Control-oriented learning for robotics and dynamical systems
Abstract/Contents
- Abstract
- Data-driven learning for robotics is ultimately done in the interest of informing decision-making and control, particularly in unseen and unstructured environments. This thesis is centered on the concept of control-oriented learning, whereby learning is attuned to this desired goal, rather than to naively fit models to data. To this end, this thesis presents novel methodologies that tailor model learning towards closed-loop feedback control for robotic systems. In Part I of this thesis, we focus on the offline learning setting. We leverage nonlinear control theory to regularize data-driven model fitting optimization problems to be better aligned with the task of trajectory tracking. In Chapter 2, we propose to do this by adding constraints to the learning problem that encourage the learned model to be stabilizable by jointly learning a dynamics model and a control-theoretic certificate which guarantees the existence of robust tracking controllers for arbitrary open-loop trajectories generated with the learned model. In Chapter 3, we take a different approach by embedding intrinsic factorized structure into our learned dynamics model that engenders the immediate synthesis of a performant tracking controller. Whether through constraints or model structure, we demonstrate that imbuing the offline learning problem with control-theoretic notions results in a model that performs better in closed-loop control online. In Part II of this thesis, we shift our focus to the online learning setting through the lens of adaptive control. In Chapter 4, we propose to learn a feature basis for adapting dynamics models offline through differentiable closed-loop simulation of classical adaptive control laws with parametric dynamics models. Ultimately, this takes the form of a meta-learning framework that links gradient descent on a trajectory tracking performance metric with a parametric representation of any unknown dynamics residual and feedback gains defining the adaptive control law. In Chapter 5, we generalize classical adaptive control concepts, namely stable concurrent estimation and certainty-equivalent cancellation of uncertain dynamics terms, to the robust model predictive control setting. Our approach allows us to couple classic online adaptation schemes, parameter uncertainty bounds, and robust model predictive control feedback that maintains safety through state and input constraint satisfaction, while decreasing conservatism in our uncertainty bounds over time. We conclude this thesis with a discussion of possible future avenues of work towards better synergy between learning modules in the autonomy stack and online closed-loop deployment objectives.
Description
Type of resource | text |
---|---|
Form | electronic resource; remote; computer; online resource |
Extent | 1 online resource. |
Place | California |
Place | [Stanford, California] |
Publisher | [Stanford University] |
Copyright date | 2023; ©2023 |
Publication date | 2023; 2023 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Richards, Spencer Mathew |
---|---|
Degree supervisor | Pavone, Marco, 1980- |
Thesis advisor | Pavone, Marco, 1980- |
Thesis advisor | Boyd, Stephen P |
Thesis advisor | Schwager, Mac |
Degree committee member | Boyd, Stephen P |
Degree committee member | Schwager, Mac |
Associated with | Stanford University, School of Engineering |
Associated with | Stanford University, Department of Aeronautics and Astronautics |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Spencer M. Richards. |
---|---|
Note | Submitted to the Department of Aeronautics and Astronautics. |
Thesis | Thesis Ph.D. Stanford University 2023. |
Location | https://purl.stanford.edu/jv434vz2493 |
Access conditions
- Copyright
- © 2023 by Spencer Mathew Richards
- License
- This work is licensed under a Creative Commons Attribution 3.0 Unported license (CC BY).
Also listed in
Loading usage metrics...