Researchers develop adaptive control method to automatically generate autonomous driving simulation scenarios

Semi-autonomous vehicles equipped with automatic systems are already on the road. As a core component of the vehicle, the software of these systems must continue to reliably meet high quality standards. According to foreign media reports, Franz Wotawa and his team from the Institute of Software Technology at Graz University of Technology (TU Graz) cooperated with the AVL cyber-physical system testing team to use adaptive control methods to automatically generate a wide range of simulation test scenarios and system internal error compensation to ensure the safety of autonomous driving.

(Image source: TU Graz)

Relying solely on test driving does not provide sufficient guarantee for the safety of autonomous driving systems. Franz Wotawa explained, "Autonomous driving vehicles must travel about 200 million kilometers to prove their reliability, especially in the event of an accident, which is 10,000 times more than the test kilometers of traditional cars. However, critical test scenarios that are life-threatening cannot be reproduced in real test driving. Therefore, the safety of autonomous driving systems must be tested through simulation."

Wotawa said, "Although current tests cover many scenarios, the question is whether these tests are sufficient and whether all possible accident scenarios are considered." Mihai Nica from AVL emphasized, "In order to test highly automated driving systems, it is necessary to rethink how the automotive industry verifies and certifies ADAS and automated driving systems. To this end, AVL and TU Graz have jointly developed a unique and efficient simulation- and test case generation-based approach and workflow to prove the intended functional safety (SOTIF), quality and system integrity requirements of automated systems."

The project team is working on developing innovative methods to simulate more test scenarios. The researchers' approach uses ontologies to describe the environment of the autonomous vehicle without having to drive millions of kilometers. An ontology is a knowledge base for the exchange of relevant information in a machine system. For example, the interfaces, behaviors and relationships of individual system units can communicate with each other. In an autonomous driving system, these are decision-making, traffic descriptions or autonomous driving (autopilot). The Graz researchers use the basic details of the driving scenario environment and enter the detailed information provided by AVL about road construction, intersections, etc. into the knowledge base. By using AVL's test case generation algorithm, the behavior of the autonomous driving system can be tested in simulation and thus the driving scenarios can be generated.

As part of the EU AutoDrive project, the researchers used two algorithms to convert ontologies into combinatorial test input models, which can then be performed using a simulation environment. "In initial experimental tests, we found serious flaws in the automated driving functions," the researchers said. "Without these automatically generated test scenarios, the vulnerabilities would not have been discovered quickly. Nine of the 319 test cases examined led to accidents." For example, in one test scenario, the brake assist system failed to detect two people coming from different directions at the same time, one of whom was hit by the braking operation. "This means that using our method, test scenarios that are difficult to test in reality can be discovered," said Wotawa.

In the event of a malfunction or a change in environmental conditions, automated systems, especially automated driving systems, must be able to correct themselves and reliably reach a given target state at any time. Franz Wotawa explains, "With current semi-automated systems, such as cruise control, the driver can intervene when the system fails. With fully automated vehicles, there is no need for driver intervention, so the system must be able to cope with a wide range of situations."

Franz Wotawa and his doctoral student Martin Zimmermann proposed a control method that can adaptively compensate for internal errors in software systems. The method selects alternative actions to reach a predetermined target state while providing a certain degree of redundancy. The action selection is based on a weight model that is adjusted over time and measures the success rate of a specific action that has been executed. In addition, the researchers proposed a Java operation and verified it through two case studies.