A new memory layer called Stigmergic Graph Memory (SGM) helps fleets of warehouse robots avoid traffic jams by learning from recent execution patterns. In benchmarks across five warehouse layouts, SGM improved task completion by 20.5% to 36.7% over the best existing many-to-many task allocation method.
What the Researchers Built
Researchers from the University of Maryland and George Mason University built Stigmergic Graph Memory (SGM), a lightweight memory system that sits between task allocation and path planning in a multi-agent warehouse system. SGM maintains decaying memory channels on both nodes (storage locations, intersections) and directed edges (aisle segments) of the warehouse graph. Each channel stores different types of recent information — node congestion, waiting events, blocking pressure, successful completions, directional traffic flow, and traversal delays.
Instead of using a single "congestion score," SGM keeps these signals separate and maps them to two control interfaces: endpoint steering and route guidance. Endpoint steering uses node memory to rank which pickup or delivery locations should be assigned to agents, while route guidance uses edge memory to influence path costs during replanning. This split design lets the system avoid problematic areas before agents even start moving.
SGM works as a plug-in layer on top of the M2M many-to-many MAPD framework (Schneider et al., 2026). It does not require changing the underlying path planner or collision-avoidance system. The researchers also built an adaptive variant, ASGM, that only activates memory-guided decisions when congestion evidence is strong, trading some throughput for lower computational overhead.
Key Results
In paired benchmarks across five warehouse layouts at three fleet sizes (28, 56, and 84 agents, representing 30%, 60%, and 90% of parking capacity), SGM outperformed both the base M2M method and its SKU-aware variant (M2M-wSKU) in every single condition — 15 out of 15 map-fleet combinations. Throughput gains ranged from 20.5% to 36.7%, all statistically significant under Holm-corrected Wilcoxon signed-rank tests.
| Condition | SGM Throughput (tasks completed) | M2M-wSKU Throughput | SGM Gain |
|---|---|---|---|
| Warehouse A, 84 agents | ~7100 | ~5700 | +24.6% |
| Warehouse B, 84 agents | ~6800 | ~5400 | +25.9% |
| Maze X, 84 agents | ~6500 | ~5000 | +30.0% |
| Maze Y, 84 agents | ~6200 | ~4500 | +37.8% |
| Average across all 15 conditions | — | — | +20.5% to +36.7% |
A component ablation revealed that endpoint steering alone achieved 99.3–100.5% of full SGM throughput, proving that the key insight is picking which sources and destinations to assign, not just how to route to them. Adding route guidance reduced planner runtime, waiting time, and blocked-motion replans in most conditions without sacrificing throughput.
In a medium-scale transfer study (larger warehouses with 200+ agents), SGM also improved completed tasks, showing the approach scales beyond the primary benchmark.
How It Works
SGM treats the many-to-many pickup and delivery problem as a two-stage decision process. First, when an agent becomes available, endpoint steering scores all feasible pickup and delivery locations using memory values. Second, route guidance modifies path costs using directional edge memory during planning.
Memory evolution formula: For each graph element (node or directed edge) and each memory channel, the stored value decays exponentially over time. A decay constant defines how quickly old information fades. New events add to the channel's value. The formula is:
M(t+1) = M(t) * decay + event_value
The decay rate and event injection logic are tuned per channel. For example, a "blocking pressure" channel on a node increments when an agent is held in place due to a predicted vertex conflict, while a "successful completion" channel increments when an agent finishes a task at that location.
Endpoint steering: When an agent needs a new goal (pickup or delivery), SGM computes a composite score for each feasible endpoint. This score is a weighted sum of memory channels relevant to that node: traffic intensity, waiting events, blocking pressure, and completion rate. High waiting or blocking pressure lowers the score; high completion rate raises it. The weighted sum is combined with the M2M assignment objective so that agents avoid hot spots while still making progress on tasks.
Route guidance: During path planning, edge memory values contribute to the travel cost. Edges with high blocking, delay, or congestion are penalized. The planner (RHCR/PBS backend) still guarantees collision-free plans; the memory costs simply steer robots toward less contested routes.
Controller variants: SGM uses memory for both endpoint steering and route guidance, and preserves queued tasks across route repairs. ASGM is an adaptive version that falls back to M2M-wSKU baseline behavior when congestion metrics are low, activating memory only when waiting, backlog, or blocking exceed a threshold. This saves computation but reduces peak throughput.
Why This Matters for Robotics
Warehouse operators running used industrial robots or warehouse robots frequently face throughput bottlenecks not from robot speed but from congestion — robots waiting at intersections, blocking each other in narrow aisles, or queuing at popular pickup stations. Traditional multi-agent planning tackles congestion reactively: route around jammed areas after they develop. SGM shows that many-to-many task allocation provides a higher-leverage control point: if you don't assign a robot to a congested aisle in the first place, you never need to route around it.
For used cobots for sale deployed in dynamic environments, a memory layer like SGM could be ported to human-robot collaboration settings where traffic patterns shift unpredictably. The key principle — separate memory for different congestion signals, used at allocation time — is domain-agnostic.
The adaptive ASGM variant is especially relevant for edge deployments with limited compute, since it only pays the memory overhead when congestion is actually a problem. Fleet operators can tune the decay constants to match their warehouse traffic rhythms without changing underlying planners or collision avoidance systems.
Limitations and Open Questions
SGM was evaluated only in simulation — no physical robot deployments. Real-world effects like sensor noise, communication delays, and hardware failures could change how memory signals behave. The memory channels and weights were manually designed; the paper does not explore learned or adaptive channel tuning. Additionally, the endpoint steering gain is so dominant that the route guidance component may be less critical; future work might simplify the architecture.
The study used a fixed set of warehouse layouts. Nonstationary environments (e.g., seasonal inventory shifts, reconfigurable storage) could degrade SGM's performance if memory decay rates are not adjusted. Finally, SGM assumes the M2M many-to-many task model — it does not directly apply to one-to-one MAPD or task allocation without inventory flexibility.
Frequently Asked Questions
What is many-to-many multi-agent pickup and delivery? In many-to-many MAPD, each task (e.g., "move item A from storage to outbound") can be fulfilled from multiple possible pickup locations and delivered to multiple possible destinations, giving the dispatcher flexibility to avoid congestion.
How does Stigmergic Graph Memory differ from standard congestion avoidance? Standard congestion avoidance is reactive — it reroutes robots after detecting a jam. SGM is proactive: it ranks endpoint choices using recent memory so that robots are assigned to less congested areas, preventing jams preemptively.
What kind of memory signals does SGM track? SGM tracks separate decaying signals for node congestion, waiting, blocking pressure, task completions, edge traversal counts, delayed traversals, and directional flow. These are kept distinct, not merged into a single congestion score.
Can SGM work with existing warehouse control systems? Yes. It is designed as a lightweight memory layer that sits between task allocation and path planning, and was evaluated on top of the standard RHCR/PBS planner. No changes to collision avoidance or robot firmware are required.
Conclusion
SGM introduces a simple but powerful idea: by maintaining decaying memory of recent traffic events and using that memory to choose which tasks to assign to which endpoints, warehouse robot fleets can avoid creating congestion in the first place. The 20-36% throughput gains across diverse layouts make a strong case for adding stigmergic memory to any many-to-many warehouse dispatch system. Future work will test this approach on physical robots and adapt it to changing warehouse layouts.