Comparison of machine learning techniques for self-collisions checking of manipulating robots

In this paper, we deal with the problem of selfcollision detection for a mobile-manipulating robot. Typically, this problem is solved by the method that precisely checks the collision between triangles in the 3D meshes. In this case, the iterative methods for collision checking use techniques like Bounding Volume Hierarchy that reduce the computation time. However, collision checking is still time-consuming during motion planning when this procedure is executed multiple times. To deal with this problem, we propose to define collision detection as a binary classification problem. Then, we show how to collect samples to train the machine learning model for classification. We systematically compare a set of techniques and evaluate them in the task of motion planning for a robotic arm taking into account the accuracy and computation time. The obtained collision classifier is implemented and verified in the Robot Operating System.