DEAR READER,
HERE YOU CAN...

… BROWSE
THE ABOUT TRUST
ARCHIVE

OUR ARCHIVE

… DOWNLOAD
THE CURRENT ISSUE
AS A PDF

DOWNLOAD

… ORDER
THE ABOUT TRUST MAGAZINE
FOR FREE

CLICK HERE

autonomous driving

A Matter of Trust

Self-driving cars and buses may soon take over the wheel from us. Yet many people have concerns about relying completely on technology while out in traffic. Whether we trust technology or not determines its fate.

Text Thomas Schmelzer  Illustration Manuel Bortoletti

When Christian Müller wants to explain how artificial intelligence (AI) in autonomous vehicles can make life-or-death decisions, he plays a video. On his computer monitor, a white car appears. Traveling at about 50 kilometers an hour, it’s approaching an intersection and is making a left turn when a man suddenly steps into the crosswalk directly in the vehicle’s path. Instead of braking or swerving, the car remains on course. There’s a crash. Müller keeps a straight face. He already knows the tragic results. “Well, the artificial intelligence wasn’t so smart there,” he says.

The video scene is a simulation, the pedes­trian, an avatar modeled on real movement data, and the auto, a projection from the underlying artificial intelligence. Only Müller, the scientist, is flesh and blood. He works at the German Research Center for Artificial Intelligence (DFKI) in Saarbrücken, Germany—nestled in the rolling hills between Luxembourg and Metz, France, on the Saar River—and is doing everything he can to ensure that such accidents happen only in simulations in the future. Müller is investigating how autonomous vehicles behave out on the roads in traffic with AI at the wheel. He is also working with TÜV SÜD and other partners to develop testing methodologies for these types of AI.

On his computer, Müller selects a different AI variant for the autonomous car and starts the scene again. The car again approaches the intersection and again begins to make a left turn. But this time the scene doesn’t end in tragedy. As the pedestrian enters the crosswalk into the vehicle’s path, the car easily swerves out of the way. “I’d be more likely to trust this sort of AI,” Müller says.

The simulation on Müller’s computer monitor is just one of countless situations that self-driving vehicles will have to cope with out in real-world traffic in the future. Yet it demonstrates what will be important no matter the circumstances: Will we trust the machine, its sensors and its enormously complex intelligence—or won’t we?

It’s the question that will determine whether autonomous vehicles will achieve a breakthrough in the coming years. Whether at some point in time there will be cars with sleeping passengers behind the wheel on the autobahns. Whether cars will be looking for their own parking spaces in the future. And whether driverless taxi fleets will be chauffeuring us through city centers any time soon. It is a question whose answer will also determine what the future of our mobility will look like.

        THE 5 STEPS of autonomous driving 

 

 

Figures vs. Feelings

For answers to this question, statistics are a good place to start. Currently, about 80 to 94 percent of all traffic accidents around the world are caused by human error. Autonomous vehicles could reduce this high rate to a minimum, so the expectation–the consulting firm McKinsey, for example, predicts for the USA that the market penetration of highly automated vehicles could enormously reduce fatal accidents by the middle of the century.

On the other hand, many people are deeply skeptical about driverless vehicles. According to a study conducted by the Baden-Wuerttemberg Cooperative State University, two of three respondents believe that a lack of trust in autonomous vehicle technology could complicate its acceptance. Almost as many study participants also expressed fear that hackers could target the extensive software in the vehicles. The German online portal for statistics, Statista, found that the idea of fully automated driving systems generated negative associations in almost half of those surveyed.

On top of all this is people’s overestimation of their abilities. “Almost everyone believes they drive much better than they actually can,” says Andreas Herrmann from the University of St. Gallen. For his book Autonomous Driving: How the Driverless Revolution will Change the World, he set man against machine in a simulator. The results: in the majority of situations, the machine was better. “However, we also saw that the machine reaches its limits in more complicated traffic situations,” Herrmann adds.

These critical situations are the crux of the matter for the future. “If we are to hand over control when driving, we must have enormous trust in the system’s safety and security,” says Global Head Autonomous Driving Dr. Houssem Abdellatif at TÜV SÜD. “That’s why we must work with manufacturers to prove that the technology can guarantee this safety and security.”

This is exactly what scientists, entrepreneurs and inventors have been working on for years. TÜV SÜD is also working on creating solutions in a variety of projects around the world. Together they’ve managed to make the technology safer, more transparent and more verifiable. And, together, they are tackling the next major challenges.

On the task are people like Marius Zöllner, who is working with TÜV SÜD testing self-driving vehicles in real traffic at the Research Center for Informatics (FZI) in Karlsruhe, Germany. Inspectors like Peter Salzberger, who is working for TÜV SÜD in the PEGASUS project to ensure that autonomous vehicles don’t make any driving errors on highways. And Christian Müller, who is unlocking the secrets of artificial intelligence in Saarbrücken.

A Matter of Trust

A Matter of Trust

A self-driving car turns a corner just as a pedestrian steps into the street—this scenario could soon be an everyday occurrence.

Demystifying AI

Müller’s computer screen is now displaying tables, graphics and schematics. As he clicks from slide to slide, words and abbreviations flash across the monitor, including Grey Box Testing, Deep Reinforcement Learning and APPL. Müller is working with his team, plus the colleagues from cooperation partner TÜV SÜD in Munich. “Basically we’re all working on demystifying artificial intelligence,” he says.

In the project openGENESIS, Müller is working towards this goal with TÜV SÜD and additional project partners to enable vehicles to make independent decisions. It was this project that made the rapid rise of autonomous driving in recent years possible in the first place.

Opening the Black Box

The breakthrough for driverless driving is mainly due to the rapid increase in computing power on microprocessors, as well as new and improved sensor technology. Only this combination has allowed artificial intelligence to finally de­velop its full potential. It is now unbeatable in finding patterns in vast amounts of data and drawing its own conclusions. There are many areas in which it has long since surpassed human brainpower.

Yet this technology still has one big catch: hardly anyone completely understands exactly how it works, what precisely is going on inside AI’s black box. Why an AI can recognize millions of pictures of dogs and then suddenly identifies a hamster as one; why it can drive for thousands of kilometers along country roads and through intersections without a mistake and then sud­denly stops in the middle of an intersection. It’s still something of a mystery.

To solve this mystery, Müller, his team and the project partners are, to put it in simple terms, proceeding as follows: first, he combs the huge datasets to find situations in which AI has previously been pushed to its absolute limits. In a second step, he recreates the situations virtually. To do this, he uses datasets that his team has gleaned from actual movement profiles from road users. Next, various AI systems compete against one another in the virtual situations. This is where Müller looks into the black box. Finally, he checks to see if the virtual situation can be transferred back into the real world.

For openGENESIS Project Manager Matthis Eicher at TÜV SÜD in Munich, each of these results is a new building block on the road to consistent and reliable testing of artificial intelligence in self-driving vehicles. “We want to build a community with openGENESIS in which each partner utilizes their various strengths to investigate and shine a light on the subsections of AI,” he explains. Gradually, the big overall picture will grow. “Our ultimate goal is to completely understand the technology and to develop the requirements for testing and certifying AI in vehicles.”

Strong Together

Strong Together

Autonomous vehicles are equipped with dozens of sensors and cameras that mutually complement one another. Together they cover all of the areas around the vehicle.

100 million – The number of traffic situations that must be played through and simulated for each and every automated driving function.

From the Computer to the Streets

To ensure approval, autonomous vehicles must demonstrate that they can drive safely under real-world conditions, in everyday traffic, not just in a simulation. This is the specialty area of Professor J. Marius Zöllner, who holds a doctorate of engineering. At the FZI in Karlsruhe, where he is also a member of the board of directors, the computer scientist heads to the tools he uses for these tasks. There are two test vehicles parked in a garage, the trunks jam-packed with processors, cables, circuitry. The car bodies are equipped with more than two dozen cameras and sensors. Where the air conditioning is normally located on the dashboard are three selector buttons: normal, assisted, autonomous.

In field tests, the cars drive themselves through intersections, stop at signals and navigate through the city’s streets. “In the middle of Karlsruhe,” Zöllner clarifies. Researchers in the cars only intervene in problematic situations during such test drives. Zöllner oversees Germany’s first testing grounds where self-driving vehicles are already being testing in urban street traffic. In order to ensure safety and collect as much data as possible, the researchers have also installed sensors and cameras at intersections, in the asphalt and on bridges.

On a touchscreen as big as a kitchen table, Zöllner shows what can be done with this data: a virtual, four-lane intersection, traffic lights, stop lines, a tramway bridge, a whole lot of colorful cars. “This isn’t a simulation, but is actually a real intersection in Karlsruhe,” Zöllner explains. The cars are moving as they do in real life. The bicyclists and pedestrians behave as they do in real life.

 

Invisible Obstacles

Observing this intersection, Zöllner and his team have learned that about one driver an hour makes an illegal U-turn. That bicyclists sometimes cross the intersection against the red light. And that pedestrians in a hurry sometimes scurry through the crosswalk at the last second. Fortunately, they can use this data to test real everyday situations. For instance, how their test vehicles would behave on a spring morning with some light fog when they are entering an intersection at exactly 34 kilometers per hour behind another vehicle that suddenly slams on the brakes. “It allows us to collect extremely valuable data,” Zöllner says. So far, however, there is no uniform worldwide system to catalog and store this data.

Another challenge is the quirks of some of the sensors. Every technology has strengths and weaknesses. Cameras have trouble when there is backlighting but can recognize colors. Radar waves fail in tunnels but are good at measuring distances. Laser sensors are very precise but cannot identify color. Together, the sensors and cameras cover all possible situations. Yet sometimes glitches crop up. On the test course in Karlsruhe, an invisible obstacle suddenly appeared in an intersection during winter. The test vehicle abruptly stopped every single time. “The technology saw a monster where there was none,” Zöllner explains. At some point his team realized that the sensors were incorrectly interpreting reflections from bushes along the roadside.

Despite all the challenges, Zöllner thinks that vehicle automation will continue to accelerate. Scenarios with multistory parking structures and company premises where cars look for empty spots and park themselves—or even semi-autonomous driving on the autobahn—is something he considers realistic in the next ten to fifteen years. But he also believes that it will still take some time before vehicles will be able to manage every situation without a human driver. There are still too many uncertainties, risks that are too great, costs that are too high.

Resource: Trust

For many people, there is also still too much skepticism. Armin Grunwald, Professor of Philosophy of Technology at Karlsruhe Institute of Technology, has been studying the societal consequences of technological advancement for years. He’s learned that the majority of people generally view technical progress in a positive light. At the same time, he also knows that for every technological innovation, there’s one resource that is more important than all the others: “Nothing works without trust,” he says.

For autonomous driving, that’s why it’s a matter of appropriate diligence and a measured marketing launch. It’s about shifting down a gear and not just banking on speed. This is the only way to learn from the mistakes that will inevitably happen at some point and to improve the technology accordingly. “I recommend that they take time and that not everyone bends to economic pressures,” Grunwald says. This will also allow customers to get a better feel for the new technology.

Author Andreas Herrmann has also observed in his research that it may indeed need some time. When he sent his subjects on a test drive in a driving simulator, at a certain point during the test he gave them the instruction to turn the driver’s seat to face backwards and let the computer keep driving. Herr­mann was surprised at how great the inhibitions were to handing over control, even in a simulator. “An independent testing and certification process for autonomous vehicle technology is undoubtedly an important step in building trust,” he says.

This next step for highly automated driving on the autobahn is currently being developed in PEGASUS. In this joint project, sponsored by the German Federal Ministry for Economic Affairs and Energy, car manufacturers, suppliers, research institutes and TÜV SÜD—as the sole testing organization—are drafting standards and regulations for the uniform testing and approval of highly automated driving functions.

“First it’s all about developing the guidelines for uniform testing,” says Peter Salzberger, who is heading the project for TÜV SÜD. It may sound trivial, but in practice it’s an extremely complex undertaking. Which driving scenarios must be tested for approval? Which of these should be real-world tests, and which simulated? How often must the highly automated vehicles run through a particular situation? How specific do the requirements have to be? Where do the risks of automation lie? And how high does the level of human and technological performance need to be for highly automated driving?

““We had to develop standards for the virtual simulations, standards for the software and testing systems that are used, standards for how we plan to design the scenarios for real-world testing sites,” Salzberger notes. Shortly before the end of the project, he’s satisfied with its results. “We’re pretty far along for highly automated driving on the autobahn,” he says. TÜV SÜD provides its own testing environment for safeguarding the highly automated driving functions on its test site, and its industrial partners’ test vehicles are already driving autonomously on the A9 Autobahn between Munich and Ingolstadt.

500 million – The estimated lines of code necessary for programming an autonomous vehicle. To compare: a Boeing 787 Dreamliner manages with about 6.5 million lines of code.

The Future Is Already Here

If you want a glimpse even further into the future, you’ll have to travel to Singapore and go to the one-north business park. A pilot project has autonomous taxis gliding along the asphalt between palm trees and high rises. The taxis currently have minders behind the wheel during this testing phase, just in case, but should soon be driving through the city fully on their own. Nanyang Technological University in the western part of the city has a built a small mini-city on a two-hectare site to test autonomous vehicles as realistically as possible. TÜV SÜD is assisting here as a strategic partner in the CETRAN project, which is working on making highly automated buses and cars as safe as possible. The city plans to have a fleet of such buses on the streets by 2022.

So Singapore is potentially becoming a place where highly automated driving may have its first real breakthrough on a larger scale. Yet here, too, consumers will ultimately have the final say about the technology’s fate. “We want to ensure throughout the world that a vehicle has been tested to the highest possible standards of technology and rigor,” Global Head Autonomous Driving Abdellatif says. “This vision will only become a reality when people trust the technology.”