Position Paper

Exhaustive state space exploration based verification of cyber-physical system designs remains a challenge despite five decades of active research into formal verification. On the other hand, models of intelligent automotive cyber-physical systems continue to grow in complexity. The testing of intelligent automotive models often uses human subjects, is expensive, and can not be performed unless the system has already been prototyped and is ready for human interaction. We propose the use of machine learning methods to learn stochastic models of human-vehicle interaction.

Automotive cyber-physical systems will need to address self-parking, advanced steering control, hazardous situation recovery, limited autonomous driving, and even more complex tasks in the coming decades. Verification of the safe behavior of these tasks for multiple vehicle configurations (weight, wheelbase, front/rear/all-wheel drive, etc.) will require significant advancements in the computational theory, as well as new approaches to compose behaviors and computational constraints with hybrid control theory and system modeling.

The demand for higher performance computing platforms has dramatically increased during the last decade due to the continuous feature enhancement process. For instance, in automotive systems new safety features like `night view assist’ and `automatic emergency breaking’ require the fusion of sensor data, video processing and real-time warnings when an obstacle is detected on the road; in the avionics domain new applications such as the helmet-mounted display systems require intensive video processing capabilities.