FAITH
A model-free safe reinforcement learning framework that approximates state-action safety values and amortizes a minimally invasive intervention filtering process, decoupling safety constraints from the task policy and supporting feasibility-aware degradation when no safe action is available.
Timeline (last 90 days)
FAITH is a model-free safe reinforcement learning framework that approximates the optimal state-action safety value and uses a feedforward network to amortize the minimal-intervention filtering process.
In FAITH, when no safe action satisfies the condition, the filter selects the action with the smallest predicted peak harm as the fallback strategy.
In the double integrator example and the Safety Gym environments, FAITH achieves the highest return among methods with no feasible-start violations, while matching the lowest harm level under infeasible starts.
On a 29-degree-of-freedom humanoid system, FAITH achieves a 99.95% safety rate in the Walking-Avoid task while retaining 97% of the unfiltered return.
In the Push-Avoid task, FAITH achieves the highest measured safety rate and learns to sacrifice balance, falling in the direction away from the protected region.
FAITH's policy was deployed on a real-world Unitree G1 humanoid robot for demonstration.
FAITH's accuracy is limited by offline data coverage and generalization ability, and the distribution of the safety value approximation error in high dimensions is not clear.
FAITH's peak-harm minimization when no safe action is available is a heuristic remedy, and its theoretical guarantees are not elaborated in the paper's abstract.
All Facts (8)
FAITH's peak-harm minimization when no safe action is available is a heuristic remedy, and its theoretical guarantees are not elaborated in the paper's abstract.
50%UnverifiedIn FAITH, when no safe action satisfies the condition, the filter selects the action with the smallest predicted peak harm as the fallback strategy.
50%UnverifiedIn the double integrator example and the Safety Gym environments, FAITH achieves the highest return among methods with no feasible-start violations, while matching the lowest harm level under infeasible starts.
50%UnverifiedOn a 29-degree-of-freedom humanoid system, FAITH achieves a 99.95% safety rate in the Walking-Avoid task while retaining 97% of the unfiltered return.
50%UnverifiedIn the Push-Avoid task, FAITH achieves the highest measured safety rate and learns to sacrifice balance, falling in the direction away from the protected region.
50%UnverifiedFAITH is a model-free safe reinforcement learning framework that approximates the optimal state-action safety value and uses a feedforward network to amortize the minimal-intervention filtering process.
50%UnverifiedFAITH's accuracy is limited by offline data coverage and generalization ability, and the distribution of the safety value approximation error in high dimensions is not clear.
50%UnverifiedFAITH's policy was deployed on a real-world Unitree G1 humanoid robot for demonstration.
50%