Multi-Agent AI Systems for Autonomous Fleet Orchestration in Smart Warehouse

Abstract

Autonomous mobile robot (AMR) fleets are increasingly used to move inventory inside warehouses, but the task-allocation logic that decides which robot serves which order is often either a fixed, non-adaptive dispatch rule or a proprietary black box. This paper reports a real, reproducible simulation study comparing three fully transparent, non-learned orchestration policies for a simulated autonomous ground vehicle (AGV) fleet performing pickup-and-delivery tasks in a warehouse grid: a static zone-assignment baseline, a reactive nearest-idle-agent heuristic, and a deterministic multi-signal market-auction mechanism that scores agents on travel distance, battery reserve, local congestion, and task-completion fairness. None of the three policies is a large language model or any other opaque learned controller; every decision rule is a plain, auditable function of observable state. Collision-free motion is guaranteed by Cooperative A* path planning over a shared space-time reservation table. We inject four real fault conditions (agent breakdown, aisle congestion, accelerated battery drain, and order-arrival surge) individually and in combination, and run 144 independent, seeded trials (three policies times six fault conditions times eight seeds, 1,200-tick horizon each) over a fixed 20x14-cell warehouse floor plan with four shelf blocks, three pack stations, and two charging stations. The market-auction policy and the reactive nearest-idle heuristic perform statistically indistinguishably on throughput and wait time under every fault condition (Welch's t-test, all p > 0.16), but the market-auction policy recovers from an injected agent breakdown in a mean of 14 simulation ticks versus 40 ticks for both other policies, a deterministic consequence of its shorter heartbeat-timeout window rather than a stochastic effect. The static zone-assignment baseline is significantly slower than both adaptive policies in every single condition, including with no faults injected at all (Welch's t-test, p < 0.001 for mean wait and cycle time in five of six conditions), and its completion rate collapses to roughly 90% under order-surge and combined-fault stress (chi-square test of independence, p < 0.001) while both adaptive policies sustain at or near 100% completion. We report these results, including the cases where the proposed market-auction policy shows no measurable advantage over the simpler reactive heuristic, as an honest, pilot-scale reference implementation rather than a claim of universal superiority, and we discuss the specific architectural property, heartbeat-timeout duration, responsible for the one place the two adaptive policies clearly diverge.