New MAWPF framework brings realistic warehouse pathfinding to AGVs
Differential-drive AGVs now get multi-step rotation costs and acceleration constraints...
A new paper from Hiroki Nagai and Keisuke Okumura, to be presented at IJCAI 2026, introduces MAWPF (Multi-Agent Warehouse Pathfinding). This framework adapts classical one-shot multi-agent pathfinding (MAPF) from idealized 2D grids to the real-world constraints of differential-drive automated guided vehicles (AGVs) in warehouses. MAWPF imposes four practical constraints: agents can only move straight or rotate in place, rotations cost multiple time steps, acceleration and deceleration are explicitly modeled, and rear-end collisions are prohibited. These changes close the gap between abstract gridworld simulations and actual warehouse automation.
The researchers benchmarked four representative suboptimal MAPF algorithms—PP, LNS2, PIBT, and LaCAM—adapted to MAWPF. Their experiments revealed that PP and LNS2 fail to handle instances with many agents, while PIBT-based approaches offer significantly better scalability, though at the cost of longer paths. This trade-off highlights that no single algorithm is optimal across all warehouse scenarios. The work provides a practical foundation for deploying lightweight one-shot pathfinding in real AGV fleets, improving collision avoidance and operational efficiency.
- MAWPF adds four realistic constraints: straight/rotate-only motion, multi-step rotation costs, acceleration/deceleration, and rear-end collision prohibition.
- PIBT-based algorithms scale to many agents but increase solution cost; PP and LNS2 fail with large agent counts.
- Paper to appear at IJCAI 2026; bridges gap between gridworld MAPF theory and operational warehouse AGV planning.
Why It Matters
Enables safer, more efficient AGV routing in warehouses by adapting proven pathfinding algorithms to real-world physics.