Specification-guided safe learning for robotic systems
J. Jiang, Y. Zhao, S. Coogan
Smarter Cyber Physical Systems: Enabling Methodologies and Applications, 2025
Abstract
This chapter considers the problems of verification and synthesis for robotic systems with respect to complex tasks. In particular, a class of problems will be considered in which uncertainties in both system dynamics as well as environmental perturbations result in in high risk of failure. Abstraction-based methods are introduced which allow for computationally tractable high-level task planning with formal guarantees on the probability of task satisfaction. Then, Gaussian process learning techniques are incorporated into the abstraction model to enable learning of the system and environmental uncertainties. Finally, control policy synthesis algorithms are introduced which allow the robot to safely traverse its environment, learning the uncertainties online until the task can be satisfied with sufficient guarantees.