• Few Shot Learning
• Human-Inspired Robotic Grasp
• Graph Neural Networks
• Grasp Learnig

Althoguh there exist an extensive literature on robotic grasping, many problems are still to be solved, especially when dealing with the interaction of robots with unstructured environments, where it is important to provide a method to learn how to generalize grasping and manipulation skills to unknown classes of objects. To address this goal, in this work I proposed a novel method for 2D vision driven grasping of objects which allows few shot learning of new human-inspired grasp primitives, relying on state-of-the-art human grasp taxonomy. This goal was achieved by means of a Gated Graph Neural Network (GGNN) and a planning module. The former allowed the embedding of the human example in the form of a Knowledge Graph that encodes the information about the relationships between grasps in a given taxonomy. The latter allowed the detection of an object in the RGB camera field of view and planned a grasp based on the extracted 2D information and the output of the GGNN.
The pipeline consists of the following three modules: Classification Module (the GGNN), which selects the grasp to be performed, Planning Module, which synthesizes a grasp strategy, Control Module, which physically implements the strategy.
The model was tested on a Franka manipulator endowed with an anthropomorphic soft underactuated end effector. The framework showed promising results with a 69% success rate in grasping execution and it was able to successfully integrate new strategies while retaining previously learned information, outperforming an ideal model provided with samples from all the classes at training time.