Animation of a robot standing confused at a road intersection. Around him is green grass. There is a tree on the left. To the right of the robot is a signpost with three illegible signs.
Illustration generated by AI. (Image: New Media Center using Adobe Firefly)

Artificial intelligence is the theme at TU Delft this year. Delta explains six pieces of AI research. Part 1: how does AI ensure that a robot without GPS knows where it is?

Lees in het Nederlands

Beep beep. ‘GPS reception not available’. So there you are, in the middle of a big city where you have never been before. It is right in the middle of a lot of tall buildings that the GPS signal may drop out. These are known as ‘urban canyons’. Google Maps has recently started asking users in these situations to help. ‘Point the camera at buildings or traffic signs in the area’. This is the first step towards GPS localisation, in other words, using sensory observations to determine where you are. Just like humans in fact.

It is not only Google that is working on this, Mubariz Zaffar, a doctoral candidate (Faculty Mechanical, Maritime and Materials Engineering, 3mE), works every day on artificial intelligence that could make GPS redundant. This will help AI agents – such as wildfire robots, self-driving cars, or the navigation app on smartphones –themselves determine where they are and how to get from A to B.

To do this, Zaffar and his colleagues feed the AI in the AI-lab 3D Urban Understanding (3DUU) huge amounts of images. And that is not all. In contrast to programmes like Google Street View, that is basically compiled from a large collection of photos which users can flick through, the activities at 3DUU are much more what Zaffar calls representation. The AI agent needs not only to connect images to real places, it needs above all to work out where it is.

King’s Day
This is quite a challenge, says Zaffar. After all, how do you teach AI that the leafy tree on the corner in August is the same as the bare tree in January? It looks really different. And without having images of a street in the middle of the night, the robot still needs to be able to find its way using information about the street in the light of day and about the differences between night and day. And what about if it is there on King’s Day? A village looks very different decorated in orange flags than on the images that the AI has previously seen.

To make it even more complicated, the AI needs to place the information in a wider context, explains Zaffar. “Not only does it need to know what location A looks like, what location B looks like and where these locations are on a map, it also needs to know their location in relation to each other. Without this information, you won’t get there.” Compare this to a human who may know the city centres of The Hague and Amsterdam like the back of his hand, but does not know that The Hague is 60 kilometres south west of Amsterdam, nor where south west is. Getting from one city to the other is then extremely difficult.

“Our goal is an AI agent that uses nothing more than just the camera. It would be like people who use no more than their eyes,” says Zaffar. “This is more reliable than GPS and can be used in countless situations.” Just think about buildings that GPS is unable to penetrate through, or urban canyons like in the first paragraph. Or places that are dangerous for people such as areas with wildfires or an exploded nuclear power station. Self-driving cars would navigate through cities more reliably if they were less dependent on GPS.

Dark sides
All technologies have pros and cons. Zaffar too is aware of the dark sides of the AI that he is working on. “All navigation robots would benefit from better localisation, including those that are deployed as weapons in armies. More to the point, GPS as is now on every telephone, was originally designed for military use.”

But you cannot stop progress, says Zaffar. “People are curious creatures. This is what has brought us to where we are today. Instead of stopping the development of AI, we had better accept it and ask ourselves what the best way to use it is and what we need to do to make sure that it is not used for bad purposes. Then AI really could solve problems.”

  • If you want to know more about the 3D Urban Understanding AI-lab, check the website for more information.
  • Part 2 in this series about artificial intelligence will be published next week: How is AI helping to make wind turbines lighter?