ADEX - Autonomous Driving Examiner

The ADEX project aims to develop an examination method for automated driving. At the core of this approach is an autonomous examiner, i.e., a programme unrelated to the specific design of the autonomous-driving controller being tested.

Short Description

Smart mobility, using highly automated vehicles, is a top priority for our modern society. It promises the ability to dramatically reduce accidents and fatalities, as well as pollution. At the same time, it also represents one of the greatest technological challenges of our time. While much progress has been made over the past few years, accidents involving automated vehicles clearly reveal the weaknesses in the current state-of-the-art of autonomous driving. The use of machine learning components and the operation of automated vehicles in complex environments, involving pedestrians, cyclists and human drivers, gives rise to serious safety concerns.

This raises the question of whether we will ever be able to fully trust autonomous-driving controllers. To answer this question, we should first ask ourselves how society has developed trust in human drivers, who must demonstrate their abilities in a driving test before they are allowed to take to the road.

The driving test consists of a large set of realistically simulated traffic situations,or scenarios, incorporating human behaviour models for pedestrians, cyclists and drivers of other vehicles, and for different weather and road conditions. These scenarios will contain both normal traffic situations and challenging edge case scenarios which allow hidden problems to be detected.

Both the concept of an autonomous examiner for automated driving and the manner in which the examiner is synthesised are highly innovative. The architecture of the proposed solution:

  • Real-world traffic accidents at hot spots are analysed using holistic accident analysis and used for synthetically creating new, realistic, and critical traffic scenarios and driving sequences for the test.
  • Actions of the autonomous-driving controller are thoroughly quantified in the form of rewards which are used by reinforcement learning algorithms to generate increasingly complex traffic situations.
  • In order to assess the behaviour of the autonomous-driving controller and create appropriate rewards, several criteria including safety, traffic regulations, ethical aspects of decision making and passenger comfort are taken into account.

As is the case for human drivers, we, as a society, will have greater trust in the reliability of autonomous-driving controllers which successfully pass this test, making us more willing to accept such systems on our roads. Failure to pass the test, on the other hand, will reveal problems that still exist with autonomous driving and provide engineers with valuable information for improvement.

Project Partners

Consortium lead

AIT Austrian Institute of Technology GmbH

Other consortium partners

  • AVL List GmbH
  • Vienna University of Technology (TU Wien)
  • SV Univ.-Prof. DI Dr. Ernst PFLEGER

Contact Address

AIT Austrian Institute of Technology GmbH
Dr. Dejan Ničković
Giefinggasse 4
1210 Vienna