Making robot pathfinding work in real warehouses

Warehouse automation relies on multi-agent pathfinding (MAPF), but standard algorithms assume simplified 2D grids where any robot can move in four directions instantly. This work bridges that gap by introducing multi-agent warehouse pathfinding (MAWPF), which models real differential-drive AGVs: straight motion and in-place rotation with multi-step costs, acceleration/deceleration, and rules against rear-end collisions. The authors adapted four existing MAPF solvers—PP, LNS2, PIBT, and LaCAM—and benchmarked them on realistic scenarios. Results show PP and LNS2 fail on many-robot instances, while PIBT-based methods maintain scalability at modest solution-quality tradeoffs. Code and benchmarks are provided to support further development.