面向养殖网箱巡检任务的强化学习训练系统

Reinforcement Learning Training System for Aquaculture Cage Inspection Tasks

网箱巡检是网箱养殖过程中的必要环节,利用机器人实现网箱无人化巡检是未来趋势。在真实水下环境中试验成本高、危险性大,目前尚缺乏适用的网箱养殖场景仿真平台。为此,论文设计实现了一个面向养殖网箱巡检任务的强化学习训练系统,给出了系统总体框架。首先基于UUV Simulator仿真实现一个网箱养殖环境。在此基础上,利用机器人操作系统(Robot Operating System,ROS)和OpenAI Gym实现强化学习训练系统。该系统可用于基于强化学习的网箱巡检控制策略训练,也可进行网箱巡检控制策略仿真和评估。最后,通过一个实验案例验证了该系统的有效性、可用性和方便性。该系统可提升网箱巡检控制算法的研发效率和安全性,进一步推动网箱养殖的智能化。

Cage inspection is a essential part of the cage culture,it is the future trend to achieve unmanned cage inspection with robots.However,the high cost and danger of experimenting in real environment,the lack of suitable simulation platform for cage culture scene hindered the development of unmanned cage inspection.To this end,a reinforcement learning training system is designed and implemented for the aquaculture cage inspection tasks with the general framework of the system.Firstly,a cage cul⁃ture simulated scene is implemented based on UUV Simulator.Based on the scene,the reinforcement learning training system is im⁃plemented using robot operating system(ROS)and OpenAI Gym.The system can be used for training reinforcement learning based cage inspection control policies,and also for cage inspection control policies simulation and evaluation.Finally,a experimental case is given to validate the effectiveness,usability and convenience of the system.The results indicate that the system can improve the efficiency and safety of research and development of cage inspection control algorithms,and furthermore,it promotes the intelli⁃gent development of cage culture.