• unknown objects
• Human Data
• Basic-Shape Decomposition
• Grasp

This thesis proposes GraspAbility (GA), a grasp planning algorithm for grasping generic unknown objects with a robot. The point cloud of the considered object is acquired and it is decomposed into cuboids of certain size through the Minimal Volume Bounding Box algorithm. At this point, exploiting a Decision Tree Regressor trained on pose, force and torque datasets (DTs), the method predicts a certain amount of grasps on all bounding boxes of the decomposition. The DTs are originally acquired in "Grasp It Like a Pro" by a human equipped with the same end-effector as the final experiments by grasping a discrete number of cuboids. Once all the predictions of grasps are obtained, an index called Graspability Index is introduced by which the best grasping pose for an unknown object is selected. Finally, the method is validated through simulations and experiments using a Franka Panda EMIKA and Pisa/IIT SoftHand.