Screenshot
One of the reasons why AI developments such as ChatGPT are flourishing and popular is the simplification and lowering of barriers in the operation of such complex and advanced models. Meta also wants the programming of robots to perform various tasks to be simple – so simple that the robots will learn on their own by imitating humans.
Meta’s AI division announced today (Friday) two very interesting developments that will allow robots to take a leap forward in terms of their ability to perform their daily tasks. The first development presented by Meta is called VC-1 (or Visual Cortex-1), which is an AI model that allows robots with cameras to imitate human vision, to better understand the area in which they are moving, the target area of the mission they were sent to, and the area where they need to complete their mission.
The model is based on another model created by Meta with several universities around the world called Ego4D. The model relies on a unique database of video taken from the first person – in contrast to most of the video taken today which is from the side – in the third person. By documenting human interactions from a specific perspective, Meta’s new development can allow robots or augmented reality glasses to better understand the environment in front of them, act accordingly and provide you with recommendations according to the conditions and actions you take in real time. The name of the new development, VC-1, refers to the visual cortex (or pulmonary cortex) which is the main area of the brain where the visual information coming from our eyes is processed. The development is supposed to improve vision through the eyes of the robots – the cameras installed on them.
The researchers of the AI division of say that their first development (VC-1) does bring a line in terms of the operation of robots with the help of AI, but this operation is limited only to short-term tasks such as lifting an object, navigating in space or holding an object – but not to complex tasks Like navigating to an object, lifting it, navigating to a second point and placing it in the new place. That’s why meta developed ASC (or adaptive skill coordination), a set of new techniques that allows robots to perform several actions one after the other while relying on models like VC-1.
ASC is based on three different components: the first is a library of basic capabilities – such as navigation, lifting and placing objects, the second is a system that determines which capabilities the robot should use at any given moment and the third is a system that determines which capabilities the robot will use if it reaches situations for which it was not trained in advance . Meta decided to use Spot, Boston Dynamics’ robo-dog, and first ran the entire process from A to Z on the computer, which included, among other things, internal 3D scans of about 1,000 houses. The researchers then downloaded the program into the robot and sent it on its way.
According to Meta’s researchers, the robots that ran the new model and system achieved 98% success in performing various tasks around the home. These are very interesting developments that could, on paper, be the basis for life alongside robots to help us with household chores – until Skynet arrives and then the situation will probably be less nice.