Object Recognition and Grasp Detection in Complex Scenarios for Home Service Robot

With the development of artificial intelligence, home service robots have been widely applied in daily life. As one of the fundamental functions of service robots, object grasping helps improve the level of unmanned and intelligent household. However, traditional robots still face challenges in flexibility and accuracy in complex home environments. In this paper, an object recognition and grasp detection system in complex scenarios for home service robots is proposed. The system consists of four modules: image acquisition, human-machine interaction software, RGB-D image processing, and robot control. The image acquisition module captures RGB and depth images of the work area and sends them to the software. The RGB-D image processing module begins with object recognition to obtain object categories and bounding box with the help of YOLOv8. The target point cloud is produced by combining the target subimage and corresponding depth information, followed by depth completion and plane expansion. Then the processed point cloud is fed into a Graspnet pose estimation network to calculate the 6D grasping pose for service robot. And finally, the robot control module is invoked to carry out the corresponding target grasping action. To evaluate the performance, an experimental platform for home service robot is built to grasp various daily necessities. The results tell that the average accuracy for object recognition reaches up to 98.3%, the success rate for single-target grasping is 94.4%, and the success rate for multi-target grasping can achieve 80%.