抄録
In this research, we develop a robot system that can find and pick up a specific object from a pile of objects with a grasp planning including sweeping motion thereby decreasing a necessary number of operations. We use local visual features for recognition and pose estimation of a target object to cope with partial occlusions. Pose estimation of unknown objects are also necessary for handling them. We obtain point cloud data by an RGB-D camera (KINECT) mounted on the robot head, and extract object surfaces by Euclidean clustering, plane detection, and normal estimation. The robot adaptively chooses grasp or sweep actions depending on the placement of the target objects and others. We implemented the proposed method on a dual-arm robot and applied it to the task of find a target object in a cluttered container. We confirmed that introducing sweep motions can effectively reduce the time to achieve the task.