Push and rotate: cooperative multi-agent path planning

In cooperative multi-agent path planning, agents must move between start and destination locations and avoid collisions with each other. Many recent algorithms require some sort of restriction in order to be complete, except for the Push and Swap algorithm [Luna and Bekris, 2011], which claims only to require two unoccupied locations in a connected graph. Our analysis shows, however, that for certain types of instances Push and Swap may fail to find a solution. We present the Push and Rotate algorithm, an adaptation of the Push and Swap algorithm, and prove that by fixing the latter's shortcomings, we obtain an algorithm that is complete for the class of instances with two unoccupied locations in a connected graph. In addition, we provide experimental results that show our algorithm to perform competitively on a set of benchmark problems from the video game industry.