Facebook Trains AI Robot To Navigate Without Using A Map

• Facebook trained bots for three days inside AI Habitat
• Mapless route-finding is necessary for next-gen robots like delivery drones

According to a report by MIT Technology Review, a team at Facebook AI has created a reinforcement learning algorithm that will let a robot find its way in an unfamiliar environment without using a map. The report said that by using a depth-sensing camera, GPS, and compass data, the algorithm can get a robot to its goal 99.9 per cent of the time along a route very close to the shortest possible path. It means no wrong turns, no backtracking, and no exploration takes place.

Mapless route-finding

According to the report, mapless route-finding is necessary for next-gen robots like delivery drones. Facebook’s algorithm does not handle outside environments yet. But this creation is a positive step towards it and can most likely be adapted to urban spaces.

The report said that Facebook trained bots for three days inside AI Habitat. It is a photorealistic virtual mock-up of the interior of a building, with rooms and corridors and furniture. In that time duration, the bots took 2.5 billion steps.

Testing its algorithm in real physical spaces

The report added that others took a month or more to train bots in a similar task, but Facebook took much lesser time by culling the slowest bots from the pool. This was done so that the faster bots did not have to wait at the finish line each round. According to the report, Facebook is now testing its algorithm in real physical spaces using a LoCoBot robot.