Facebook has Trained its AI so that They Can Navigate Without a Map

Facebook AI

We all know about the amazing robots of Boston Dynamics and their diverse stunts. But, in reality, they are lost without a map. They do not know which direction to go from one point to another, they are always consulting an internal map that tells them where to move.

Facebook claims to have a solution to this problem in AI. That solution is the artificial intelligence that allows robots to move without the help of a map and intuiting the path.

A team of Facebook engineers has created a new artificial intelligence algorithm with reinforcement learning. This AI, applied to a robot, allows you to find a way to reach a certain place by analyzing your environment even if it is an unknown environment. You only need a camera with a depth sensor, a GPS and a compass.

How do you teach an AI how to learn to locate? Through constant practices, thousands and millions of practices. Facebook created AI Habitat, which is a kind of photorealistic virtual model of the interior of a building. AI Habitat has rooms, offices, corridors, furniture, doors … all as if it were a real building.

With the created environment they placed inside a few bots that move inside with the aim of reaching points of the building without knowing the building beforehand.

Facebook has Trained its AI

For three days the bots walked (virtually) 2.5 billion steps, the equivalent of the human walking for 80 years, according to Facebook. They also comment that in order to speed up the process, they started to take the slower bots out of the building, so the fastest ones didn’t have to “wait” to find the new point while the slow ones ended.

As a result, a robot is now able to reach the target point of almost 100% of the time, specifically 99.9% according to the researchers. The bots do not go in the wrong directions, they just have to go back on their own steps and generally do not engage in exploring before going to the requested point.

The robot is given a point in the GPS and this seeing how the rooms and corridors are correctly determined whether or not to cross there without having to know all the rooms and corridors in detail.

Facebook is now testing artificial intelligence in real environments, with real robots like the LoCoBot. The next step will be to teach the AI ​​to move in outdoor environments, although here there are many more variables to consider other than doors, corridors or rooms that can be found inside a building.

How exactly artificial intelligence manages to learn to move is a complicated thing to know. It is generally not known exactly what procedures an AI has followed to learn a new concept.

Facebook indicates that they are investigating how to make AI can also be located without GPS or depth camera, because in real environments, for example, the GPS may not work properly indoors.

Advertisement

LEAVE A REPLY

Please enter your comment!
Please enter your name here