How AI helps autonomous driving achieve a breakthrough

A man on a bike wearing an Easter bunny costume emerges suddenly and rides across the road. For the car driver behind there is a moment of shock, a glance in the rearview mirror, the slam on the brakes. The driver jerks the steering wheel away from the cyclist. An impulsive process. Whether it’s the Easter Bunny or a man in fancy dress is irrelevant in this situation, the driver sees an obstacle and reacts accordingly.

It's a completely different story when artificial intelligence is behind the wheel. It has not "learned" the scenario "Easter bunny on bicycle" and thus cannot clearly identify the object in front of it. How it will react is unclear. In the worst case, the AI is "confused" and makes the wrong decision.

A famous driving test run by US scientists showed what can happen when AI is confused. After a sticker was placed on a stop sign, the AI interpreted the sign not as a command to stop - but as a speed limit. The system chose an option it was familiar with instead of issuing a warning. Incorrectly calculated decisions like this can have fatal consequences.

Confusing reality

"A perception AI that has never seen a skater has no chance of identifying him correctly," is how Sven Fülster, one of the four founders of the Berlin-based start-up Deep Safety, sums it up. This is a challenge the entire industry is facing. Founded in 2020, the company has set itself the goal of solving the biggest problem of autonomous driving: making artificial intelligence fit for the confusion of real life.

In some situations, AI doesn't stand a chance.Sven Fülster

Fortunately, cycling Easter Bunnies are rare. Because in principle, AI can lead to greater safety on the road. Human error is blamed for more than 90 percent of all traffic accidents. AI systems can read, calculate and interpret an almost unimaginable amount of data simultaneously. They don't get distracted by smartphones, radios or their passengers. They don't tire as long as the power supply and technology are working properly. And the more new data fodder they process, the more precisely they work.

But real life presents an immeasurable combination of possibilities so not every eventuality can be trained and tested. The most dangerous scenario is that the technical systems misinterpret unforeseen traffic situations: in traffic, at a stop sign. Or the Easter bunny.

A driving school for artificial intelligence

So, will it ever be possible to drive safely through London, Cologne, Paris or Berlin in the confusing hustle and bustle of a rush hour while relaxing at the wheel and reading the newspaper? "Sure," say the entrepreneurs at Deep Safety, who are sending their AI to driving school. "We're developing an AI that can say it doesn't know something." 

Sven Fülster, CEO of the start-up, explains how: "With our technology, a driverless car can understand the world on a much deeper level. We've implemented what humans learn in driving school: to think ahead while understanding and anticipating the movements of others."

We're developing AI that can say it doesn't know something.Sven Fülster

Deep Safety’s product is called BetterAI. “We know that AI, unlike humans, will interpret unknown situations in unpredictable ways. BetterAI is the first AI certifiable to the ISO26262 security standard that recognises unknown situations, unknown things, and people doing something unknown," the entrepreneurs explain.

For example, Deep Safety's Perception AI can reliably handle unknown situations and borderline cases on the road. It can also recognise the Easter Bunny on a bicycle - perhaps not as a man in disguise, but still as an unidentifiable object from which distance should be kept. AIs in current vehicle models cannot do this.

Data analysis in real time

Sebastian Hempel, Chief Technology Officer at Deep Safety, explains why this seemed impossible for a long time: "The challenge is that the analysis of the perceptual data – what the camera “sees” – has to run in real time. It takes a very long time to calculate an image. And 30 images have to be calculated per second." Deep Safety's AI has solved this problem. The app even works on a smartphone.

The analysis of the perceptual data has to run in real time.Sebastian Hempel

Competition for the best technology for autonomous driving is in full swing around the world. For years, even the major automotive and tech companies have failed to find a comprehensive solution to these problems. Self-driving vehicles have certainly been used sporadically. But after incidents, they were suspended again.

The founders of Deep Safety are convinced that with their technology, they can avoid such misinterpretations by AI systems in the future. And they have a big vision: "Our short-term goal is to improve the driver assistance systems currently on the road," says Fülster. "In the medium term, our BetterAI will make the driver superfluous. In the third step, we want to actually bring autonomous driving to the city."

We want to bring autonomous driving to the city.Sven Fülster

For the dream of reading a newspaper behind the wheel to become reality, a lot still needs to happen. Deep Safety is part of it. "This is a breakthrough," Fülster says. "We're helping to get people to accept autonomous driving and embrace the technology."