Robot Planning and Control combined with Machine Learning Techniques

Robots and Artificial Intelligence (AI) have been growing in various contexts, such as automotive, healthcare and manufacturing. In this thesis, the adoption of Machine Learning (ML) and control techniques is explored to solve relevant tasks in the robotics field. The focus is on the task of retrieving an object target from cluttered environment; as it is well-known, the problem is combinatorial and hard to solve in reasonable time. Here, it is solved by using a two layers architecture characterized from an high level Task Planner (TP) made of a Reinforcement Learning (RL) agent, combined to a low level based on Motion Planner (MP) and Inverse Kinematics (IK) Control. The architecture is validated via simulation and using the KINOVA Jaco2 7-DoFs robot manipulator. During the computation of the path from the Motion Planner, it could be useful to check the collision quickly. Thus, the collision detection problem in unstructured environment by using a learning-based approach is considered. The idea is to present an architecture that could reduce the computational time spent by the planner. Going into detail, it is addressed employing depth images and point clouds for adapted neural networks, i.e. CNN and PointNet. The proposed approach is validated with an industrial robot at the Technology & Innovation Center (TIC) of KUKA Deutschland GmbH in Augsburg (Germany). Furthermore, also an optimization in the control law is considered. When a robot manipulator is redundant, it is possible to exploit the additional degree of freedoms for maximizing different functionals into the null space of the Jacobian matrix, e.g. maximization of manipulability and distance from joint limits. The idea is to show that, through a Supervised Learning (SL) approach, it is possible to enlarge the dextrous workspace of the robot.