DMAPF: A Decentralized and Distributed Solver for Multi-Agent Path Finding Problem with Obstacles

Abstract

ion (e.g., planning for each agent independently then combine the partial plans to obtain the solution – while resolving any conflict), dealing with MAPF problems with a large number of agents (greater than a thousand) still remains a challenge. We have designed a decentralized and distributed MAPF algorithm, named DMAPF – Distributed Multi-Agent Pathfinder – with scalability in mind. DMAPF is able to take advantage of distributed computing to cope with the possibility of having an ever-increasing problem size. An input problem to DMAPF is encoded in answer set program as generated by the ASPRILO project [6] and solved using mainly the answer set programming, which will be introduced in Section 2.2. The communication between distributed components in DMAPF is facilitated by the Robot Operating System (ROS), which will be introduced in Section 2.3. DMAPF is an improvement over our original system, ros-dmapf [10]. We address the issues in rosdmapf where it (i) only works in maps without obstacles; and (ii) has a low success rate with dense maps. DMAPF shares the same idea with ros-dmapf in that, given a MAPF problem, it divides the problem spatially and assigns each divided subproblem to a distributed solver, which is a ROS node. The difference is that DMAPF also further divides each subproblem into disconnected regions called areas. This, together with a few other changes, allow DMAPF to deal with having obstacles. The details of problem division will be explained in Secion 3.1. After the subproblems have been distributed, each solver individually creates an abstract plan – a sequence of areas that an agent needs to visit to reach the area that contains its goal – for each agent in the given subproblem. The details of abstract plan creation will be explained in Section 3.2.1. After the plans have been made for every agent, the solvers interleave communicating with neighboring solvers to send/receive migrating agents (details in Sections 3.2.2, 3.2.3, and 3.2.5) with movement planning (details in Section 3.2.4) along each round. This differs from ros-dmapf where it does the communication from beginning to end first, then finds a movement plan for each round in a single attempt. If ros-dmapf is unable to make a plan during the movement planning phase, it would be difficult to adjust the border assignment since assignments for the next rounds have already been agreed upon, whereas DMAPF can easily adjust the assignment and retry. This, together with changes in migration protocol, allow DMAPF to achieve a higher success rate and better solution quality than ros-dmapf. DMAPF has been implemented and compared with other state-of-the-art solvers. Section 4 shows results of the comparison. Section 5 concludes with discussion on the results, advantages, limitations, and future work.