Facebook A.I. Learns How to Guide Tourists Through NYC

One of artificial intelligence’s defining skills is that it is unwaveringly patient. But it still has a lot of room for improvement in the way of communication skills. That’s why Facebook’s A.I. research lab decided to present A.I. with what may be its greatest challenge yet: Guiding a clueless tourist through the streets of New York City.

The company’s Talk to Walk project hopes to improve A.I.’s mastery of communication by having it guide a virtual tourist by describing landmarks via text messages. University of Montreal Ph.D. student Harm de Vries who was involved in the study told Inverse that this task could be exactly the training needed to bring about next-generation chatbots or the assistant robots of the future.

“The goal of the research project is to explore a more human-like approach to language learning,” he explained. “Previous work mostly focused on language learning from large [collections of text], like Wikipedia. [This] project aims to explore language learning in a contextual environment, making it possible to link words and concepts to perception and action in that environment.”

The tourist bot and guide bot interacting with each other.

Facebook Research

This is #16 on Inverse’s list of the 20 Ways A.I. Became More Human in 2018.

After compiling multiple 360-degree photos of New York City neighborhoods and sample conversations of humans guiding other humans, the researchers contrived two A.I. systems: one to play the tourist and one to play the seasoned local.

To be the tourist, the A.I. had to learn how to describe the surrounding area so the guide could understand where it was. From there, the guide began issuing directions in hopes of getting the lost computer to a preset destination. Unfortunately, the duo never fully completed its task. But there was a silver lining.

The guide bot needs to try and guide the tourist using only descriptions.

Facebook Research

The guide became very good at understanding exactly where the tourist was based on its descriptions. That’s something that humans are actually worse at (“Which McDonald’s are you in front of?!?”)

Even this partial success was evidence that A.I. can take a page out of humans’ playbook to learn how to walk the walk and talk the talk.

Related Tags