No abstract

Background: Autonomous cars could make traffic safer, more convenient, efficient and sustainable. They promise the convenience of a personal taxi, without the need for a human driver. Artificial intelligence would operate the vehicle instead. Especially deep neural networks (DNNs) offer a way towards this vision due to their exceptional performance particularly in perception. DNNs excel in identifying objects in sensor data which is essential for autonomous driving. These networks build their decision logic through training instead of explicit programming. A drawback of this technology is that the source code cannot be reviewed to assess the safety of a system. This leads to a situation where currently used methods for regulatory approval do not work to validate a promising new piece of technology. Objective: In this paper four approaches are highlighted that might help understanding black box technical systems for autonomous cars by focusing on its behaviour instead. The method of experimental psychology is proposed to model the inner workings of DNNs by observing its behaviour in specific situations. It is argued that penetration testing can be applied to identify weaknesses of the system. Both can be applied to improve autonomous driving systems. The shadowing method reveals behaviour in a naturalistic setting while ensuring safety. It can be seen as a theoretical driving exam. The supervised driving method can be utilised to decide if the technology is safe enough. It has potential to be developed into a practical driving exam.

No abstract