Safe Artificial Mental Models for Cyber-Physical Systems Learning

Recently, machine learning techniques related to large language models and reinforcement learning have witnessed tremendous advances.

However, to safely and robustly incorporate these techniques into complex cyber-physical systems—such as humanoid robotics, autonomous vehicles, or industrial automation—is non-trivial.

Specifically, large language models may return incorrect results, and training reinforcement learning algorithms on physical systems can lead to unsafe actions.

Moreover, even if a combined pipeline of large language models and reinforcement learning of cyber-physical systems would work, it is hard to explain why it acts in a certain way: the explainability problem of using machine learning on cyber-physical systems.

This project aims to develop a new theoretical foundation, robust design, and practical implementation for a new concept called artificial mental models, defined as domain-specific language programs.

The purpose is to enable (i) safe user interaction using large language models, (ii) safe actuation using efficient reinforcement learning, and (iii) explainable actions validated in a formal setting, resulting in an overall trustworthy learning-based cyber-physical system.