Project 1: Assured autonomous decision-making with machine learning

Next-generation autonomous systems (e.g., drones and robots) should operate smartly in contested environments. They should be able to get high-level instructions from humans and automatically create control policies to carry them out even under unknown environmental disturbances and adversarial attacks. The high-level instructions may vary depending on the application and can contain numerous (sub)tasks that are logically dependent on each other. For example, a service robot may need to investigate different terrains in differential time constraints while remaining safe from possible hazards. This project developed general-purpose algorithms to create control policies automatically, applicable to a wide class of autonomous systems applications. These algorithms can understand any high-level human instructions expressable by a class of formal logical formulas and create optimal control policies from machine learning. The use of machine learning ensures that high-level instructions are fulfilled even in the presence of unknown environmental disturbances and adversarial attacks. The proposed algorithms are implemented by open-source software tools, which can successfully solve control problems for service and industrial robots.

Project 2: Data-driving verification of cyber-physical systems

Modern control systems rely on computer programs on cyber devices (e.g., embedded computers) to control physical plants (e.g., drones or robots) to perform various complex tasks. These computer programs are typically developed in ideal simulated environments and tend to fail in real physical environments due to disturbances and attacks that are unknown beforehand. To fill this sim-to-real gap, engineers must verify whether the controlled systems can still fulfill the tasks under the possible disturbances and attacks. Previously, such a process is performed case-by-case with high costs of human labor. This project developed general-purpose data-driven verification methods to verify a wide class of controlled systems designed for different tasks. The proposed methods are based on machine learning and logical reasoning from the controlled system paths. They can provide provable statistical guarantees on the result without knowing the system’s dynamic model. The project also implemented the proposed methods with open-source software tools and demonstrated its superior efficiency over existing methods on a wide range of benchmarks.