Researchers at Facebook, which is no stranger to artificial intelligence, have created Habitat, a platform that allows A.I. agents to rapidly learn about the physical world by living in realistic environments within virtual homes.
Teaching A.I.-powered robots to accomplish tasks within the real world takes a significant amount of time. While it may be possible to have actual robots move about in physical spaces for their training, it can take hundreds of hours, an even years, for the A.I. agents to learn how to move from one place to another, determine objects, and answer questions about their surroundings.
In comparison, the A.I. may be placed within virtual homes that represent real ones, with the speed of training dependent on how fast a computer can run the calculations of the 3D world. This means that the A.I. agents will be able to learn thousands of hours training in as short as a few minutes, instead of robots woefully bumping into walls when trained in the real world. In addition, while other simulation engines run at about 50 to 100 frames per second, Habitat is optimized to run at 10,000 frames per second, enabling the rapid training for A.I. agents.
Habitat itself is not the simulated homes, but rather the platform that can host the virtual environments. It is already compatible with several 3D datasets, such as MatterPort3D, Gibson, and SUNCG, but to make them more realistic, Facebook created a database for the platform named Replica.
Replica is a collection of photorealistic rooms that were created through a combination of photography and depth mapping of real homes. The 3D modules of the different rooms may be pieced together in any kind of configuration, depending on the training desired for A.I. agents.
The worlds within Habitat are only visual though, which means that the A.I. agents will not be able to interact with objects in the virtual homes. They may be able to learn how to move from the bedroom to the kitchen, but they will not be able to pick up a spoon, for example. The lack of the ability to interact with the environment is a glaring limitation for the platform, but the Facebook researchers are looking to add the function soon.
- Evolving, self-replicating robots are here — but don’t worry about an uprising
- Why teaching robots to play hide-and-seek could be the key to next-gen A.I.
- Exoskeletons with autopilot: A peek at the near future of wearable robotics
- Image-recognition A.I. has a big weakness. This could be the solution
- Facebook’s new image-recognition A.I. is trained on 1 billion Instagram photos