Autonomous vehicles have driven millions of miles on public roads, but even the simplest scenarios, such as a lane change maneuver, have not been certified for safety. As there is no systematic method to bound and minimize the risk of decisions made by the vehicle’s decision controller, the insurance liability of autonomous vehicles currently is entirely on the manufacturer. We have developed a tool for autonomous vehicle plan verification and execution across a variety of driving scenarios. The goal was to develop the safety foundations, scalable verification algorithms and tools for AV software. System-wide safety objectives are achieved in terms of:
+ discrete decisions (e.g. switch from lane follow to lane change),
+ low-level control (e.g. find and follow a feasible trajectory to change lanes),
+ incorporating machine learning with safety constraints,
+ incorporating energy constraints with perception while ensuring safety,
+ goal revision (e.g. choosing to satisfy some properties at the cost of others) and
+ incorporating it all on a testbed of 1/10 scale autonomous racing cars.