A Learning Method to Determine How to Approach an Unknown Object to be Grasped

We present a learning algorithm to determine the appropriate approaching pose to grasp a novel object. Our method focuses on the computation of valid end-effector orientations in order to make contact with the object at a given point. The system achieves this goal by generalizing from positive examples provided by a human operator during an offline training session. The technique is feature-based since it extracts salient attributes of the object to be grasped rather than relying on the availability of models or trying to build one. To compute the desired orientation, the robot performs three steps at run time. Using a multi-class Support Vector Machine (SVM), it first classifies the novel object into one of the object classes defined during training. Next, it determines its orientation, and, finally, based on the classification and orientation, it extracts the most similar example from the training data and uses it to grasp the object. The method has been implemented on a full-scale humanoid robotic torso equipped with multi-fingered hands and extensive results corroborate both its effectiveness and real-time performance.