Simulation Driven Design and Test for Safety of AI Based Autonomous Vehicles

An autonomous vehicle (AV) integrates sophisticated perception and localization components to create a model of the world around it, which is then used to navigate the vehicle safely. Machine learning (ML) based models are pervasively used in these components to extract object information from noisy sensor data. The requirements for these components are primarily set to achieve as high accuracy as possible. With modern AVs deploying many sensors (cameras, radars, and LiDARs), processing all the data in real-time leads to engineers making trade-offs which might result in a sub-optimal system in certain driving situations. Due to the lack of precise requirements on individual components, modular testing and validation also becomes challenging. In this paper, we formulate the problem of deriving abstract world model accuracy needed for safe AV behavior from top level driving scenario simulations. This is computationally expensive as the world model can contain many objects with several attributes and an AV extracts a world model every time-step during the simulation. We describe approaches to efficiently address the problem and derive component-level requirements and tests.