execution, but classical designs require an analytic safety function and dynamics model, and standard minimal-intervention filters are myopic to long-horizon task return because they minimize only instantaneous action deviation. Hard projections are also undefined when no safe action exists.
We present FAITH, a feasibility-aware, model-free framework that approximates the optimal state-action safety value and amortizes minimal-intervention filtering with a feedforward network. The task policy optimizes the task return through the filtered dynamics, which recovers the feasible constrained problem without competing safety terms. When no action satisfies the learned safety condition, the same filter approximation predicts the action minimizing peak harm.
Results
- Double integrator and Safety Gym: FAITH achieves the highest return among methods with no violations from feasible starts, and matches the lowest harm from infeasible starts.
- 29-DoF humanoid: The Walking-Avoid task reaches a 99.95% safety rate while retaining 97% of the unfiltered return.
- Push-Avoid: The policy learns to sacrifice balance and fall away from the protected region, achieving the highest measured safety rate.
- The same policy was also demonstrated on a real Unitree G1 humanoid robot.