Researchers led by Ye Lin Tun, with Choong Seon Hong among the authors, have built FedDIA, a method for robot fleets that learn a shared navigation policy without exchanging raw experience, according to a paper on arXiv. Usual methods weight every robot equally or by its rewards, which favours robots in easy places that have already stopped learning.
FedDIA scores each robot by how much its local training improves the shared model, weighted by how hard its environment is. On a navigation task without maps, with eight robots in different settings, it reached 75.5% average success against 68.7% for the strongest federated baseline. The advantage held in warehouse and home layouts and with one faulty robot.



