Authors

Abstract

Given that picking robots are faced with many picking tasks in the field operation environment and the target and obstacles are located at random and uncertain positions, an obstacle avoidance planning method for the picking path of virtual robots based on deep reinforcement learning was proposed to achieve rapid route planning of robots under a lot of uncertain tasks. Next, the random motion strategy of virtual robots was set according to the physical structure of robot bodies. By comparatively analyzing the advantages and disadvantages of the observed values input by different networks, an environmental observation set was established in combination with actual picking behaviors as the network input; then, a reward function was established by introducing the idea of target attraction and obstacle repulsion contained in the artificial potential field method, aiming to evaluate the behavior of virtual robots and increase the success rate of obstacle avoidance. The results of the simulation experiment showed that the success rate obtained by virtual robots in completing the picking task reached 95.5% under obstacles set at different positions. The coverage path length of the deep reinforcement learning algorithm was reduced by 272.79 in compared with that of genetic algorithm, with a reduction rate of 5.09%. The total time consumed by navigation was 1549.24 s, which was 83.15 s shorter than that of the traditional algorithm. The study results manifest that the system can efficiently guide virtual robots to rapidly reach the random picking points on the premise of avoiding obstacles, meet picking task requirements and provide theoretical and technical support for the picking path planning of real robots.

Abstract in Chinese

针对采摘机器人在野外作业环境中，面临采摘任务数量多，目标与障碍物位置具有随机性和不确定性等问题，提出一种基于深度强化学习的虚拟机器人采摘路径避障规划方法，实现机器人在大量且不确定任务情况下的快速轨迹规划。根据机器人本体物理结构设定虚拟机器人随机运动策略，通过对比分析不同网络输入观测值的优劣，结合实际采摘行为设置环境观测集合，作为网络的输入;引入人工势场法目标吸引和障碍排斥的思想建立奖惩函数，对虚拟机器人行为进行评价，提高避障成功率。仿真实验结果显示，不同位置障碍物设置情况下虚拟机器人完成采摘任务成功率达95.5%，深度强化学习算法覆盖路径长度相比于遗传算法减少了272.79in，缩短了5.09%，整体导航用时1549.24s，相比于传统算法缩短了83.15s。研究结果表明，本系统能够高效引导虚拟机器人在避开障碍物的前提下快速到达随机采摘点，满足采摘任务要求，为真实机器人采摘路径规划提供理论与技术支撑。