Mikael Johansson develops autonomous vehicles and AI solutions at Scania

After two years at the Volkswagen Group’s central research division in Wolfsburg, Mikael Johansson, predevelopment of autonomous vehicles, knows everything there is to know about how to get an autonomous vehicle to drive the required route. He has been working on the development of electrical systems at Scania for nine years. The opportunity to work in Germany on autonomous vehicles and AI (artificial intelligence) came at just the right time!

“I wanted to try something new, and preferably abroad. Now I get a lot of use out of everything I learned.” The projects under way at group research in Germany are at the very forefront of the VW Group’s research into developing concepts for finished projects, although a lot of work remains to be done before any launch is on the cards.

Mikael Johansson spent his two years working on a joint project involving VW Group Research, Scania and MAN. In the last year up until last summer, he worked solely on AI. This can involve image processing programs that interpret and classify camera images in order to produce interesting information:

“We train neural networks to be able to classify what pixels are road and which are something else so that the vehicle will understand where it is going.” Mikael is now back at home in Södertälje, but is continuing to work on autonomous trucks and buses. He has a particular focus on AI and wants to continue developing what he was working on in Wolfsburg.

“AI can be found in all industries – medicine, advertising, finance. The basic algorithm is the same. It’s just how it develops that’s different.”

Scania is growing quickly within the field and a key driver is learning a lot from one another within the VW Group. There are various ways of developing autonomous vehicles. Either you feed in information such as “if a particular thing happens, the vehicle should respond in this way; if something different happens, the vehicle should do this.”

“This has advantages, as it is predictable, but it also has limits, because it is difficult to capture all the scenarios and it may also be very difficult to formulate the precise behaviour in some complex cases.”

Like a black box

With the AI method, a large volume of data is fed in and the software adapts the algorithm and captures a great deal more information in a more abstract way.

“It’s like a black box. It can handle a lot of information and it is a powerful tool, but it’s not as predictable. So it’s a question of combining different technologies in order to make use of their respective strengths and compensate for one another’s weaknesses. With this variant of AI, which has had a huge impact lately, you say precisely what is right and what is wrong, and the algorithm is trained to imitate this.”

During tests on the test track, things have gone relatively well, according to Mikael.

“No disasters. Instead the vehicle has just not started at all, or has clearly driven badly, such as very jerkily.”

Learns good behaviour

“There are various ways for AI to learn. Using methods that have not had much of an impact so far, the algorithm can be trained by specifying in a more general way what is right and what is wrong, rather than in detail. For an autonomous vehicle, for instance, you can potentially program in what is good – arriving at a requested destination – and what is bad – crashing – driving behaviour. Then the algorithm learns itself the precise behaviour needed to fulfil these general objectives.”

Almost like magic

If the algorithm is then left to test this millions of times, you get useful results. But naturally that has to be done in a simulation, and not out on the roads, and a sufficiently realistic simulation environment is still not available.

“Al technology is not actually new or difficult to use. Its breakthrough has come now because it is only now that there are sufficiently powerful computers to cope with the demanding calculations. It is almost like magic when you see what AI can manage, and in the future development work will be more about using large quantities of data and training AI algorithms to find new functions. It’s a brand new way of working!”