(ORDO NEWS) — Meta was able to create a model that accurately navigates in space without the use of GPS, maps, other positioning systems, and additional reinforcement learning. The new technology will be useful in the design of robots, metaverses and augmented reality.
In its research, the company focused on a “point-to-target navigation model” – a system that can navigate new environments without any map or GPS sensor. This technology uses algorithms that mimic human thought processes in a simplified format.
At the scale of a pair of physical objects, the difference between such an algorithm and one using GPS is insignificant. However, when it comes to digitizing buildings or even cities, the new Meta algorithm can easily outperform the competition due to the smaller amount of data it processes.
AR glasses that show us where we left our keys, for example, require fundamentally new technologies that help AI understand the location and dimensions of an unfamiliar, ever-changing environment without computationally intensive and preloaded maps.
As humans, we don’t need to know the exact location or length of our coffee table in order to be able to walk around it without bumping into its corners. the engineers explained.
The same principle is used in the new model. This system uses a technique known as visual adometry. It allows AI to determine its location based on visual data – for example, if a robot with the support of this system approaches a wall and then turns 90 degrees, it will remember that it no longer makes sense to move backward and take this into account when laying a new route.
The company claims that this technique can be used to create efficient neurons without human annotations.
To further improve AI performance, Meta created the Habitat-Web training data collection. This library provides more than 100 thousand different methods of object-targeted navigation.
The technology links the Habitat simulator, running through a regular web browser, to the Mechanical Turk service and allows remote control of virtual works at any scale.
According to the developer, AI agents trained on this data can achieve the most “state of the art” – for example, they learn to look for objects by looking into rooms and checking hidden corners.
New modular approach
In addition, the Meta team has created a plug and play modular approach. This solution helps papers generalize diverse sets of semantic navigational tasks through a zero-experience learning system.
The idea is to help AI agents adapt on the fly without resource-intensive maps and training. The model “captures” basic visual navigation data once and then applies it to different tasks in the 3D environment without retraining.
In the results of the study, engineers said that new models require 12.5 times less training data and show successful results 14% more often than competitors. In the future, developers want to implement such models in the company’s metaverse, as well as in future AR / VR headsets.
Contact us: [email protected]