CF6D: A Visual Perception Network for Collision-Free Grasping and Manipulation in Complex Industrial Environments

Accurate 6-D pose estimation is critical for robots to perform grasping tasks successfully. In recent years, numerous 6-D pose estimation algorithms have been widely applied in robotic grasping experiments. However, in complex environments, industrial robots often struggle to extract effective image features, leading to inaccurate 6-D pose estimations and subsequent grasping failures. In addition, improper grasping sequences caused by stacked parts increase the risk of collisions, further elevating the probability of grasping failure. To overcome these challenges, we presented CF6D, a collision-free grasping network based on 6-D pose estimation. Specifically, we devised a keypoint prediction network that utilizes a least squares fitting algorithm to exactly detect object keypoints and estimate 6-D poses. The network also outputs an optimal grasping sequence, enabling the robot to effectively grasp stacked parts based on precise poses and a reasonable order. On the relevant dataset, our method is obviously better than the existing methods in terms of accuracy and robustness. Furthermore, test results on a real-world robotic platform demonstrate that this method excels in practical grasping tasks, meeting the demands of industrial applications.