A paper posted Oct. 8 describes a learned safety filter that changes a robot’s proposed action before it reaches the motors. In simulation, the filter helped a 29-degree-of-freedom Unitree G1 avoid a marked zone and, after a hard push, fall away from it. The authors also showed the same controller recovering from a light push and falling away from boxes on a physical G1, but those hardware demonstrations were qualitative, not a statistical safety test.
Many safe-reinforcement-learning methods ask one policy to improve a task reward while also avoiding unsafe states. When the goals conflict, a weak safety penalty can allow violations, while a strong one can make the policy overly cautious. FAITH separates those jobs: the task policy proposes an action, and a learned filter corrects it before execution.
Rather than checking only the robot’s current position, the filter scores a proposed move by the largest future incursion of any modeled robot link into the protected area, assuming it can choose later moves to reduce that incursion. A score within a learned safety margin is treated as acceptable; otherwise, the filter changes the move to try to bring the score within that margin. The score is learned from experience, so it is not a safety certificate. The task policy is trained through the corrected actions, allowing it to adapt to the filter while optimizing task reward. At deployment, both components run as feed-forward networks rather than solving a new optimization problem for every move.
If every available action fails the learned test, the filter instead favors the one with the lowest predicted worst intrusion. That means none passed the model’s test; it does not prove that a safe move was physically impossible.
In 4,096 simulated humanoid trajectories, the paper reports that 99.95% of FAITH’s Walking-Avoid episodes stayed outside the protected region, matching the best safety rate among the tested methods. Its task return was 2,365, about 97% of the unfiltered policy’s 2,428. In the harder Push-Avoid simulation, FAITH’s safety rate was 85.45%, slightly above the next listed baseline at 84.52%, but its average return was lower: 80 versus 115.5. The controller sometimes chose a fall to keep the robot’s links away from the zone.
The test marks flat protected regions as stand-ins for areas occupied by people; it does not measure impact severity or human injury. The physical demonstrations report no safety-rate statistics, and the authors did not quantify the effects of LiDAR noise, control latency or communication delay. They say discounting, a finite safety penalty and learning errors leave the filter without a formal safety guarantee. FAITH is a promising way to make a robot’s trade-off explicit, but the paper does not certify a humanoid for work near people.
Sources
FAITH paper, arXiv · Author publication page · Illustrative photo and license
Illustrative photo, not FAITH test imagery: “Unitree G1 at Uav expo 2024” by Sayanesy, CC0 1.0.