Solving an industry-inspired generalization of lifelong MAPF problem including multiple delivery locations

Abstract

In this study, we worked on a new industry-inspired generalization of the multi-agent path finding (MAPF) problem. In this new problem, incoming jobs should be assigned to agents in a way that minimizes the total path cost. In this problem, definition of a job is visiting a certain coordinate by the agent. In addition, the agents must keep the following tasks on their own lists and decide the order of visiting the tasks assigned to them by minimizing the total distance traveled. This new structure can be applied to many real-world problems like cleaning an area with autonomous robots, automated warehouses and area protection with autonomous robots. In solving this problem, three different problems must be solved; to which agent a new job will be assigned, in what order the agents will handle that job, and finally, how agents will act without colliding with each other when the agents have multiple destinations. This problem is different from the job-distribution problems in the MAPF literature since the optimization criterion is total path cost and agents have to decide the order of the jobs to be visited. For this reason, new job-distribution methods that meet these goals are presented in this study. Five new heuristic algorithms that produce fast solutions for job-distribution are presented. Furthermore, we presented one brute force algorithm to find the optimal path-cost solution and compare with the heuristics created. Experiments conducted on benchmark maps have demonstrated that the job-distribution algorithm yielding the best performance resulted in path costs that were at most three to three and a half percent higher compared to the brute force approach. Additionally, the heuristics developed were found to operate at least ten times faster in terms of processing time when compared to the brute force approach. The CBS-D*-lite algorithm was previously presented in the literature and shown to produce good results in incremental environments. With these modifications, agents can have multiple destinations and collisions between the agent paths are avoided.