Assuring Safety under Uncertainty in Learning-Based Control Systems

Citation

Cheng, Richard (2021) Assuring Safety under Uncertainty in Learning-Based Control Systems. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/9kye-rn93. https://resolver.caltech.edu/CaltechTHESIS:01052021-195655093

Abstract

Learning-based controllers have recently shown impressive results for different robotic tasks in well-defined environments, successfully solving a Rubiks cube and sorting objects in a bin. These advancements promise to enable a host of new capabilities for complex robotic systems. However, these learning-based controllers cannot yet be deployed in highly uncertain environments due to significant issues relating to learning reliability, robustness, and safety.

To overcome these issues, this thesis proposes new methods for integrating model information (e.g. model-based control priors) into the reinforcement learning framework, which is crucial to ensuring reliability and safety. I show, both empirically and theoretically, that this model information greatly reduces variance in learning and can effectively constrain the policy search space, thus enabling significant improvements in sample complexity for the underlying RL algorithms. Furthermore, by leveraging control barrier functions and Gaussian process uncertainty models, I show how system safety can be maintained under uncertainty without interfering with the learning process (e.g. distorting the policy gradients).

The last part of the thesis will discuss fundamental limitations that arise when utilizing machine learning to derive safety guarantees. In particular, I show that widely used uncertainty models can be highly inaccurate when predicting rare events, and examine the implications of this for safe learning. To overcome some of these limitations, a novel framework is developed based on assume-guarantee contracts in order to ensure safety in multi-agent human environments. The proposed approach utilizes contracts to impose loose responsibilities on agents in the environment, which are learned from data. Imposing these responsibilities on agents, rather than treating their uncertainty as a purely random process, allows us to achieve both safety and efficiency in interactions.

Thesis Files