Learning end-effector orientations for novel object grasping tasks

We present a new method to calculate valid end-effector orientations for grasping tasks. A fast and accurate three-layered hierarchical supervised machine learning framework is developed. The algorithm is trained with a human-in-the-loop in a learn-by-demonstration procedure where the robot is shown a set of valid end-effector rotations. Learning is then achieved through a multi-class support vector machine, orthogonal distance regression, and nearest neighbor searches. We provide results acquired both offline and on a humanoid torso and demonstrate the algorithm generalizes well to objects outside the training data.