鱼群涌现机制下多机器人运动强化的迁移控制
作者:
作者单位:

1.上海理工大学;2.国立台湾大学

作者简介:

通讯作者:

中图分类号:

TP.242

基金项目:

国家自然科学基金项目(面上项目,重点项目,重大项目)


The Transfer Control of Multi-Robotics Motion Reinforcement Employing Fish Schooling Emergency Mechanism
Author:
Affiliation:

University of Shanghai for Science and Technology

Fund Project:

The National Natural Science Foundation of China (General Program, Key Program, Major Research Plan)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    采用鱼群交互模型驱动多智能体系统可以涌现出优良的集群运动特性,具体表现为集群规模可缩放,集群单体局部通信,集群运动自组织等。但是由于轮式机器人与真实鱼类相比具有较大的差异性,使得鱼群数据训练的控制模型难以迁移应用于真实机器人系统。为此,提出了一种结合深度学习与强化学习的迁移控制方法,该方法首先使用鱼群运动数据训练深度网络(Deep Neural Network, DNN)模型,以此作为机器人成对交互的基础,然后向后串联强化学习的深度确定性策略梯度方法(Deep Deterministic Policy Gradient, DDPG)来修正DNN模型的输出,以保证机器人具有沿墙轨道跟踪和安全运动的能力,在上述DNN+DDPG模型的基础上,设计集群最大视觉尺寸方法挑选关键邻居,从而将DNN+DDPG模型拓展到多智能体的运动控制。集群机器人运动实验表明:所提方法能使机器人仅利用单个邻居信息就能形成可靠、稳定的集群运动,与单纯DNN直接迁移控制相比,所提DNN+DDPG控制框架既保存了原有鱼群运动的灵活性,又增强了机器人系统的安全性与可控性,使得该方法在集群机器人运动控制领域具有较大的应用潜力。

    Abstract:

    The multi-agent system driven by the interaction model of fish schooling can emerge excellent characteristics of collective motion, such as scalable group size, local communication of individuals and self-organized collective motion. However, due to the individual differences between robots and real fish, it is difficult for the control model trained by the data of fish schooling to be directly applied to the actual robotics system. Hence, a transfer control method combined with deep learning and deep reinforcement learning is proposed. Firstly, a Deep Neural Network (DNN) model is trained by the data of fish schooling, which is the basement for the interactive control of robots. Then, a deep reinforcement learning method (Deep Deterministic Policy Gradient, DDPG) is connected to the output of DNN model for ensuring the ability of a robot following the wall and moving safely in the group. Finally, based on the above DNN+DDPG model, a key neighbor selection method of the maximum group visual size is designed to expand DNN+DDPG model to multi-agent motion control. Collective motion experiments show that the proposed method can formulate reliable and stable collective motion of the robots via individual information. Compared with the pure DNN transfer control, the proposed DNN+DDPG control frame not only preserves the flexibility of the collective motion of fish schooling, but also enhances the safety and controllability of the robotics system. Thus, there exists strong potential application of the proposed method for the swarm robotics motion control.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-09-05
  • 最后修改日期:2021-12-19
  • 录用日期:2021-12-30
  • 在线发布日期: 2022-02-01
  • 出版日期: