The rising levels of autonomy exhibited by complex cyber-physical systems have brought questions related to safety and adaptation to the forefront of the minds of controls and robotics engineers. Often, such autonomous systems are deemed to be safety-critical in the sense that failures during operation could significantly harm the system itself, other autonomous systems, or, in the worst-case, humans interacting with such a system. Complicating the design of control and decision-making algorithms for safety-critical systems is that they must cope with various degrees of uncertainty as they are deployed autonomously in increasingly real-world environments. These challenges motivate the use of learning-based techniques that can adapt to such uncertainties while adhering to safety-critical constraints.
The main objective of this dissertation is to present a unified framework for the design of controllers that learn from data online with formal guarantees of safety. Rather than using a controller trained on an a priori dataset collected offline that is then statically deployed on a system, we are interested in using real-time data to continuously update the control policy online and cope with uncertainties that are challenging to characterize until deployment. We approach the problem of designing such learning-based control algorithms for safety-critical systems through the use of certificate functions, such as Control Lyapunov Functions (CLFs) and Control Barrier Functions (CBFs), from nonlinear control theory. To this end, we first discuss how modern data-driven techniques can be integrated into traditional adaptive control frameworks to develop classes of CLFs and CBFs that facilitate the design of both controllers and learning algorithms that guarantee, respectively, stability and safety by construction. Next, we shift from the problem of safe adaptive control to safe reinforcement learning where we demonstrate how similar ideas from adaptive control can be extended to safely learn the value functions of optimal control problems online using data from a single trajectory. Finally, we discuss an extension of the aforementioned approaches to richer control specifications given in the form of temporal logic formulas, which provide a formal way to express complex control objectives beyond that of stability and safety. / 2025-05-30T00:00:00Z
| Identifer | oai:union.ndltd.org:bu.edu/oai:open.bu.edu:2144/46288 |
| Date | 30 May 2023 |
| Creators | Cohen, Max H. |
| Contributors | Belta, Calin A. |
| Source Sets | Boston University |
| Language | en_US |
| Detected Language | English |
| Type | Thesis/Dissertation |
Page generated in 0.0016 seconds