Abstract
Computer systems are increasingly coming to be relied upon to augment or replace human operators in controlling mechanical devices in contexts such as transportation systems, chemical plants, and medical devices, where safety and correctness are critical. A central problem is how to verify that such partially automated or fully autonomous cyber-physical systems (CPS) are worthy of our trust. One promising approach involves synthesis of the computer implementation codes from formal specifications, by software tools. This project contributes to this "correct-by-construction" approach, by developing scalable, automated methods for the synthesis of control protocols with provable correctness guarantees, based on insights from models of human behavior. It targets: (i) the gap between the capabilities of today's hardly autonomous, unmanned systems and the levels of capability at which they can make an impact on our use of monetary, labor, and time resources; and (ii) the lack of computational, automated, scalable tools suitable for the specification, synthesis and verification of such autonomous systems.
The research is based on study of modular reinforcement learning-based models of human behavior derived through experiments designed to elicit information on how humans control complex interactive systems in dynamic environments, including automobile driving. Architectural insights and stochastic models from this study are incorporated with a specification language based on linear temporal logic, to guide the synthesis of adaptive autonomous controllers. Motion planning and other dynamic decision-making are by algorithms based on computational engines that represent the underlying physics, with provision for run-time adaptation to account for changing operational and environmental conditions. Tools implementing this methodology are validated through experimentation in a virtual testing facility in the context of autonomous driving in urban environments and multi-vehicle autonomous navigation of micro-air vehicles in dynamic environments. Education and outreach activities include involvement of undergraduate and graduate students in the research, integration of the research into courses, demonstrations for K-12 students, and recruitment of research participants from under-represented demographic groups. Data, code, and teaching materials developed by the project are disseminated publicly on the Web.
Performance Period: 09/01/2015 - 09/30/2018
Institution: University of Texas at Austin
Sponsor: National Science Foundation
Award Number: 1550212