中国科学院自动化研究所机构知识库

This paper presents a vision-based robotic grasping approach with complex obstacles environments. A deep learning-based object detection algorithm is used to detect object in the image and obtain the object's category and position. Then the Euclidean cluster extraction algorithm is adopted to segment scenes composed of 3D point clouds to obtain the positions and size of obstacles. According to the acquired information of the object and obstacles, one can judge whether the object can be directly grasped. If there is no direct solution, the obstacles that interfere
with the grasping shall be firstly moved to other positions, then the object is grasped. These new positions of interference obstacles are selected based on artificial potential field. The experimental results on the Kinova MICO2 arm demonstrate that the approach can effectively achieve the grasping of target object even with severe interference from obstacles.