Learning Task Oriented Grasp for objects of common use

Recent advances in Robotics research are pushing the boundaries of machines applicability more and more toward the deployment of autonomous dexterous robots in our everyday life. However, this comes with significant conceptual and practical difficulties related to the interaction with the humans and the surrounding word. Of note, the latter is shaped to be functional to its usability by humans: just think of objects of daily living, that are made to afford a specific grasp type for a selection of possible tasks. However, state of the art method usually limit the robot-objects interactions to a mere pick and place of objects, while little has been done to learn how to grasp an object depending on the envisioned task. Yet, a human that is grasping a knife would implement a completely different strategy depending if the intention is to cut a meal or just pass it to a fellow. This thesis represents a step toward the development of a neural architecture capable of identifying the proper grasp strategy that a robot should implement considering not only the shape and the position of the object, but also the commanded task. To this purpose, as a first contribution of this work, the OCID dataset has been extended to provide also grasp-task pairs for each object in the image: the dataset extension allowed to perform task oriented grasping in extremely difficult circumstances, as each image in the dataset contain several objects. Then, state-of-the-art architectures have been adapted to predict task oriented grasps by developing a trainable CNN-based architecture that can produce high-quality grasp detection for a parallel-plate gripper. Particular attention was given to the reduction of the network footprint and of the inference time of state of the art architectures, obtaining a high-performing task oriented grasping proposal generator that relies on 2D data only.