Reinforcement learning-based motion control for tooth belt drive systems

Abstract

This thesis focuses on improving motion control in tooth belt drive systems characterized by nonlinear and flexible dynamics. Traditional controllers, such as cascade PID-type controllers, are commonly used but often face limitations in these types of systems due to their complete reliance on fixed parameters and simplified linear models used for their tunning. To address this problem, the thesis analyzes the use of reinforcement learning (RL), specifically the twin delayed deep deterministic policy gradient (TD3) algorithm, as an alternative control approach. In the proposed method, the RL agent replaces the conventional cascaded control loop (outer P-type position and inner PI-type speed controller). A detailed Simulink model obtained from LUT University is used to test and validate the approach through simulations.

Additionally, the study investigates the reward function design and its influence on the training performance and final control behavior. The results show that the RL-based controller adapts effectively to different system conditions and operates successfully without manual tuning, making it a suitable approach to address the identified challenges. The results show that the RL-based controller adjusts well to different system conditions and works effectively without manual tuning, making it a suitable option to overcome the mentioned problems correctly.