To address the problems of high complexity and low efficiency in the cooperative pursuit of maneuvering targets by multiple unmanned vehicles in continuous and dynamic environments, a counterfactual multi-agent policy gradient algorithm based on curriculum experiential learning was proposed. The cooperative pursuit efficiency of unmanned vehicles under non-stationary and partially observable conditions was thereby improved. First, a pursuit environment with dynamically moving targets and a dynamic model of unmanned vehicles was constructed, along with the criteria for successful target capture. Then, the cooperative pursuit problem of unmanned vehicle formations was modeled as a decentralized partially observable Markov decision process (Dec-POMDP). A counterfactual policy gradient algorithm within a curriculum learning framework was proposed to guide unmanned vehicles to complete the pursuit task step by step. The complex cooperative pursuit task was decomposed into three progressively difficult stages: target approaching, target capturing, and target encircling. For each stage, corresponding step-wise and terminal rewards were designed to progressively guide the agents’ learning process. Simulation results demonstrate that the proposed method, based on curriculum learning and counterfactual policy gradients, can adapt to complex environmental changes and autonomously adjust formations and path planning. In scenarios involving dynamic targets with varying maneuverability, the multi-agent cooperative pursuit strategy exhibits strong robustness and generalization ability.
LI Zhi, WANG Hong-peng, YU Xue-wei.
Multi-unmanned Vehicle Cooperative Pursuit Method Based on Curriculum Learning and Counterfactual Multi-agent Policy Gradient[J].
Science Technology and Engineering, 2026, 26(25): 10932-10943 DOI:10.12404/j.issn.1671-1815.2505395
针对多无人车协同围捕任务中每辆无人车对环境的局部观测限制、多辆无人车策略非平稳性和协同决策的需求,提出EL-COMA采用集中训练、分布执行(centralized training with decentralized execution, CTDE)的策略学习框架。训练阶段利用所有智能体的观测和动作来评估每个智能体的联合价值,从而提供更准确的反事实基线计算。反事实基线衡量在其他无人车动作不变的环境下,无人车选择当前动作和其他动作的平均效果,从而明确无人车每个动作的真实独立贡献。反事实基线类似于优势函数基线,能减少策略梯度的方差,使学习过程更加平稳、收敛更快。中心化的评估网络可以精确指导个体策略优化。在执行阶段,每辆无人车不依赖于通信,根据自己的局部观测进行独立去中心化执行,但是可以实现接近中心化执行的性能。
WangXiaohui, ChenSitong. Review of robot localization and navigation research based on BIM[J]. Science Technology and Engineering, 2025, 25(21): 8761-8772.
[3]
Ab WahabM N, NazirA, KhalilA, et al. Improved genetic algorithm for mobile robot path planning in static environments[J]. Expert Systems with Applications, 2024, 249: 123762.
[4]
ZhuC, HuangB, LuY, et al. Distributed affine formation maneuver control of autonomous surface vehicles with event-triggered data transmission mechanism[J]. IEEE Transactions on Control Systems Technology, 2023, 31: 1006-1017.
[5]
ZhangH, GuoJ, LuoG, et al. Emerging trends in intelligent vehicles: the IEEE TIV perspective[J]. IEEE Transactions on Intelligent Vehicles, 2023, 8: 3983-3995.
XiongJun, ZhangWenbo, XiongZhi, et al. Survey of cooperative multi-agent path finding[J]. Journal of System Simulation, 2025, 37(12): 3033-3049.
[8]
CaiW Y, ChenH, ZhangM Y. A survey on collaborative hunting with robotic swarm: key technologies and application scenarios[J]. Neurocomputing, 2024, 598: 128008.
LiY, ZhaoJ Y, ChenZ H, et al. A robot path planning method based on improved genetic algorithm and improved dynamic window approach[J]. Sustainability, 2023, 15(5): 4656.
[12]
LiuY, ZongC, DaiC, et al. Behavioral decision-making approach for vehicle platoon control: two noncooperative game models[J]. IEEE Transactions on Transportation Electrification, 2023, 9: 4626-4638.
[13]
WeiW, WangJ, DuJ, et al. Differential game-based deep reinforcement learning in underwater target hunting task[J]. IEEE Transactions on Neural Networks and Learning Systems, 2023, 36(1): 462-474.
[14]
GaoY, BaiC, ZhangL, et al. Multi-UAV cooperative target encirclement within an annular virtual tube[J]. Aerospace Science and Technology, 2022, 128: 107800.
[15]
SunW, TsiotrasP, LollaT, et al. Multiple-pursuer/one-evader pursuit-evasion game in dynamic flowfields[J]. Journal of Guidance, Control, and Dynamics, 2017, 40(7): 1627-1637.
[16]
LiuL X, WangX, YangX, et al. Path planning techniques for mobile robots: review and prospect[J]. Expert Systems with Applications, 2023, 227: 120254.
JiangBo, HuXiaohan, ZhengYuan, et al. Local high-density UAV collision avoidance decision based on reinforcement learning[J]. Science Technology and Engineering, 2025, 25(30): 13167-13178.
[19]
EcoffetA, HuizingaJ, LehmanJ, et al. First return, then explore[J]. Nature2021, 590(7847): 580-586.
[20]
VinyalsO, BabuschkinI, CzarneckiWM, et al. Grandmaster level in StarCraft II using multi-agent reinforcement learning[J]. Nature, 2019, 575(7782): 350-354.
[21]
SilverD, HubertT, SchrittwieserJ, et al. A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play[J]. Science, 2018, 362(6419): 1140-1144.
[22]
BakerB, KanitscheiderI, MarkovT, et al. Emergent tool use from multi-agentautocurricula[J]. arXiv prePrint, arXiv: 1909. 07528, 2019.
PanYunwei, LiMin, ZengXiangguang, et al. AUV obstacle avoidance and path planning based on artificial potential field and improved reinforcement learning[J]. Acta Armamentarii, 2025, 46(4): 72-83.
[25]
NolanB, JakobN F, SarathC, et al. The Hanabi challenge: a new frontier for AI research[J]. Artificial Intelligence, 2020, 280: 103216.
MiaoTianyue, WangLu, HeJiaxiao, et al. Research on digital twin simulation method of industrial robot integrated with reinforcement learning[J]. Journal of System Simulation, 2024, 36(12): 2971-2983.
[28]
WangT, GuptaT, MahajanA, et al. RODE: learning roles to decompose multi-agent tasks[J]. arXiv prePrint, arXiv: 2010. 01523, 2020.
[29]
XiaJ W, LuoY S, LiuZ K, et al. Cooperative multitarget hunting by unmanned surface vehicles based on multi-agent reinforcement learning[J]. Defence Technology, 2023, 29: 80-94.
FuYanfang, LeiKailin, WeiJianing, et al. A hierarchical multi-agent collaborative decision-making method based on the actor-critic framework[J]. Acta Armamentarii, 2024, 45(10): 3385-3396.
[32]
CaoX, LiM Y, TaoY T, et al. HMA-SAR: multi-agent search and rescue for unknown located dynamic targets in completely unknown environments[J]. IEEE Robotics and Automation Letters, 2024, 9(6): 5567-5574.
[33]
ZhangK, YangZ, BasarT. Multi-agent reinforcement learning: a selective overview of theories and algorithms[J]. Handbook of Reinforcement Learning and Control, 2021, 325: 321-384.
[34]
ChenL, WangY N, MoY, et al. Multiagent path finding using deep reinforcement learning coupled with hot supervision contrastive loss[J]. IEEE Transactions on Industrial Electronics, 2023, 70(7): 7032-7040.
[35]
SkrynnikA, AndreychukA, NesterovaM, et al. Learn to follow: decentralized lifelong multi-agent path finding via planning and learning[C]// Proceedings of the AAAI Conference on Artificial Intelligence. Vancouver: AAAI, 2024: 17541-17549.