From model-based to data-driven control: applications to self-balancing robots

Details

Supervisors

Bianchin, Gianluca

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électromécanicien, à finalité spécialisée: mécatronique

Abstract

Data-driven control has emerged as a promising paradigm for dynamical systems, enabling the construction of feedback controllers directly from historical data without the need for system model identification. This paper investigates the practical application of data-driven control by comparing it with traditional model-based control methods on a real balancing robot, the Pololu Balboa 32U4. The study confirms the unstable nature of the robot through mathematical modeling and underscores the need for active control to stabilize its motion. Model-based control methods, including LQR, Bessel, and ITAE pole placement, ensure stability across simulations and experiments. However, a noticeable steady-state error in reference tracking indicates potential for improvement, particularly in navigation-oriented applications. Data-driven control shows promise for self-balancing applications, producing promising results in simulations. In experimental contexts, these methods are effective in scenarios where data is gathered using slower controllers that manage natural disturbances. However, their effectiveness decreases with faster controllers due to inaccuracies in sensor data measurement and significant system noise, which compromise the integrity of the data-driven results. The research identifies avenues for future exploration to enhance stability and the quality of data-driven results. These include implementing sensor fusion, using model-based Kalman filters, or exploring alternative non-model-based filtering techniques. These strategies aim to advance the data-driven approach for the Balboa 32U4, paving the way for enhanced stability and performance in real-world applications.