Energy-Constrained Hybrid Repair for Lifelong Multi-Agent Path Finding in Smart Warehouses

Smart warehouses require autonomous mobile robots to complete lifelong tasks while avoiding conflicts, respecting battery constraints, and sharing charging stations. Existing MAPF methods provide strong conflict resolution, but energy, charging, and online action repair are commonly evaluated separately. We present ECR-HR, an energy-constrained hybrid repair framework that combines a normalized energy model, charging-aware goals, risk-informed priorities, and bounded local conflict repair. The scientific contribution is a coupled execution and evaluation interface rather than a new complete MAPF solver or a claim of dominance over MAPF-LNS2. In reproducible simulation, we compare ECR-HR with classical, repair-based, lazy-search, conflict-based, and learning-based baselines. In 40-seed nominal evaluation, ECR-HR reduces candidate conflict rate relative to WHCA* from 0.0479 to 0.0255 (p=3.89×10−6) while MAPF-LNS2 achieves the strongest raw success. A 30-seed study using MovingAI map geometry, priority and repair comparisons, module-level runtime profiling, simulated disturbance tests, 25-seed energy coefficient sensitivity, and preference weight sensitivity further define the framework’s operating boundary. Enhanced GNN-PPO-HR increases held-out success from the original 0.188 to 0.753±0.174 but remains below mature search baselines. All evidence is simulation-based, the energy coefficients are normalized rather than hardware-calibrated, and real-robot validation remains necessary.