← All work

Research · Multi-agent systems · 2026

Stigmergic Graph Memory

A first-author paper (under review at AAAI-27) that treats endpoint selection in many-to-many warehouse pickup and delivery as a congestion-control decision: a bounded, decaying graph-memory layer reads recent execution history to steer which source–destination pairs enter the planner, before any path is fixed.

Schematic of the Stigmergic Graph Memory execution cycle, from warehouse state through shortlisting, endpoint steering, route guidance, and the RHCR/PBS planner.
The SGM decision layer: it updates typed execution memory, ranks feasible endpoints, and derives route costs before RHCR/PBS plans a collision-free step.

Problem

In many-to-many pickup and delivery, a warehouse request names a stock-keeping unit, not a fixed pickup and drop-off. The controller has to choose an agent, a source, and a destination before any path is planned. Existing graph-guidance methods only shape routing after those goals are fixed, so endpoint selection stays blind to recent traffic — and individually cheap source–destination pairs keep piling work onto the same corridors and stations.

Approach

SGM is a bounded, decaying memory layer over the warehouse graph, inspired by stigmergy: the indirect, trace-based coordination behind ant-colony optimization. It records recent execution signals on nodes (waiting, endpoint pressure, completions) and on directed edges (traversal, delay, blocking, flow), then uses them to rank which feasible source–destination pairs enter the planner and to bias route costs. It sits on top of an RHCR/PBS planner and only re-scores feasible candidates and legal moves, so a feasibility proof guarantees every executed plan stays collision-free and task-feasible.

Line charts of cumulative completed tasks over time on the Maze A and Maze B layouts, with SGM curves rising above the stronger M2M baseline at every fleet level.
Cumulative completed tasks on the two maze layouts (mean of 25 paired seeds): SGM (vermilion) stays ahead of the stronger M2M baseline (charcoal) across low, medium, and high fleet levels.

Outcome

Across paired request streams on five layouts, three fleet sizes, and 25 seeds, SGM beats two reconstructed many-to-many baselines in all 15 map–fleet conditions, with throughput gains of 20.5–36.7%. Component ablations pin the gain on endpoint steering — deciding which work enters the planner — while route guidance further trims planner time, waiting, and blocked moves. The paper is first-authored with Joon-Seok Kim and is under review at AAAI-27.