With increase in the applications of autonomous systems, in both civilian and military domains, it has become increasingly important to provide formal guarantees on correctness of the control architecture used for the autonomy. It becomes even more crucial in the safety-critical applications of these systems, for example, when these systems operate in a close vicinity of humans. Furthermore, as these systems and the environment where they operate become more and more complex, the traditional control design methods become impractical. Neural networks, on the other hand, have shown promising results in the field of autonomy for complex systems, such as autonomous driving in unseen environments. However, we are far from formally verifying a control architecture that involves a neural network component. Researchers have started exploring various methods of verifying correctness of neural networks, but much work is yet to be done.

The goal of this research topic is to collect recent developments on the topic of formal verification methods for neural network (NN)-based control architectures as it provide one-of-a-kind resource for the researchers as well as the end-users of such control frameworks. The collection of articles in this research topic will then serve as the starting point for researchers looking for the recent advancements, available tools and technologies, and the open problems and challenges in the field of verification of NN-based controllers.

The research topic invites original contributions in terms of new theory, algorithm, methods, tools, or new ways of analysis to aid the existing formal verification tools. The scope of this research topic includes, but is not limited to, the following areas:

- Neural network verification methods with applications in robotics
- Statistical tools, such as conformal prediction, for probabilistic safety guarantees
- Novel training mechanisms and data-labeling methods for safety-critical robotics
- Computationally efficient formal verification tools for complex control architectures
- Reachability tools for neural network architectures used in control design
- Safety filter and certificate learning mechanisms addressing the challenge of generalization and/or scalability
- Bridging the sim2real gap in one-shot or few-shot learning methods

Are you interested in contributing a manuscript to this Research Topic?

Formal Verification of Neural Networks-Based Control Architecture for Safety-Critical Autonomous Systems