Google Is Teaching A.I. How to Travel Through Cities Like a Human

No maps included.

Unsplash / Dennis Kummer

Navigating familiar areas like the streets of your childhood neighborhood is less about carrying around a map and more about remembering what’s around the corner. This comes almost naturally to humans and animals, but can a computer learn to walk the streets of New York City the same way? Google thinks so.

A team of computer scientists at the tech company’s artificial intelligence wing, DeepMind, have developed an A.I. that can travel city streets by simply exploring them on foot. The group exclusively used images from Google Street View to turn their A.I. into a bonafide city slicker by just setting it loose in picture renditions of cities like Paris, London, and NYC.

“The task of navigation can be solved by answering two questions,” Piotr Mirowski, first author of this research, tells Inverse. “Where are you? And how do you get to where you want to go? This can be a child walking in a neighborhood without a smartphone, a bird learning to fly back to its nest, or a robot. So there’s room to get inspired by real life.”

Five areas of Manhattan used in this study.

Google DeepMind

Mirowski and his colleagues explain how they were the first to teach an A.I. to cruise through cities around the world without using map or GPS data in a paper published by the Cornell University Library. While this human-like computer vision technique still isn’t ready for real-world application, it could see uses in aiding self-driving cars navigate areas without reliable map data.

This research made use of neural networks, or artificial replicas of the human brain. These electronic craniums start off like completely clueless tourists, getting lost in bustling city streets, but soon become experts in urban travel after a few million trials.

“We train the neural network to navigate through Central Park, the West Village, Midtown, and Harlem,” says Mirowski. “It is able to memorize a map of the environment without ever seeing a map of the environment. It does this by exploring the area at random in the beginning, but then it receives a reward after getting to a destination. It’s establishing a connection with that [reward] signal and its perception.”

Our agent navigates in visually diverse environments, without having access to the map of the environment.

Google DeepMind

While Mirowski and his team have successfully trained an A.I. almost exactly like a human, there are still some kinks to be worked out. The system needs to be completely retrained every time it’s dropped into a new city, which is keeping it from seeing real-time implementation.

So once the team is able to figure out how to carry the navigational skills the computer learned in one city to other cities, this could become an extremely simple way to train smart cars, drones, and other devices.

Related Tags