Robotic Picking of Cylindrical Fuel Pellets from a Boat Using 3D Range Sensor

Fabrication of nuclear fuels comprises many intricate tasks that are difficult to automate. One such task is picking of cylindrical pellets from a container, called a boat. In the boat, pellets can be randomly oriented, touching and partly or wholly occluded. They have poor contrast with respect to each other and the boat. Hence, vision based techniques with either monocular or stereo camera configurations are not suitable for identification of pellet poses. This paper describes a novel technique in which 3D point cloud data is acquired by a 3D range sensor, and processed with optimized RANSAC (O-RANSAC) algorithm for estimating pellet poses for the robot to pick. First, cylinders are identified from point cloud data, and thereafter scores indicating the extent of matching is computed. We also present the improvement in processing time for O-RANSAC algorithm as compared to RANSAC. We demonstrate that O-RANSAC algorithm is robust even in presence of outliers and noise by testing in simulation with synthetic data. Proposed method was tested in an experimental setup consisting of an articulated robot, a 3D range sensor and dummy densely packed pellets in multilayer.