Methods for quantifying, representing, and utilizing uncertainty in learning-enabled autonomy
Abstract/Contents
- Abstract
- In order to leverage the advances of machine learning (ML) to enable reliable robot autonomy in the unpredictable and unstructured real world, we must design ML models that can effectively quantify and represent uncertainty, and design planning algorithms that can leverage these estimates of uncertainty to ensure safe performance during deployment. However, standard ML methods often fail to effectively capture key sources of uncertainty that arise when deploying robots in the real world; in particular, the environmental uncertainty arising from partial observability, and the epistemic uncertainty arising from limited training data from test-time conditions. In the first part of this thesis, we develop tools for quantifying and representing uncertainty in ML models. Our algorithms combine a Bayesian perspective on uncertainty quantification with the expressive modeling capabilities of deep neural networks to yield efficient and dynamic representations of uncertainty. In the second part, we address how a learning-enabled autonomy stack should leverage uncertainty estimates when planning actions to take. In particular, we explore the notion of optimality when considering risk while planning with model uncertainty, and propose a Monte-Carlo planning algorithm which introduces a notion of robustness onto the classic explore-exploit tradeoff of reinforcement learning. We also extend these ideas to the continuous control setting, where we develop an approach for planning and learning in uncertain environments while maintaining probabilistic guarantees on safety, and demonstrate this approach in hardware. We conclude with a discussion of next steps and a broader discussion of the role of uncertainty modeling in learning-enabled autonomy, including key challenges and opportunities for future research.
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 | 2022; ©2022 |
| Publication date | 2022; 2022 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Sharma, Apoorva |
|---|---|
| Degree supervisor | Pavone, Marco, 1980- |
| Thesis advisor | Pavone, Marco, 1980- |
| Thesis advisor | Finn, Chelsea |
| Thesis advisor | Kochenderfer, Mykel J, 1980- |
| Degree committee member | Finn, Chelsea |
| Degree committee member | Kochenderfer, Mykel J, 1980- |
| Associated with | Stanford University, Department of Aeronautics and Astronautics |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Apoorva Sharma. |
|---|---|
| Note | Submitted to the Department of Aeronautics and Astronautics. |
| Thesis | Thesis Ph.D. Stanford University 2022. |
| Location | https://purl.stanford.edu/rv570gt1642 |
Access conditions
- Copyright
- © 2022 by Apoorva Sharma
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...