Facebook’s Habitat 2.0 in action.
Facebook
You know Facebook best for its social networking services, but tucked away behind the scenes are many projects the company is working on to advance technology in myriad ways. One of those projects involves training robots to navigate home environments to better serve us, their future owners, and perform the chores we find most loathsome.
Back in 2019, Facebook released AI Habitat, an open-source simulation platform of photo-realistic 3D home environments. Researchers could use the platform to train robots to navigate these environments. But now, the company is taking it to the next level.
On Wednesday, the company announced Habitat 2.0, the next version of that platform, which features vastly improved speeds, new benchmarks and a reconstructed dataset that will allow robots to be trained in interactive 3D environments. Not only will researchers be able to train robots to navigate these environments much faster and more efficiently than before, but they will also now be able to complete tasks within that environment.
We’re talking stocking the fridge, setting the table for dinner, loading the dishwasher and even taking out the garbage. If you’re not already sold on the idea of a robot butler, knowing that you could cross these tasks off your to-do list could well be the incentive you need.
The key to this second version of Facebook’s AI platform is the reconstructed dataset, which it calls ReplicaCAD. Habitat 2.0 uses a mirror image dataset to its predecessor, but the previously static images have now been converted into individual 3D models, which means robots can be trained to move around and manipulate them in a whole new way.
To ensure that robots are being taught effectively, the dataset includes information about the material composition, geometry and texture of all the individual objects. This allows them to treat a fridge door differently to a book to a sink to a sofa cushion, for example. In total the dataset includes 92 objects that took 3D artists 900 hours to create and that can be arranged in 111 unique setups of a single living space.
The speed of the new platform also shows a marked improvement over the past version. Facebook now claims Habitat 2.0 over 100 times faster than its predecessor, dropping experimentation time down from 6 months to just two days.
The company is hoping the platform will provide a research framework for training embodied AI (that is, AI embedded within a physical system, such as a robot) for years to come. “We hope that the ability to perform more complex tasks in simulation will bring us closer to the AI that can help make our everyday lives easier and better,” said Facebook research scientist, Dhruv Batra.