MODEL PREDICTIVE CONTROL OF SKID-STEERED MOBILE ROBOT WITH DEEP LEARNING SYSTEM DYNAMICS

Abstract

This thesis project presents several model predictive control (MPC) strategies for control of skid-steered mobile robots (SSMRs) using two different combinations of software environment, optimization tool and machine learning framework. The control strategies are tested in WeBots simulator. Spatial-based path following MPC of SSMR with static obstacle avoidance is developed in MATLAB environment with ACADO optimization toolkit using spatial kinematic model of SSMR. It includes static obstacle and border avoidance strategy based on artificial potential fields. Simulations show that the controller is effective at driving SSMR on a track, while avoiding borders and obstacles. Several more MPCs are developed using Python environment, ACADOS optimisation framework, and Pytorch-Casadi integration framework. Two time-domain controllers are made in Python environment, one based on SSMR kinematic model and another based on data-driven state-space model using Pytorch-Casadi framework. Both are setup to reach a goal point in simulation experiment. Experiments show that both versions reliably reach a target point. Standard and data-driven versions of spatial path following MPC are developed. Standard is a reimplementation of MPC designed in MATLAB with modifications to cost function and border avoidance, without static obstacle avoidance. Data-driven path following MPC is an extension of standard variant with state-space model replaced with a hybrid of spatial kinematics and data-driven model. Simulation of both spatial controllers confirm their effectiveness in following reference path.