Grasping learning, optimization, and knowledge transfer in the robotics field.

Service robotics is a fast-developing sector, requiring embedded intelligence into robotic platforms to interact with the humans and the surrounding environment. One of the main challenges in the field is robust and versatile manipulation in everyday life activities. An appealing opportunity is to exploit compliant end-effectors to address the manipulation of deformable objects. However, the intrinsic compliance of such grippers results in increased difficulties in grasping control. Within the described context, this work addresses the problem of optimizing the grasping of deformable objects making use of a compliant, under-actuated, sensorless robotic hand. The main aim of the paper is, therefore, finding the best position and joint configuration for the mentioned robotic hand to grasp an unforeseen deformable object based on collected RGB image and partial point cloud. Due to the complex grasping dynamics, learning-from-simulations approaches (e.g., Reinforcement Learning) are not effective in the faced context. Thus, trial-and-error-based methodologies have to be exploited. In order to save resources, a samples-efficient approach has to be employed. Indeed, a Bayesian approach to address the optimization of the grasping strategy is proposed, enhancing it with transfer learning capabilities to exploit the acquired knowledge to grasp (partially) new objects. A PAL Robotics TIAGo (a mobile manipulator with a 7-degrees-of-freedom arm and an anthropomorphic underactuated compliant hand) has been used as a test platform, executing a pouring task while manipulating plastic (i.e., deformable) bottles. The sampling efficiency of the data-driven learning is shown, compared to an evenly spaced grid sampling of the input space. In addition, the generalization capability of the optimized model is tested (exploiting transfer learning) on a set of plastic bottles and other liquid containers, achieving a success rate of the 88%.