Task Reconfiguration and Congestion Self-Healing Decisions for Warehouse Robot Fleets under Communication-Failure Disturbances
-
摘要: 针对通信丢包、通道阻塞与机器人故障耦合条件下仓储机器人群任务状态不可达、局部拥堵扩散和服务恢复迟滞问题，提出通信可用性约束事件触发重构方法(communication-availability-constrained event-triggered reconfiguration, CAR-ETR)。其核心是以实际送达分组构造动态通信可用集，并由该集合同时约束故障确认、可分配机器人和任务匹配可行域；事件触发匹配、衰减占用拥堵场与不动点冲突消解分别支撑在线更新、交通疏导和安全执行。在结构化仓库、24台机器人和复合扰动条件下开展20组配对仿真。与SC-PR相比，CAR-ETR的完成任务数提高19.88%，平均流动时间、服务损失和单位任务全部分组尝试数分别降低24.61%、42.83%和37.54%；与RHCR-LNS2基线相比，完成任务数差异不显著，网关相关分组尝试数和决策时间分别降低48.85%和74.75%，但全部分组尝试数增加24.47%。0–12台同时临时故障试验表明，在恢复率不低于0.80且完成量不低于无故障条件90%的判据下，已验证容错下界为12台。参数敏感性和规模实验揭示了长最小重构间隔及48台比例负载下的性能退化。全部正式仿真均未出现离散安全约束违例。CAR-ETR适用于集中协调或计算资源受限的中等规模仓储机器人系统，实物有效性仍需进一步验证。Abstract:
Objective Warehouse robot fleets allocate streaming pickup-and-delivery tasks in narrow aisles. Packet loss can hide a healthy robot, a physical failure can invalidate its assignment, and a blockage can propagate queues. Existing studies often separate communication, task allocation, and multi-agent path finding. Scheduling may therefore use simulator-side states that were never delivered, while evaluation may omit failed transmissions, retries, or relay hops. This inconsistent information boundary obscures whether recovery is deployable. This study develops a method whose reconfiguration feasible domain is constrained by delivered-packet evidence. Methods A communication-availability-constrained event-triggered reconfiguration method, termed CAR-ETR, is proposed. Its central contribution is a cross-layer hard gate rather than a new event trigger, Hungarian matcher, congestion weight, or conflict resolver. Delivered heartbeat, retransmission, or relay packets form the availability set $ {\mathcal{A}}_{t} $; this set determines the assignable-robot set $ \mathcal{R}_{t}^{a} $, which determines which variables $ {x}_{ik,t} $ may enter task reconfiguration. Undelivered or unprobed states cannot be restored from simulator truth. The mechanism therefore changes the feasible domain instead of only modifying a matching cost or update frequency. An occupancy-history congestion field and a fixed-point executor support traffic redirection and discrete safety without being claimed as independent innovations.The warehouse is a grid graph, and each task has a release time, pickup node, delivery node, and due time. Availability is updated during heartbeat rounds. Direct delivery permits one retry; after failure, the nearest reachable robot within radius $ {R}_{c} $ may relay the packet. Source-to-relay loss increases with normalized Manhattan distance, relay-to-gateway delivery is sampled independently, and each hop permits one retry. A source enters $ {\mathcal{A}}_{t} $ only after end-to-end delivery. Every attempted heartbeat, retry, relay hop, assignment command, acknowledgement, centralized-planning packet, and local-conflict packet is counted.Reconfiguration uses fresh snapshots and is triggered by fault changes, backlog, prolonged waiting, or a maximum silent interval. The matching cost combines cached distance, projected lateness, task age, and an exponentially decayed occupancy field. The window $ H=16 $ covers four heartbeat periods and two minimum reconfiguration intervals; $ \beta =0.88 $ gives a 5.42-step half-life and retains 87.1% of the infinite-kernel mass. A fixed-point resolver keeps one priority winner for duplicate targets and removes moves into failed or stationary robots and opposite-edge swaps. Runtime assertions check vertex, swap, and failed-node invariants after every step.Experiments use a $ 31\times 23 $ warehouse with 24 robots, eight stations, and 480 steps. Compound disturbances combine correlated packet loss, four temporary blockages, and robot failures. Baselines are Greedy-A*, periodic Hungarian assignment, selective-conflict prioritized replanning, RHCR-PP with space-time reservations, and RHCR-LNS2 with large-neighborhood conflict repair. All methods share task, disturbance, communication, and safety settings. The primary experiment uses 20 paired seeds, Student-$ t $-based 95% confidence intervals, paired Wilcoxon tests, and global Holm correction over 40 comparisons. Additional experiments vary peak loss, fleet size, component removal, 0-12 simultaneous failures, a $ 5\times 5 $ $ H $-$ \beta $ grid, event thresholds, communication radius, and heartbeat period. Results and Discussions CAR-ETR completes $ 128.75\pm 6.55 $ tasks, with a mean flow time of $ 62.99\pm 1.50 $ steps, a restricted recovery duration of $ 27.85\pm 15.65 $ steps, and a recovery rate of 0.95. Relative to selective-conflict prioritized replanning, completed tasks increase by 19.88%, mean flow time decreases by 24.61%, service loss decreases by 42.83%, and all packet attempts per task decrease by 37.54%; these differences remain significant after global Holm correction. The 55.65% lower mean recovery duration is not significant after correction and is not treated as a confirmed advantage.Relative to RHCR-LNS2, the completed-task difference is -0.23%, and completed tasks, flow time, service loss, and resilience do not differ significantly. CAR-ETR reduces gateway-related packet attempts by 48.85% and decision time by 74.75%, while increasing all packet attempts by 24.47%. Relative to RHCR-PP, completed tasks are 0.73% lower, gateway-related attempts are 48.74% lower, and decision time is 66.35% lower, while all packet attempts are 23.48% higher. CAR-ETR therefore exchanges robot-to-robot local messages for lower centralized coordination and online computation rather than universally dominating rolling-horizon planning.When peak packet loss rises from 10% to 70%, completed tasks decrease by 3.00%, while all packet attempts increase by 20.96% and retry-plus-relay attempts increase sevenfold. The operational fault-tolerance criterion requires a recovery rate of at least 0.80 and completed tasks of at least 90% of the no-failure mean. All tested levels through 12 simultaneous temporary failures satisfy it, establishing a verified lower bound of 12 of 24 robots for the tested 96-step duration. Service loss nevertheless rises from 35.00 to 62.51, so this is not a theoretical threshold and cannot be extrapolated to permanent failures. An overly long minimum reconfiguration interval degrades throughput and recovery; removing congestion feedback produces the largest ablation loss. Weak scaling remains effective through 36 robots, but saturation and local-message amplification appear at 48 robots. No vertex collision, opposite-edge swap, or active-to-failed-robot overlap occurs in 1476 runs.Conclusions CAR-ETR establishes a communication-evidence gate from delivered packets to task-reconfiguration feasibility, with congestion guidance and fixed-point conflict resolution supporting execution. It improves service over local periodic replanning and reduces gateway coordination and online computation relative to strong rolling-horizon baselines, at the cost of more total packet attempts. Evidence remains limited to discrete-event simulation; continuous dynamics, wireless contention, localization error, energy constraints, permanent failures, and component interactions require hardware-in-the-loop and physical-platform validation. -
表 1 主要仿真参数
参数 取值 参数 取值 仓库规模 $ 31\times 23 $ 机器人数量 24 仿真时域 480步 工作站数量 8 任务到达率 $ 0.0135N $/步 任务停止释放 第450步 通信半径$ {R}_{c} $ 7格 心跳周期 4步 故障检测延迟 4步 最大重传次数 1 背景丢包率 0.06 主要峰值丢包率 0.50 扰动区间 第172–316步 封闭起始 第200步 故障起始 第220步 临时封闭节点 4 故障机器人 2 拥堵窗口$ H $ 16 衰减系数$ \beta $ 0.88 A*拥堵权重$ \eta $ 1.8 $ {\lambda }_{l},{\lambda }_{a},{\lambda }_{c} $ $ 2.5,0.35,10.0 $ $ {B}_{\min },{\theta }_{w} $ $ 3,2.5 $ $ {T}_{\min },{T}_{\max } $ $ 8,24 $步 中继距离指数$ \kappa $ 2 
下载: 导出CSV
表 2 主要服务、通信与计算结果(均值$ \pm $95%置信区间，$ n=20 $)
方法 完成任务数 流动时间/步 受限恢复
时长/步恢复率 服务损失 韧性指数 全部分组/
任务网关相关分组/
任务决策时间/
(ms·步$ {}^{-1} $)Greedy-A* $ 96.55\pm 5.72 $ $ 107.25\pm 7.87 $ $ 94.40\pm 34.49 $ 0.55 $ 85.49\pm 20.83 $ $ 0.692\pm 0.075 $ $ 152.54\pm 13.98 $ $ 43.68\pm 2.92 $ $ 1.192\pm 0.063 $ Periodic-HA $ 109.15\pm 2.81 $ $ 81.24\pm 3.37 $ $ 60.05\pm 24.11 $ 0.85 $ 61.30\pm 15.12 $ $ 0.780\pm 0.054 $ $ 131.41\pm 5.29 $ $ 38.37\pm 0.98 $ $ 1.368\pm 0.084 $ RHCR-PP $ 129.70\pm 6.57 $ $ 63.19\pm 2.18 $ $ 36.15\pm 18.16 $ 0.95 $ 35.31\pm 12.45 $ $ 0.873\pm 0.045 $ $ 68.80\pm 3.22 $ $ 64.34\pm 3.02 $ $ 4.007\pm 0.151 $ RHCR-LNS2 $ 129.05\pm 6.34 $ $ 61.61\pm 1.69 $ $ 39.85\pm 22.71 $ 0.95 $ 35.02\pm 12.47 $ $ 0.874\pm 0.045 $ $ 68.25\pm 3.27 $ $ 64.48\pm 3.07 $ $ 5.340\pm 0.180 $ SC-PR $ 107.40\pm 3.89 $ $ 83.56\pm 2.95 $ $ 62.80\pm 28.46 $ 0.85 $ 62.82\pm 13.96 $ $ 0.774\pm 0.050 $ $ 136.02\pm 8.02 $ $ 39.16\pm 1.55 $ $ 1.374\pm 0.083 $ CAR-ETR $ 128.75\pm 6.55 $ $ 62.99\pm 1.50 $ $ 27.85\pm 15.65 $ 0.95 $ 35.91\pm 12.24 $ $ 0.871\pm 0.044 $ $ 84.95\pm 6.29 $ $ 32.98\pm 1.67 $ $ 1.349\pm 0.095 $
下载: 导出CSV
表 3 CAR-ETR与3种关键基线的配对检验
指标 对比方法 配对均值差 $ {d}_{z} $ 全局$ {p}_{\text{Holm}} $ 完成任务数 RHCR-PP –0.950 –0.864 0.039805 完成任务数 RHCR-LNS2 –0.300 –0.146 1.000000 完成任务数 SC-PR 21.350 1.399 0.002362 流动时间 RHCR-PP –0.203 –0.049 1.000000 流动时间 RHCR-LNS2 1.381 0.357 1.000000 流动时间 SC-PR –20.568 –2.818 0.000076 受限恢复时长 RHCR-PP –8.300 –0.303 1.000000 受限恢复时长 RHCR-LNS2 –12.000 –0.318 1.000000 受限恢复时长 SC-PR –34.950 –0.684 0.119335 服务损失 RHCR-PP 0.600 0.090 1.000000 服务损失 RHCR-LNS2 0.896 0.120 1.000000 服务损失 SC-PR –26.905 –1.158 0.001001 全部分组/任务 RHCR-PP 16.152 2.331 0.000076 全部分组/任务 RHCR-LNS2 16.702 2.412 0.000076 全部分组/任务 SC-PR –51.065 –2.159 0.000076 网关相关分组/任务 RHCR-PP –31.355 –10.305 0.000076 网关相关分组/任务 RHCR-LNS2 –31.495 –10.096 0.000076 网关相关分组/任务 SC-PR –6.177 –1.333 0.000076 决策时间 RHCR-PP –2.659 –8.166 0.000076 决策时间 RHCR-LNS2 –3.992 –9.349 0.000076 决策时间 SC-PR –0.025 –0.331 1.000000
下载: 导出CSV
表 4 消融结果(均值，$ n=15 $)
配置 完成任务数 流动时间/步 受限恢复时长/步 服务损失 韧性指数 全部分组/任务 网关相关分组/任务 拥堵比例 可达比例 $ -\text{RA} $ 126.53 64.10 46.67 37.51 0.865 80.32 29.40 0.310 0.917 $ -\text{ET} $ 122.20 67.38 59.33 49.01 0.824 91.84 35.38 0.337 0.962 $ -\text{CG} $ 105.40 80.52 75.13 75.06 0.730 138.85 40.28 0.424 0.965 完整CAR-ETR 126.33 65.06 48.87 40.42 0.855 86.25 33.70 0.319 0.962
下载: 导出CSV
-
[1] SHARON G, STERN R, FELNER A, et al. Conflict-based search for optimal multi-agent pathfinding[J]. Artificial Intelligence, 2015, 219: 40–66. doi: 10.1016/j.artint.2014.11.006. [2] LI Jiaoyang, TINKA A, KIESEL S, et al. Lifelong multi-agent path finding in large-scale warehouses[C]. Proceedings of the 35th AAAI Conference on Artificial Intelligence, Vancouver, Canada, 2021: 11272–11281. doi: 10.1609/aaai.v35i13.17344. (查阅网上资料,无法确认出版地信息是否正确). [3] CHEN Zhe, HARABOR D, LI Jiaoyang, et al. Traffic flow optimisation for lifelong multi-agent path finding[C]. Proceedings of the 38th AAAI Conference on Artificial Intelligence, Vancouver, Canada, 2024: 20674–20682. doi: 10.1609/aaai.v38i18.30054. [4] ZHENG Han, MA Yining, ARAKI B, et al. Learning-guided prioritized planning for lifelong multi-agent path finding in warehouse automation[J]. Journal of Artificial Intelligence Research, 2026, 85: 28. doi: 10.1613/jair.1.20611. [5] WANG Yutong, DUHAN T, LI Jiaoyang, et al. LNS2+RL: Combining multi-agent reinforcement learning with large neighborhood search in multi-agent path finding[C]. Proceedings of the 39th AAAI Conference on Artificial Intelligence, Philadelphia, USA, 2025: 23343–23350. doi: 10.1609/aaai.v39i22.34501. [6] 王昱, 张旭秀. 一种结合选择性通信与冲突解决的多智能体路径规划方法[J]. 电子与信息学报, 2025, 47(8): 2830–2840. doi: 10.11999/JEIT250122.WANG Yu and ZHANG Xuxiu. A multi-agent path finding strategy combining selective communication and conflict resolution[J]. Journal of Electronics & Information Technology, 2025, 47(8): 2830–2840. doi: 10.11999/JEIT250122. [7] MAOUDJ A and CHRISTENSEN A L. Improved decentralized cooperative multi-agent path finding for robots with limited communication[J]. Swarm Intelligence, 2024, 18(2): 167–185. doi: 10.1007/s11721-023-00230-7. [8] VERMA A, GAUTAM A, SHEKHAWAT V S, et al. Empirical study on multi-robot task allocation under communication range restrictions[J]. IEEE Access, 2025, 13: 171818–171830. doi: 10.1109/ACCESS.2025.3614892. [9] GIELIS J, SHANKAR A, and PROROK A. A critical review of communications in multi-robot systems[J]. Current Robotics Reports, 2022, 3(4): 213–225. doi: 10.1007/s43154-022-00090-9. [10] SHIBATA K, JIMBO T, and MATSUBARA T. Deep reinforcement learning of event-triggered communication and consensus-based control for distributed cooperative transport[J]. Robotics and Autonomous Systems, 2023, 159: 104307. doi: 10.1016/j.robot.2022.104307. [11] AZNAR F, PUJOL M, and DÍEZ Á. Event-triggered adaptive consensus for multi-robot task allocation[J]. Computer Communications, 2026, 251: 108499. doi: 10.1016/j.comcom.2026.108499. [12] MAYYA S, D’ANTONIO D S, SALDAÑA D, et al. Resilient task allocation in heterogeneous multi-robot systems[J]. IEEE Robotics and Automation Letters, 2021, 6(2): 1327–1334. doi: 10.1109/LRA.2021.3057559. [13] SHEN Danqing, CHEN Xiaoming, QI Wenhai, et al. Task allocation for UAV swarms under communication attacks: An approach based on game theory and negotiation mechanism[J]. Journal of the Franklin Institute, 2025, 362(1): 107417. doi: 10.1016/j.jfranklin.2024.107417. [14] 连远锋, 田天, 陈晓禾, 等. 数字孪生辅助强化学习的燃气站场巡检任务分配算法[J]. 电子与信息学报, 2025, 47(7): 2285–2297. doi: 10.11999/JEIT241027.LIAN Yuanfeng, TIAN Tian, CHEN Xiaohe, et al. Gas station inspection task allocation algorithm in digital twin-assisted reinforcement learning[J]. Journal of Electronics & Information Technology, 2025, 47(7): 2285–2297. doi: 10.11999/JEIT241027. [15] 黄高勇, 宋俊, 方旭明, 等. 面向铁路双无人机协同巡检轨迹优化的多智能体强化学习方法[J]. 电子与信息学报, 2026, 48(5): 1936–1947. doi: 10.11999/JEIT251321.HUANG Gaoyong, SONG Jun, FANG Xuming, et al. Multi-agent reinforcement learning method for trajectory optimization in dual-UAV cooperative railway inspection[J]. Journal of Electronics & Information Technology, 2026, 48(5): 1936–1947. doi: 10.11999/JEIT251321. [16] FARUQ F, LACERDA B, HAWES N, et al. A framework for simultaneous task allocation and planning under uncertainty[J]. ACM Transactions on Autonomous and Adaptive Systems, 2024, 19(4): 21. doi: 10.1145/3665499. [17] MA Ziyuan, LUO Yudong, and PAN Jia. Learning selective communication for multi-agent path finding[J]. IEEE Robotics and Automation Letters, 2022, 7(2): 1455–1462. doi: 10.1109/LRA.2021.3139145. [18] WANG Yutong, XIANG Bairan, HUANG Shinan, et al. SCRIMP: Scalable communication for reinforcement- and imitation-learning-based multi-agent pathfinding[C]. 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems, Detroit, USA, 2023: 9301–9308. doi: 10.1109/IROS55552.2023.10342305. [19] SARTORETTI G, KERR J, SHI Yunfei, et al. PRIMAL: Pathfinding via reinforcement and imitation multi-agent learning[J]. IEEE Robotics and Automation Letters, 2019, 4(3): 2378–2385. doi: 10.1109/LRA.2019.2903261. [20] CHOI H L, BRUNET L, and HOW J P. Consensus-based decentralized auctions for robust task allocation[J]. IEEE Transactions on Robotics, 2009, 25(4): 912–926. doi: 10.1109/TRO.2009.2022423. [21] 司明, 邬伯藩, 胡灿, 等. 智能仓储交通信号与多AGV路径规划协同控制方法[J]. 计算机工程与应用, 2024, 60(11): 290–297. doi: 10.3778/j.issn.1002-8331.2310-0113.SI Ming, WU Bofan, HU Can, et al. Collaborative control method of intelligent warehouse traffic signal and multi-AGV path planning[J]. Computer Engineering and Applications, 2024, 60(11): 290–297. doi: 10.3778/j.issn.1002-8331.2310-0113. [22] WANG Zhaofei, WENG Zihao, WANG Jing, et al. An original medical logistics robot system design for the complex environment of hospitals with dynamic obstacles[J]. Robotic Intelligence and Automation, 2025, 45(2): 210–227. doi: 10.1108/RIA-03-2023-0029. [23] LI Jiaoyang, CHEN Zhe, HARABOR D, et al. MAPF-LNS2: Fast repairing for multi-agent path finding via large neighborhood search[C]. Proceedings of the 36th AAAI Conference on Artificial Intelligence, 2022: 10256–10265. doi: 10.1609/aaai.v36i9.21266. (查阅网上资料,未找到对应的出版地信息,请确认). -
下载:
图(7) / 表(4)
计量
- 文章访问数: 10
- HTML全文浏览量: 0
- PDF下载量: 1
- 被引次数: 0
