Year: 2025
Type: dissertation
Abstract: Autonomous systems operating in the real world must stay live and safe. Here, liveness denotes its ability to complete tasks, e.g., reaching a goal or tracking a desired output signal, while safety means satisfying constraints. Typical tasks include real-time navigation in a priori unknown environment, and multi-agent collision avoidance with humans involved. Further, it becomes more challenging when there are unexpected (unmodeled) disturbances in the environment. Control Lyapunov functions (CLFs) have been widely used to solve the liveness problem by stabilizing the system to its equilibrium point, while control barrier functions (CBFs), which guarantee forward control invariance, are used to maintain safety. However, finding a valid CLF or CBF is hard, especially for high-dimensional nonlinear systems with state and/or input constraints. Hamilton Jacobi reachability (HJR) analysis formulates the liveness and safety problems as optimal control problems and has been used in aerospace, autonomous driving, and more. It computes a value function whose level sets provide liveness and safety information, and the gradient informs the liveness and safety controller. This value function can be obtained numerically on a discrete grid using dynamic programming, and therefore suffers from the ``curse of dimensionality.''The first chapter of the dissertation proposes the robust control Lyapunov-value function (R-CLVF), which is a Lipschitz continuous function with CLF properties, and can be computed using numerical tools from HJR analysis. The R-CLVF can exponentially stabilize nonlinear systems with bounded control and disturbance to any point of interest if possible, and to a robust control invariant set around the point of interest otherwise. The second chapter focuses on the system decomposition, which is one potential solution to the ``curse of dimensionality'': the value functions for low-dimensional subsystems are computed first, and then the value function for the original system is reconstructed. Exact reconstruction can be achieved under certain conditions. The third chapter focuses on the reach-avoid-stay problem, which combines the time-optimal property of HJR and the infinite-time stability of the R-CLVF. The last chapter applies the previous theories to the navigation task, providing frameworks that can survive unmodeled disturbance and adapt the speed based on the environment.
Author: Gong, Zheng
Institutions: Duke University, University of California San Diego
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7070782655
