Making robot cars understand humans

Intelligent vehicles cruise the campus and the Delft inner city. Professor Dariu M. Gavrila (1967), who develops self-driving cars, will present his inaugural speech on Friday June 23 2017.

Professor Dariu Gavrila (Photo: JW)
Professor Dariu Gavrila (Photo: JW)

I read you worked with Daimler for almost twenty years, from 1997 to 2016? What did you work on?

"I was there exactly in the period that this whole field of intelligent vehicles and driver-assistance started to happen. It was a wonderful time. At the beginning at Daimler, we were thinking about systems that use cameras in the car. And many people said: this is never going to happen, we will never have camera sensors in cars. Radar sensors are good enough and camera sensors will never be used in cars. That was in 2000."

And now?

"You can buy a Mercedes that has six or eight cameras on board. It is stacked with sensors. There has been an amazing change in perception among people that come from the traditional car industry. They have seen a movement towards software, a less mechanical but more intelligent type of technology. It has been a great time, seeing the field emerge from its humble beginnings in 1997 to road tests by Uber and Google. But the future is going to be even better."

In the meantime, a lot must have happened in computer vision?

"Yes indeed. I can make it a little more precise. When I started at Daimler, I became interested in the plight of the vulnerable road user, meaning pedestrians and cyclists. This was in 1998 and 90% of my colleagues at Daimler were telling me that it was not possible to reliably detect a pedestrian from a camera image. The rest were saying that people would never buy this product because it would not help the guys in the car who don't care about the people outside."

Which is a bit crude of course.

"Yeah, but half told me it was not needed, and the other half that it could not be done. Only in 2007 did the mood start to change a little. And now it seems strange that people thought we didn't need to be aware of pedestrian safety."

After such an adventure, what brought you to Delft?

"I was given the opportunity to create a new group here, the intelligent vehicles group (at the Faculty of Mechanical, Maritime and Materials Engineering, jw ). It was a new challenge. At Daimler I had pretty much achieved everything I could as a technical person. One year before I came here, I became a distinguished scientist, which is the highest honour if I may say so myself. Between 2013 and 2015, my algorithms for pedestrian safety were reintroduced in the S, E and C Class Mercedes Benz. So for me, setting up a new group meant having a broader perspective. We're not just working on pedestrian safety, we're covering the spectrum of the self-driving car. Plus it's kind of a nice idea to give something back to the new generation. I have learned a lot and I actually enjoy giving classes. There is a gap in technical education in this field. There was no course on intelligent vehicles in Delft, at least not from the perspective of inside vehicles themselves. There are colleagues at the Faculty of Civil Engineering and Geosciences who work on intelligent transportation systems. I thought they were a good match. Our group is more focused on on-board technology, while our colleagues at Civil Engineering – Professors Bart van Arum and Serge Hoogendoorn for example – work more at the macro-level, the transportation infrastructure, thinking about traffic flow and cooperation. This is also complementary to our work. If we unite our strengths, we can build critical mass. These were the main reasons for me to come here."

'Students will always take the right of way anyway'

Inside the group's demonstrator vehicle (Photo: TU Delft)
Inside the group's demonstrator vehicle (Photo: TU Delft)

You have been quoted as saying that the technology for self-driving cars is 95% ready. So what is the remaining 5% and why is that so hard?

"Well, first of all, this 95% is not a hard figure. What I mean is that if I look at the different types of driving environment, there are areas where the technology is not yet ready to drive in a socially acceptable way."

To make it concrete: if you have a complex traffic situation, would a self-driving car at present be able to cope and stop?

Well, let me show you something from my inaugural speech. Most of the auto industries are working on the highway scenario, which is kind of easy and predictable. The urban environment is harder because you have many more things to worry about like crossing traffic, vulnerable road users, unclear markings and traffic lights. So this environment is already much more complex. And then there are different urban scenarios. There is a big difference between the broad lanes in Phoenix, Arizona, where Google works, and the inner city of Delft, where we did a test drive. Here you are in close contact with cyclists and pedestrians all the time, there are bridges, you have to negotiate access to make out who goes first. There is a zebra crossing and just ahead of it a guy crossing the road without even bothering to look. Then we get into a dark street with a car parked the wrong way, can we pass it? There are cyclists here that are on collision course. And this is just half-an-hour's worth of Delft city traffic."

Is automated driving in such an environment achievable?

"Yes, of course it is. It is just much harder. Going from Phoenix to Delft and then possibly on to Delhi is what the last percentage of automated driving is all about."

I understand that automated vehicles from the Researchlab Automated Driving Delft (RADD) will appear on campus shortly.

"This is an initiative of our colleagues from Civil Engineering. They have the WePod driving around and there will be a demo on Tuesday June 27. As I understand it, it is the purpose of this lab to see these technologies in action. They will have to make sure that the technology is safe, that we're not going to use students as guinea pigs. I don't expect we're going to do tests without a safety driver any time soon."

What do you recommend students to do if they encounter a self-driving car?

"I don't think it will matter a lot whether it is autonomous or not – students will always take the right of way anyway, ha ha! I think these vehicles will drive extra defensively at first. That means that the danger is rather that the car won't move forward anymore. You would need the yellow traffic wardens to help it out. If you want a robotic system to integrate in a city or a populated area like the campus, you cannot expect humans to adapt to the robotic system. The regular users should not have to do anything special. So students should behave in their 'normal' way, otherwise it would just not be feasible."

Related content:
Gavrila's homepage:
Intelligent Vehicles Group

The inaugural speech is entitled 'The Intelligent Vehicle (R)evolution', 15:00, Aula TU Delft, June 23 2017.