Autonomous Driving Algorithm Distributes Risk Fairly

Technical University of Munich, Germany

With the help of ethically acting software, autonomous vehicles are to bring more safety to traffic in the future. (Image: tum.de via istockphoto.com)

Technical realization is not the only obstacle before autonomously driving vehicles can become ubiquitous. Ethical questions play an important role in the development of the corresponding algorithms: Software must be able to handle unforeseeable situations and make the necessary adjustments to avoid an accident.

Now, researchers at the Technical University of Munich (TUM) have developed the first ethical algorithm to fairly distribute the levels of risk rather than operating on an either/or principle. Around 2,000 critical-situation scenarios were tested across various types of streets and regions such as Europe, the U.S., and China. The research was published in Nature Machine Intelligence.

“Until now, autonomous vehicles were always faced with an either/or choice when encountering an ethical decision,” said Maximilian Geisslinger, Scientist, TUM. “But street traffic can’t necessarily be divided into clear-cut, black-and-white situations; much more, the countless gray shades in between have to be considered as well. Our algorithm weighs various risks and makes an ethical choice from among thousands of possible behaviors — and does so in a matter of only a fraction of a second.”

The basic ethical parameters on which the software’s risk evaluation is oriented were defined by an expert panel as a written recommendation on behalf of the EU Commission in 2020. To translate these rules into mathematical calculations, the research team classified vehicles and persons moving in street traffic based on the risk they present to others as well as the respective willingness to take risks. A truck, for example, can cause serious damage to other traffic participants, while in many scenarios the truck itself will only experience minor damage; the opposite is true in the case of a bicycle.

In the next step, the algorithm was told not to exceed a maximum acceptable risk in the various respective street situations. In addition, the research team added variables to the calculation which account for responsibility on the part of the traffic participants — e.g., the responsibility to obey traffic regulations.

Previous approaches treated critical street situations with only a small number of possible maneuvers; in unclear cases the vehicle simply stopped. The new risk assessment results in more possible degrees of freedom with less risk for all. For example: An autonomous vehicle wants to overtake a bicycle, while a truck is approaching in the oncoming lane. All the existing data on the surroundings and participants is now utilized. Can the bicycle be overtaken without driving in the oncoming traffic lane and at the same time maintaining a safe distance to the bicycle? What is the risk posed to each respective vehicle, and what risk do these vehicles constitute to the autonomous vehicle itself?

In unclear cases, the autonomous vehicle with the new software always waits until the risk to all participants is acceptable. Aggressive maneuvers are avoided, while the autonomous vehicle doesn’t simply freeze up and abruptly slam on the brakes. Yes and No are irrelevant, replaced by an evaluation containing a large number of options.

“Until now, often traditional ethical theories were contemplated to derive morally permissible decisions made by autonomous vehicles,” said Franziska Poszler, scientist, TUM. “This ultimately led to a dead end, since in many traffic situations there was no other alternative than to violate one ethical principle. In contrast, our framework puts the ethics of risk at the center. This allows us to take into account probabilities to make more differentiated assessments.”

The researchers emphasized the fact that even algorithms that are based on risk ethics still cannot guarantee accident-free street traffic. In the future, it will be necessary to consider further differentiations such as cultural differences in ethical decision-making. Also in the future, the software will be tested on the street using the research vehicle EDGAR.

For more information, contact Andreas Huber at This email address is being protected from spambots. You need JavaScript enabled to view it..