Formal Design of Semi-autonomous Cyber-Physical Transportation Systems

Abstract:

The goal of this project is to develop fundamental theory, computationally efficient algorithms, and real- world experiments for the analysis and design of safety-critical cyber-physical transportation systems with human operators. We envision a nearby future in which roads will be populated by networks of smart vehicles that will cooperate with each other, with the surrounding infrastructure, and with their drivers to make transportation safer, more enjoyable, and more efficient. To reach this vision, there is an urgent need for modeling, analysis and design techniques for provably safe cyber-physical transportation systems with human-in-the-loop. We propose to meet this need through an ambitious modeling, theoretical, and experimental collaborative effort combining human factors, control theory, and computer science. As crashes at traffic intersections account for about 40% of overall vehicle crashes, we will focus on intersection crashes in this project. Specifically, our grand-challenge application is to design provably safe driver-assist systems that understand driver's intentions and provide warnings/overrides to prevent collisions at traffic intersections. With this focus, we propose to construct, from four human factors experiments hybrid automata models for the driver-vehicles- intersection system, which incorporate driver behavior and performance as an integral part. Due to the probabilistic nature of driver behavior, we propose to construct a partial order of these hybrid automata models, ordered according to confidence levels on the model parameters. These hybrid models will have imperfect state information because of uncontrollable and unobservable driver's decisions, sensor noise, and communication limitations. We propose to formulate the driver-assist design problem as a set of partially ordered hybrid differential games with imperfect information, in which games are ordered according to parameters confidence levels. This novel approach to address safety specifications allows to formally establish a tradeoff between conservatism of the design and safety confidence. This is especially crucial for driver assist systems, in which the frequency of warnings and overrides should be carefully tuned based on driver's expectations, government regulations, and industrial and international safety standards. We propose to validate our designs experimentally in the UMTRI driving simulator and in large-scale computer simulations leveraging the software developed by the SimMobility project at MIT.