基于鱼群涌现行为启发的集群机器人硬注意力强化模型  

Hard attention reinforcement model for swarm robotics inspired by fish school emergence behavior

作　　者：刘磊[1,2] 葛振业[2] 林杰 陶宇 孙俊杰 Liu Lei;Ge Zhenye;Lin Jie;Tao Yu;Sun Junjie(School of Management,University of Shanghai for Science&Technology,Shanghai 200093,China;School of Optoelectronics,University of Shanghai for Science&Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学管理学院,上海200093 [2]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《计算机应用研究》2024年第9期2737-2744,共8页Application Research of Computers

基　　金：上海市自然科学基金资助项目(22ZR1443300)。

摘　　要：生物集群运动模型能使集群机器人涌现秩序,但是所形成的机器人自然集群秩序难以有效地被人工控制,为此提出鱼群硬注意力模型来解析实验鱼群数据中的交互行为。该模型通过编码器网络、图注意力网络、信息聚合网络、预解码网络以及最终解码网络等结构来获取焦点单体的重要邻居;再利用深度确定性策略梯度技术设计轨道强化网络与安全强化网络,以实现集群的人工控制。多智能体仿真与集群机器人实验结果表明:所提方法能够实现集群的人工轨道、安全控制,重要邻居信息为解决集群运动的强化学习难题提供了新思路,所提控制模型在无人机群空中协作、智慧农机集群作业、物流仓储多体搬运等领域具有较大的应用潜力。The biological swarm motion model enables the emergence of order in robot collectives,but controlling the natural swarm order formed by robots is challenging.To address this issue,this paper proposed the fish school hard attention model to analyze interaction behaviors in experimental fish school data.This model utilized structures such as an encoder network,graph attention network,information aggregation network,pre-decoding network and a final decoding network to capture crucial information about the focal individual s important neighbors.Subsequently,it employed deep deterministic policy gradient techniques to design trajectory reinforcement networks and safety reinforcement networks to achieve artificial control of the swarm.Results from multi-agent simulations and experiments with swarm robotics demonstrate that the proposed method can realize artificial trajectory and safety control of collectives.The utilization of high-attention neighborhood information for resolving reinforcement learning challenges in collective motion provides a novel approach.The proposed control model exhibits substantial potential applications in areas such as collaborative aerial operations of drone swarms,intelligent agricultural machinery operations,and multi-robot material handling in logistics and warehousing.

关 键 词：自然秩序人工控制 集群硬注意力机制 多智能体运动强化学习 集群机器人任务控制 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]