Abstract: In order to harvest citrus fruits smoothly, the problem of obstacle perception and obstacle avoidance needed to be resolved in the movement of the robot arm. In this paper, the branches were segmented according to the characteristics of the branches, and the deep learning Mask R-CNN neural network was used for training and recognition. Then, the three-dimensional information of the key points of the branch obstacles was obtained with the Kinect v2 camera for reconstruction. Improved algorithm of rapidly-exploring random trees (RRT) was used to plan the obstacle avoidance motion of the manipulator. The simulation and control platform was built, and the citrus harvesting robot developed by the research group was verified in the laboratory environment. The results showed that the successful rate of obstacle avoidance is was 90.7% and the average planning time  was 1.5 s. This study laid a foundation for further practical environmental harvesting.

Key words: harvesting, robot, robotic arm, obstacle avoidance, motion planning

摘要： 利用机器人采摘柑橘果实需要解决机械臂运动过程中对障碍物的感知与避障问题。根据枝干的特征对枝干进行分段标记，使用深度学习Mask R-CNN神经网络进行训练、识别，然后与Kinect v2相机得到枝干障碍物关键点的三维信息进行重建。应用快速扩展随机树（rapidly-exploring random trees，RRT）的改进算法进行机械臂的避障运动规划。搭建了仿真及控制平台，并在实验室环境下通过课题组自行研制的柑橘收获机器人进行了验证，结果表明，样机避障成功率为90.7%，平均规划时间为1.5 s。上述结果为进一步进行实际环境采摘奠定了基础。

关键词: 采摘, 机器人, 机械臂, 避障, 运动规划