Learning-based Control for Autonomous Mobile Robots

Abstract

As mobile robots leave structured indoor environments to operate in challenging outdoor environments, their motion controllers require advanced techniques to mitigate the effects of unmodelled surface materials (e.g., snow, sand, grass), terrain topography (e.g., side-slopes, inclines), and complex robot dynamics. This thesis investigates learning-based control within the context of path-tracking autonomous mobile robots. Learning-based algorithms alleviate the need for significant engineering work in identifying and modelling all disturbances that a controller may be required to mitigate. Furthermore, they are capable of predicting and acting in anticipation of repeatable effects and disturbances not modelled prior to deployment. The learning-based algorithms reduce tracking errors, increase operational speed, and increase localization reliability. Specifically, the thesis presents four approaches: 1) Iterative Learning Control (ILC), 2) Learning-based Nonlinear Model Predictive Control (LB-NMPC), 3) Robust Min-Max LB-NMPC (MM-LB-NMPC), and 4) Robust Constrained LB-NMPC (RC-LB-NMPC). ILC generates an acausal feedforward signal that reduces the path-tracking errors using information from any previous trial. While the approach is computationally appealing, ILC typically assumes that the robot is initialized with identical initial conditions for each trial and tracking the same desired path (i.e., generalization is non-trivial). On the other hand, LB-NMPC is a technique that uses a learned process model directly, enabling interpolation and extrapolation from experiences. In this case, the current control action is obtained by solving a finite-horizon optimal control problem using the current state of the plant as the initial state at each time-step. Finally, this thesis investigates two recent results in Robust NMPC in order to guarantee controller stability throughout the learning process in spite of model uncertainty. For MM-LB-NMPC, the control problem is altered to optimize for plausible worst-case scenarios. For RC-LB-NMPC, tightened constraints are applied to nominal predictions such that all plausible predicted sequences satisfy the given constraints. The resulting RC-LB-NMPC algorithm is a robust, learning controller providing safe, conservative control during initial trials when model uncertainty is high and converging to high-performance, optimal control during later trials when model uncertainty is reduced with experience.