Learning-based control is a dynamic confluence of control theory, machine learning, and optimization, targeted at addressing the problem of data-driven sequential decision-making. This paradigm offers a compelling solution to ensure the stability and optimality of closed-loop dynamical systems, even in the presence of unknown system dynamics. However, the implementation of learning-based control algorithms, reliant on data collection through sampling and experiments, faces perturbations due to noisy data and inadequate samples. This raises a critical question: Can learning-based control algorithms converge to a suboptimal solution despite the perturbations? To tackle this question, we will introduce the techniques from advanced control theory, such as input-to-state stability (ISS), to guarantee the robustness of learning-based control algorithms. Our findings reveal that under some conditions, learning-based control algorithms can indeed converge to a small neighborhood of the optimal solution, as long as the perturbations are relatively small. Furthermore, the presentation will highlight the practical benefits of learning-based control through its application in the balance control of wheel-legged robots, showcasing its efficacy in managing model uncertainties and parameter variations.

Leilei Cui is a fifth-year Ph.D. candidate in the Department of Electrical and Computer Engineering at New York University, advised by Prof. Zhong-Ping Jiang. He received the B.S. degree in Automation from Northwestern Polytechnical University in 2016, and the M.S. degree in Control Science and Engineer from Shanghai Jiao Tong University in 2019. His research focuses on the intersection of reinforcement learning, control theory, and optimization, with applications in robotics, intelligent transportation and human sensory motor control.