Decentralized Adaptive Path Selection for Multi-Agent Conflict Minimization

Consider multiple robots moving towards individual goals in a cluttered environment. While contacts between robots in these situations can be averted by reactive collision avoidance methods, deadlocks may arise in tight spaces if robots move along precomputed, conflicting paths. To resolve these issues, methods have been proposed which consider robots that employ communication, or centralized planning, or follow predefined rules. This work considers only decentralized planning solutions that employ minimum information, i.e., each robot has access only to the current position of its neighbors, without using any form of prediction, intent recognition or agent modeling. This leads to a study of several methods for minimum-conflict path selection among dynamic obstacles. The evaluation of these methods in varying simulated benchmarks, provides the following insights: (a) considering the “minimumconflict” path given the other agents’ current positions is critical for deadlock avoidance, (b) reasoning over a diverse set of paths that provide multiple alternatives improves path quality, and (c) accumulating cost over alternative paths by finding the agent’s best path selection in hindsight allows robots to learn effective highlevel strategies in a computationally efficient way that is adaptive to the other agents’ behavior, reducing task completion time and average path lengths. 1