RA-STGNN-AMR: Forecast-filter-rank Spatio-temporal Graph Learning for Resilient and Energy-feasible AGV/AMR Fleet Coordination

Abstract

Automated Guided Vehicles (AGVs) and Autonomous Mobile Robots (AMRs) are central to flexible material handling, yet conventional shortest-path and reactive planners cannot anticipate congestion and typically treat battery status, workstation availability, human-zone risk and disruption recovery as separate decisions.This paper presents the Resilience-Aware Spatio-Temporal Graph Neural Network for AGV/AMR coordination (RA-STGNN-AMR), a forecast-filter-rank framework coupling a Bidirectional LSTM (Bi-LSTM) edge predictor, a Spatio-Temporal Graph Neural Network (ST-GNN), hard feasibility screening and a fixed resilience-aware path cost.The factory is modeled as a time-varying graph whose node/edge states update from queue, utilization, occupancy, obstacle, workstation and battery information: the forecast stage estimates near-future travel time and congestion, the feasibility stage rejects routes violating battery, capacity, conflict or workstation constraints, and the ranking stage selects only among feasible candidates.Public MovingAI, SafeLog and 3D-warehouse MAPF instances are evaluated via a seedindexed manufacturing-state stream replayed identically for every method, across fleets of 20-100 robots, five disruption levels and 30 repeated seeds.Relative to Dijkstra, reported simulation outputs show a 31.76%reduction in average travel time, 56.78% reduction in waiting time, 51.18% reduction in recovery time and 41.20% increase in throughput; the conflict metric falls from 14.7 to 4.9 and the normalized energy-efficiency index rises from 1.00 to 1.31.Against PF-DDQN, DRL+RNN and GNN-only, travel-time reductions are 8.48-13.55%and throughput gains 9.71-13.81%.Battery ablation raises the Energy-Infeasible Route Rate from 1.8% to 11.6%.These are benchmarkadapted simulation outputs: the energy index is not measured physical energy, and physical-testbed safety, paired seedlevel inference, public code replication and absolute real-time performance remain outside the current evidence.