Collision avoidance with potential fields based on parallel processing of 3D-point cloud data on the GPU

In this paper we present an experimental study on real-time collision avoidance with potential fields that are based on 3D point cloud data and processed on the Graphics Processing Unit (GPU). The virtual forces from the potential fields serve two purposes. First, they are used for changing the reference trajectory. Second they are projected to and applied on torque control level for generating according nullspace behavior together with a Cartesian impedance main control loop. The GPU algorithm creates a map representation that is quickly accessible. In addition, outliers and the robot structure are efficiently removed from the data, and the resolution of the representation can be easily adjusted. Based on the 3D robot representation and the remaining 3D environment data, the virtual forces that are fed to the trajectory planning and torque controller are calculated. The algorithm is experimentally verified with a 7-Degree of Freedom (DoF) torque controlled KUKA/DLR Lightweight Robot for static and dynamic environmental conditions. To the authors knowledge, this is the first time that collision avoidance is demonstrated in real-time on a real robot using parallel GPU processing.