The new Habitat 2.0 simulation platform allows researchers to accelerate the training of their robots in a virtual environment
Facebook has announced a new step in what it calls “artificial intelligence in the body,” a technology expected to make robots perform routine tasks in a mundane environment, such as loading the refrigerator with groceries or taking out the trash. The development method promises fast results and will probably catalyze the development of all types of robotic “servants”.
The social media platform introduced Habitat 2.0 – an upgrade of the Habitat simulation platform, which allows researchers to train their robots faster – using a virtual environment, ZDNet reported. This training covers simulations of situations with extremely accurate reproduction of the smallest details that machines would encounter in an ordinary environment such as a kitchen or living room.
Training robots through a simulated, virtual environment brings a number of benefits in terms of cost and time savings compared to training them in a real-world setting. This could mean seeing real robot assistants in action soon who can help with housework.
Such machines will be able, for example, to help the person by taking items on command – put goods in the refrigerator, load the dishwasher, do the laundry. More sophisticated applications can be expected, such as a robot to guide a visually impaired person when walking outdoors.
In order for machines to perform useful real tasks, it is necessary to provide them with “experience” in hundreds of different real environments – all the way to baby toys scattered on the floor and the folded corners of the carpet.
In this respect, the simulation changes the game. Instead of physically bringing the robot into various apartments, houses and offices for months and years, Facebook scientists believe that a much more pragmatic approach is to place the robot in a virtual environment. This speeds up his learning. For this purpose, they use a set of data Replica – a compilation of 18 3D scans of real situations, ranging from office conference rooms to two-story houses.
Replica is described as an ultra-realistic library that incorporates some of the finest details in any real setting, including mirror reflections and carpet textures. Until recently, this had its limitations: Replica was a static data set, ie. while the robot can move through cyberspace, it cannot interact with any of the objects.
Now, however, with the new version of the Habitat 2.0 framework, this challenge has been overcome. The robots can not only rotate around the virtual environment in the new platform, but can also interact with objects they would find in an ordinary kitchen, dining room or other commonly used space.
To make the virtual environment so realistic, the project involved the work of over 900 hours of work by 3D artists who created 111 unique layouts of living space, including 92 objects. Special attention is paid to the material composition, geometry and texture of the objects, as well as whether they have specific mechanisms, such as the mechanisms for opening and closing doors and refrigerators. Even trifles such as kitchen utensils, books and furniture have been recreated.
Facebook isn’t the only company interested in virtual AI in-body simulations. Earlier this year, Seattle-based Alan Institute for Artificial Intelligence released ManipulaTHOR, an advanced, virtually robotic arm capable of manipulating objects in more than 100 simulated environments.
However, a significant improvement in the Facebook project is the speed of learning through the new virtual environment. This allows orders of magnitude faster training of artificial intelligence. The speed significantly reduces the experimentation time, allowing researchers to complete trials that would normally take several months in just two days.
In turn, the shortened experiment time allows researchers to try new ideas much faster and more often. And this leads to the completion of a much larger number of simulations and paving the way for more significant progress in the field.
The approach is also relevant to another area of robot development – training a machine to perform many different tasks. Typically, robots currently perform “individual” skills such as picking fruit, placing or opening a drawer. However, home assistants must “know” how to perform many tasks and arrange them in a chain without accumulating mistakes.
In addition, Facebook hopes to expand its data set and environments to include more types of spaces, typical of different cultures, layouts and types of objects. To this end, the company opened Habitat for 3D assets to third parties and announced a partnership with the spatial data company Matterport.
For a start, this partnership will mean the inclusion in the platform of a set of 1000 digital “twins” of real objects – bedrooms, bathrooms, kitchens and corridors of different styles, sizes and complexity.
