Title: Robust Control of Contact-Rich Robots via Neural Bayesian Inference
Program: Doctor of Philosophy in Electrical and Computer Engineering
Advisor: Dr. Aykut Satici, Mechanical and Biomedical Engineering
Committee Members: Dr. John Chiasson, Electrical and Computer Engineering, Dr. Kurtis Cantley, Electrical and Computer Engineering, and Dr. Hao Chen, Electrical and Computer Engineering
In this work, we provide several data-driven control design frameworks for contact-rich robotic systems. These systems exhibit continuous and discrete state transitions, which makes them difficult to control with one policy. Hence, we design a control learning mechanism that provides a mixture of expert controllers parameterized by neural networks. This architecture also learns a gating network, which determines the control switching scheme based on the observed states. Additionally, we address the adverse effects of model uncertainties in the control of contact-rich robots. Lack of accurate environmental models can misrepresent the effects of contact forces on the system. Optimal policies designed from such models can lead to poor performance or even instability. In particular, we demonstrate the effects of system parameter uncertainties and measurement error on the overall performance of the system. Then, we leverage the stability properties of passivity-based control in conjunction with the robustness properties of Bayesian learning to infer stochastic controllers.