Multi-agent pathfinding for deadlock handling in rotational movements

Abstract

Listen

Translate article

Multi-agent pathfinding (MAPF) is a challenging problem especially in environments with a large number of agents and small map size. Although some recent studies revealed that the reinforcement learning approach is feasible to train the agents for resolving the deadlock problem at large scales in terms of the number of agents, the robots, which commonly appear in warehouses, are ignored in their design. As the robots in warehouses using zero-radius right-angle turns (e.g., KIVA robots) usually stay in the same location for a long time with the same rotation movements, the deadlock problem becomes more severe especially in dense areas. In this paper, a reinforcement and imitation learning algorithm with a deadlock avoidance scheme with reinforcement and imitation learning, a deadlock breaking scheme and a deadlock resolving scheme are proposed for resolving the deadlock problem by considering the configurations for a rotation-based warehouse environment including moving forward, rotating clockwise, and rotating anticlockwise only. In addition, a light model of the proposed algorithm is also presented to speed up the inference process. In the experimental results, the proposed algorithm and its light version are shown to be effective and useful for handling different scenarios of the deadlock problem.