Planning over MAPF Agent Dependencies via Multi-Dependency PIBT

A team of researchers has introduced a new algorithmic framework, Multi-Dependency PIBT (MD-PIBT), that represents a major advance in Multi-Agent Path Finding (MAPF). MAPF is the critical computational problem of planning collision-free paths for dozens to thousands of agents, like robots in a warehouse or autonomous vehicles. The work, led by Zixiang Jiang, Yulun Zhang, Rishi Veerapaneni, and Jiaoyang Li, directly addresses a scalability bottleneck in one of the field's most popular algorithms, Priority Inheritance with Backtracking (PIBT), and its recent extension, Enhanced PIBT (EPIBT).

The core innovation of MD-PIBT is planning over "agent dependencies"—a more sophisticated model of how agents' planned paths influence each other. While PIBT and EPIBT are restricted to resolving conflicts involving only one other agent at a time, MD-PIBT's framework can search over multiple, complex dependencies simultaneously. This makes it far more general and powerful; specific configurations of MD-PIBT can reproduce the older algorithms, while new configurations enable novel planning strategies impossible before.

In rigorous experiments, MD-PIBT demonstrated remarkable performance, effectively planning for massive swarms of up to 10,000 homogeneous agents. It was tested under various real-world "kinodynamic" constraints that model actual robot movement, including pebble motion, rotation, and differential drive robots with speed and acceleration limits. The researchers found the algorithm is particularly effective for MAPF problems involving "large agents" that occupy multiple grid cells, a common scenario in logistics and manufacturing.