Making Decisions in the Blink of an eye

Min Reading Time

It’s impossible to program for everything that happens on the road. This is why autonomous vehicles need supercomputers that can learn and make decisions. The secret to artificial intelligence, and what it can already do today.

Stefan Schrahe, December 07, 2017

Stefan Schrahe has been writing about everything four-wheeled for three decades now. In his leisure time, he enjoys traveling by bike - though he also prefers motorized ones.

David is eleven and with the curiosity of a child, tries to find his way in the world. As a humanoid robot in Steven Spielberg’s movie A.I. Artificial Intelligence, he is constantly learning to deal with new situations. In fact, artificial intelligence (AI) works in very much the same way as young children learn. A two-year-old, for example, who shoots down a slide with wooden stairs and a straight slide for the first time will also immediately recognize the spiral-shaped slide with metal stairs at the nearby playground as a slide. The brain connects relevant images of both objects in neural networks – the stairs, the semicircular chute, the location of both slides on a sandy playground, and children who are standing at the top or already sliding down. The skill of neural networks is their ability to generalize and correctly evaluate new combinations.

Training supercomputers

Artificial intelligence also involves learning from images. It’s all about training rather than programming. In the process known as “deep learning,” the computer draws on and analyzes huge amounts of data. This allows the AI to make its own decisions and means that it no longer has to rely on every conceivable possibility being programmed beforehand. And with the infinite number of different situations that can occur on the road, this would be simply impossible. But this ability requires immense computing power.

For AI to be able to recognize each situation, interpret it based on experience, and then take action, neural networks need to be simulated and vast amounts of data have to be processed within the shortest amount of time. Only computers with extremely fast processors have this capability, such as the ZF ProAI . This mobile supercomputer was created in cooperation with Nvidia, the world’s leading supplier of graphics processing units. The US company specializes in creating artificial worlds like the scenarios found in computer games. The ZF ProAI computer platform uses this ability the other way around. Instead of creating an artificial image, the computer translates the real world into a data model in which the algorithm gets increasingly better at finding its way around. Every single learning experience is shared in the cloud and is made available to all other systems the next time the algorithm is updated. As such, every ZF ProAI benefits from swarm intelligence.

By 2030, up to

15 %

of newly registered vehicles in Germany could be autonomous.

©ZF

The car detects distracted drivers and reacts automatically.

Ever more AI in cars – three examples

The goal of external airbags is to protect electric vehicles’ batteries in the event of a collision. Thanks to images from the vehicle’s cameras, ZF ProAI can determine whether the object involved in the collision is a little Thuja tree or a concrete pillar, and react accordingly. This means the airbag will deploy in a collision with a lamppost, but not with a decorative plant.
Using images of the driver’s face and the computer’s knowledge of how people focused on the road situation look, the computer can make an assumption about the driver’s level of concentration with extremely high accuracy. If the driver has become distracted, they receive a warning and the car can even be brought to a stop, if necessary.
Since autonomous vehicles allow more freedom for flexible seating positions, seat belts and airbag systems like those we know today are either ineffective or even dangerous. In the event of a collision, artificial intelligence helps analyze the situation in the blink of an eye, evaluate it, and initiate the most suitable course of action for that specific case.

Concept vehicle conducts driving maneuvers autonomously

Using a simple example, Adam Coates explains the extent to which artificial intelligence has already become a fixed part of our daily lives: “In 2017, around 100 million people around the world got internet access for the first time; almost all of them via a mobile device. In doing so, they had already interacted with AI as part of their first online experience.” Coates is director of the Silicon Valley AI Lab established by the Chinese Internet firm Baidu.

The fact that artificial intelligence is by no means a distant vision when it comes to autonomous driving is demonstrated by the team around Oliver Briemle, project manager for ZF ProAI. The team devised a plan early on to take a production vehicle with artificial intelligence to level 3). Since August 2017, a concept car equipped with the ZF ProAI supercomputer has been conducting initial autonomous driving tests. In this context, the computer processes camera, radar, and lidar information to control the compact vehicle, which remains virtually unchanged on the outside. The vehicle is scheduled to hit public roads in early 2018.

The Supercomputer ZF ProAI jointly developed by Nvidia and ZF

Playing battleship with radar scans

The fact that artificial intelligence processes information visually, but doesn’t necessarily need to rely on photos gives Dr Jochen Abhau, expert for machine learning at ZF, a new perspective: “We can copy radar scans of the immediate vicinity onto a data grid and process them as patterns, for example. The principle is very similar to the concept of the game Battleship played on paper. The system can use the pattern to ascertain whether a playground or parking lot is located behind a row of houses along the side of the road.”

Thanks to this semantic segmentation, the car knows whether it can park nearby, and will navigate its own way there, if necessary. However, the AI experts can only speculate as to whether the algorithm will be as pleased to find the parking space as the young child is when it discovers an empty slide.

When it comes to AI, training – not programming – is the name of the game.

—

In a nutshell: Autonomous driving is only possible with artificial intelligence, since driving on public roads involves numerous unpredictable situations that simply cannot be programmed for. This means computers need to learn to interpret situations and make the right decisions. Together with Nvidia, ZF has developed a supercomputer for autonomous driving: the ZF ProAI. And in early 2018, a car equipped with the system will begin driving on public roads.

Autonomous Drive - in Five Steps to a Self-Driving car

Since the 1950s, not just automotive engineers have been fascinated by the dream of fully autonomous driving. What stage are we at today? When will it become a reality?

Share this page on:

Get in touch

Technology

Making Decisions in the Blink of an eye

Training supercomputers

Training supercomputers

Ever more AI in cars – three examples

Concept vehicle conducts driving maneuvers autonomously

Concept vehicle conducts driving maneuvers autonomously

Playing battleship with radar scans

Playing battleship with radar scans

Related Topics

Autonomous Drive - in Five Steps to a Self-Driving car

Autonomous Drive - in Five Steps to a Self-Driving car