Inverse Models for Trajectory Control Aided by Data, Machine Learning Models, and GPUs

Abstract

This dissertation investigates how the easy access to large amount of data and cheap computation power will benefit the usage of the inverse models for trajectory control. In general, model-based inversion methods have been used to achieve high precision trajectory tracking in the past, and iterative methods with inverse models tend to achieve some of the highest precision possible for output tracking. However, it is challenging to get plant models for many practical systems such as robots with complex environmental interactions. In this context, this thesis explores three scenarios using both theory and experiments: (1) the use of data-based inverse models for improving precision in iterative control, (2) identifying the type of observables needed to develop machine-learning-based (neural-net-based) inverse models to enable precise tracking, and (3) the use of inverse-models in improving performance of model predictive path integral control (MPPI) to regulate the trajectory, relying on parallel computation powered by graphic processing units (GPUs).