Neural Networks and Sliding Modes: Control, Observation, and Application to Robotics

The aim of this dissertation is to present and analyze several control and observation schemes that rely on the joint use of neural network and sliding modes. In particular, it presents a novel framework which exploits deep neural networks (DNN)s and integral sliding mode (ISM) to design control schemes able to control perturbed nonlinear systems with fully unknown dynamics. Differently from other methodologies present in the literature, the DNNs are not trained offline, but, inspired by the adaptive control framework, their weights are adjusted online while the system is being controlled via adaptation laws that are derived from Lyapunov stability analysis. Such a framework is then extended to the case in which the system must satisfy some state or input constraints, presenting three modifications: one which relies on a modified sliding variable, one that embeds model predictive control, and one that makes use of barrier functions. The joint use of DNNs and sliding modes is also explored in the domain of fault diagnosis (FD). In particular, two FD schemes are presented. The former relies on the aforementioned DNN based ISM framework to build an unknown input observer (UIO) for the estimation of fault affecting a system. As for the latter, it consists of a deep reinforcement learning (DRL) agent that aims to estimate the sensor fault affecting the joints of a robotic manipulator. Such an estimate is used to clear the faulted signal and build a battery of second order sliding mode UIOs that allows to estimate the actuator fault acting on the robot joints. Finally, application of DNN and sliding modes in the domain of physical human-robot interaction are presented. In particular, an adaptive version of the DNN based ISM control framework is developed to control a robotic manipulator so that it performs the so-called ergonomic handover, i.e., exchanges object with the human operator, adapting to her/his pose to reduce psychophysical stress. Moreover, a collision avoidance architecture that relies on convolutional neural networks for obstacle detection and ISM for obstacle avoidance is presented. The control and observation methodologies present in this dissertation have been theoretically analyzed and their validity is assessed in simulation or experimentally obtaining more than satisfactorily results. The experiments are performed on a real Franka Emika Panda robot, present in the Intelligent Robotics Lab at the University of Pavia.