Robots that learn by watching videos of human activities and simulated interactions. This is the premise of Goalthe company that owns the social platforms Facebook, Instagram and WhatsApp, which is increasingly dedicated to Artificial Intelligence (AI).
American technology thus announces two important advances in Artificial Intelligence and Robotics. By the way, this is the news that was recently revealed in Facebook AI Research (FAIR).
In short, Meta researchers are advancing the technology of embedded artificial intelligence agents capable of performing sensorimotor tasks in the real world. Take, for example, this new approach that allows robots to perform everyday activities, such as manipulating objects.
Artificial intelligence powered by the new artificial visual cortex
The two announced innovations, Adaptive Skill Coordination (ASC) and Artificial Visual Cortex, will allow robots and other artificial intelligence assistants to operate in the real world, without using real-world data.
An artificial visual cortex (called VC-1): An individual perception model that, for the first time, supports a wide range of sensorimotor abilities, environments, and impersonations.
Thus we have improvements in language creation systems, something similar to or at least inspired by the knowledge acquisition models of the human brain. Next, we can see a clearer demonstration of this concept.
The VC-1 is formatted to learn through videos about people performing daily tasks, based on an Ego4D data set (a mega platform that aggregates over 3,000 hours of recording of daily life), created by Meta AI. .
This, along with other academic partners. In this way, the VC-1 equals or exceeds the best known results in 17 different sensorimotor tasks in virtual environments.
For the Meta group, this is the digital “brain” that will make it possible to bring new advances to the field of robotics, among other applications for our day to day.
AI-enhanced movement coordination
A new approach called adaptive skill coordinationcommonly known as sensorimotor (ASC), it achieves near perfect performance, with a 98% success rate.
Considered to be a highly challenging task, robotic mobile reproduction in physical environments consists of moving towards an object, grabbing it, moving to another location, placing the object, and repeating.
The novelties were announced by the Meta group in Publication himself with several videos illustrating the advances in robotics provided by AI.