Automation and Robotics

Unmanned vehicles, drones, self-driving cars and other sorts of advanced autonomous vehicles are being announced on an almost daily basis. Uber is working on fl ying taxis, every car company has a self-driving car in the works, and drones are the hottest Christmas toy for people of all ages. Inside these, autonomous vehicles are systems based on advanced artifi cial intelligence, including artifi cial neural networks (ANN), machine learning based systems, probabilistic reasoning, and montecarlo models providing support for complex decision making. One of the common concerns about autonomous vehicles, be they fl ying or driving, is for safety. Safety testing is usually based on deterministic behavior, the aircraft or car or boat, which faced a similar situation, behaves the same way every time. But, what happens when the vehicle is learning from its environment, just as we humans do. Th en it may behave diff erently each time based on experience. How then to predict and evaluate in advance? How safe an autonomous system might be? Th is paper will present two complementary approaches to this problem. One is a stochastic model for predicting how an autonomous system might behave as it learns over time, providing a range of behavioral responses, to be used as a risk assessment tool? Th e other is a set of methods and standards for writing test procedures for such vehicles.