The Effectiveness Index Intrinsic Reward for Coordinating Service Robots

Modern multi-robot service robotics applications often rely on coordination capabilities at multiple levels, from global (system-wide) task allocation and selection, to local (nearby) spatial coordination to avoid collisions. Often, the global methods are considered to be the heart of the multi-robot system, while local methods are tacked on to overcome intermittent, spatially-limited hindrances. We tackle this general assumption. Utilizing the alphabet soup simulator (simulating order picking, made famous by Kiva Systems), we experiment with a set of myopic, local methods for obstacle avoidance. We report on a series of experiments with a reinforcement-learning approach, using the Effectiveness-Index intrinsic reward, to allow robots to learn to select between methods to use when avoiding collisions. We show that allowing the learner to explore the space of parameterized methods results in significant improvements, even compared to the original methods provided by the simulator.