A Learning-Based Approach to Safety in Robotic Systems
Presentation | April 26 | 1-2 p.m. | 337B Cory Hall
Abstract: From drones to self-driving vehicles, the near future has promise of seeing a larger integration of robotic systems in civilian spaces. As these systems become more pervasive, the need for algorithms that can ensure their safety becomes paramount. Hamilton-Jacobi (HJ) reachability analysis, a tool from optimal control and differential game theory, has been useful for handling safety. The approach can also be extended to systems with uncertainty through a robust worst-case treatment. Unfortunately, the robust treatment can lead to restrictive behavior of the underlying system, and result in poor performance with respective to other objectives.
In this talk we begin by presenting the HJ framework, and demonstrating its utility in safety critical applications. We then propose methods for mitigating the restrictiveness of the traditional HJ framework, by incorporating data-driven methods that learn about the system's safety through observations. Inspired by the field of reinforcement learning (RL), we give two methods, a model-based method based on Gaussian processes (GPs) and a model-free method based on temporal difference learning (TD). Finally, we conclude the talk by introducing a novel formulation for HJ reachability analysis that is more amenable to these data-driven methods, which results in improved computation time.