Can we make robots learn motor skills in the real world safely and efficiently?
Based on mathematical descriptions of real-world problems, i.e., models, iterative control algorithms can provide safe and efficient solutions to those problems in simulations. Learning from data can handle unmodeled dynamics and unstructured situations. The integration of learning and iterative control algorithms will lead to intelligent robots learning motor skills in practice, toward becoming extensions of both the human body and the human brain.
In this talk, I will discuss my work on integrating learning into control to teach robots motor skills. I will focus on two types of control algorithms, model predictive control with iterative linear-quadratic-regulator for trajectory planning and iterative learning control for trajectory tracking, and how they can be applied for snake robot locomotion. I am interested in how to learn motor skills in a way that is generalizable across individual robots, and how to use models and data under different circumstances.