Heisei Yonezawa, Ansei Yonezawa, Itsuro Kajiwara
This study introduces a curriculum-based continual learning framework to improve robust control of mechanical systems with multiple uncertainties, enhancing learning efficiency and sim-to-real transfer in applications like automotive powertrains.
Controlling mechanical systems with multiple uncertainties, like those found in automotive powertrains, is a challenging task. This research proposes a new approach that breaks down complex control problems into manageable tasks, allowing a learning system to handle different uncertainties one at a time. By gradually expanding the complexity of these tasks and using a model-based controller as a performance baseline, the system can learn more efficiently and effectively. The method has been successfully tested in real-world scenarios, showing promise for improving the robustness of control systems in industry.
