Learning Action-oriented Grasping for Manipulation

Complex manipulation tasks require grasping strategies that simultaneously satisfy the stability and the semantic constraints that have to be satisfied for an action to be feasible, referred as action-oriented semantic grasp strategies. This study develops a framework using machine learning techniques to compute action-oriented semantic grasps. It takes a 3D model of the object and the action to be performed as input and provides a vector of action-oriented semantic grasps. We evaluate the performance of machine learning (particularly classification techniques) to determine which approaches perform better for this problem. Using the best approaches, a multi-model classification technique is developed. The proposed approach is evaluated in simulation to grasp different kitchen objects using a parallel gripper. The results show that multi-model classification approach enhances the prediction accuracy. The implemented system can be used as to automate the data labeling process required for deep learning approaches.