Researchers led by Srikar Babu Gadipudi have built the protection of a safety filter into reinforcement-learning policies for robot teams, so that the filter is no longer needed once the robots are deployed, according to a paper on arXiv. Such filters prevent collisions by correcting unsafe actions, but for a team they become a central optimisation whose constraints grow with the square of the team size, the authors write.
The filter also hides from training what the policy needs to learn, so the method learns only from safe actions that were actually carried out and distils the filter into the policy at made-up situations close to a collision. The authors prove that this keeps errors from building up.
Acting only on local observations and without a filter, the policies had no collisions in 1,800 test episodes with several agents. A policy trained with three agents worked without retraining in teams of 20 at the same computing cost per agent and stayed collision-free with up to eight. The team also flew the policies on as many as five Crazyflie quadrotors at once.



