Abstract: In order to solve the problems of long picking paths, low efficiency and low picking quality caused by old leaves, stems and other interferences in the picking process of famous tea leaves, an improved deep reinforcement learning method based on target recognition is proposed. After the image target is preprocessed, the HIS color model is used to obtain target objects of different depths, the picking position of the shoot is obtained through the setting of parameter channels, the shape characteristics of the picking object are analyzed, and the speed, angular velocity, and distance error are used as the guiding factors of the reward function to realize the improvement of deep reinforcement learning. The path planning design of the picking process is accomplished by establishing objective functions, objective networks, and empirical recovery to achieve intensive training of the planned paths. Gazebo simulation platform is used to carry out  reinforcement learning training of picking path, simulate obstacles to achieve the optimization of picking path, complete the verification of the planning algorithm, and get with the increase of training times, the improved deep reinforcement learning method is effective for picking path optimization, the localization cutting accuracy is controlled within 0.005m, and the efficiency of path optimization is improved by 3.6%.

Key words: tea picking, image processing, improved DQN, path planning, Gazebo simulation, picking machinery

摘要： 为解决名优茶叶采摘过程中老叶、茎梗等干扰物导致采摘路径较长、效率低、采摘品质低等问题，提出一种基于目标识别的改进型深度强化学习方法。图像目标经过预处理后，利用HIS颜色模型获取不同深度的目标对象，通过参数通道的设置获取嫩芽的采摘位置，分析采摘对象的外形特征，利用速度、角速度、距离误差等作为奖励函数的导向因素，实现对深度强化学习的改进。通过建立目标函数、目标网络以及经验回收实现规划路径的强化训练，从而完成采摘过程的路径规划设计。利用Gazebo仿真平台对采摘路径进行强化学习训练，模拟障碍物实现采摘路径的优化，完成规划算法的验证，并得到随着训练次数的增加，改进型深度强化学习方法对采摘路径优化有效，定位切割精度控制在0.005m范围内，路径优化效率提高3.6%。

关键词: 茶叶采摘, 图像处理, 改进DQN, 路径规划, Gazebo仿真, 采摘机械