Massively multi-agent pathfinding made tractable, efficient, and with completeness guarantees

Pathfinding is an important underlying task for many autonomous agents. Abstracting the environment into a navigation graph (e.g., a grid map) enables a mobile unit to plan its path to goal using heuristic search. For example, an A* search finds an optimal path. With multiple units moving simultaneously inside a shared space, the goal is to navigate each unit to its target without colliding into static obstacles or other units. This problem is much harder. Even without motion constraints, finding optimal solutions in a fully known, two-dimensional environment is NP-complete [1, 6]. With both branching factor and number of states growing exponentially in the number of units, a centralised search in the combined state space of all units is intractable in practice even on relatively small collections of mobile units. However, problems in applications such as robotics, logistics, military operations planning, disaster rescue, and computer games often involve ‘massively’ large numbers of agents. Traditional multi-agent path planning approaches each has its particular strengths. Centralised methods preserve solution optimality and completeness by planning globally, and sharing information centrally. Decentralised methods decompose the problem into a series of smaller searches, which can be much faster, and scale up to much larger problems. However, each approach also has serious drawbacks. For instance, the optimality requirement is very costly in practice. [4] incorporates decentralised planning for non-interfering subgroups of units (ID) to an improved centralised planning (OD), and scales much better than a standard centralised A*. But as reported in the paper, the