Researchers led by Angelo Moroncelli have proposed CAST (Certified Action Selection at Test time), a way to make robot policies learned from demonstrations respect new constraints without retraining, according to a paper on arXiv. A usual safety filter edits the policy's action to meet the constraints but ignores the task, and a large edit can leave the robot unable to continue, the authors write.

CAST instead draws a few action sequences from the unchanged diffusion policy, checks each against a safety condition and runs the certified one that a learned set of value estimates ranks highest. The authors prove that a certified step is safe whatever the ranking does, and add a viability check so the robot does not run into dead ends where no safe action is left.

In simulated tests on robomimic manipulation tasks, a planar quadrotor, a MuJoCo humanoid and a unicycle model, CAST beat the output filter on success at both the task and its constraints in nearly every setting.

Source: arXiv Robotics (cs.RO)

Background

Angelo Moroncelli

Robotics researcher

Robotics researcher.