Stigmergic Graph Memory boosts warehouse robot throughput 20-36%

Researchers Aditya Dutta and Joon-Seok Kim from arXiv have introduced Stigmergic Graph Memory (SGM), a novel approach to improving many-to-many multi-agent pickup and delivery (MAPD) in automated warehouses. Unlike traditional methods that optimize routing after goals are fixed, SGM adds a lightweight memory layer that records recent execution signals—like congestion and successful picks—on warehouse nodes and directed edges. This bounded, decaying memory helps rank feasible endpoints and route preferences before goals are instantiated, effectively informing the allocation process with real-time traffic history. SGM does not alter collision constraints or the planner's validity, making it a low-risk drop-in enhancement.

Across extensive experiments on five warehouse layouts, three load levels, and 25 random seeds per condition, SGM outperformed two reconstructed many-to-many allocation baselines in all 15 map-load combinations. The paired throughput gains ranged from 20.5% to 36.7%, demonstrating that recent execution memory can significantly improve warehouse throughput by guiding which feasible goals enter the planner—not just by optimizing paths to already fixed goals. This work offers a practical, scalable way to boost fulfillment efficiency without expensive hardware changes.