Facebook develops a robot to navigate without maps
The company created by Mark Zuckerberg designed a machine capable of moving around the world without the need to use a guide
Researchers of Artificial Intelligence (AI) of Facebook developed a robot capable of navigating in the real world without using maps.
The company created a reinforcement learning algorithm, called Decentralized Distributed Proximal Policy Optimization (DD-PPO), capable of reaching a specific destination 99.9% of the time without using maps.
Facebook highlighted in a statement that the robot can reach destinations with only 3% deviation from the real road. DD-PPO only needs a standard RGB camera with depth data, as well as a GPS and a compass.
One of the main problems of the maps is that the moment they are created they become obsolete, since most of the real-world environments are evolving. By learning to navigate without maps, trained DD-PPO agents will accelerate the creation of new AI applications for the physical world.
The company recalls that previous systems only reached a success rate of 92%, but stresses that “even failing 1 out of 100 times is not acceptable in the physical world, where a robot could harm itself or others if it makes a mistake.“
The Facebook AI team explained that it has used a training method through which it teaches a virtual agent to handle point-to-target navigation equivalent to 80 years of human experience, which is 2,500 million steps.