Facebook Updates Habitat Environment to Train ’Embodied AI’

0
36

In 2019, Facebook released AI Habitat, a simulator tool that allows AI systems to practice to operate in surroundings that are designed to resemble real-world situations, such as residences and offices. Facebook stated on 30th June 2021, that it has enhanced Habitat’s capabilities to make it “orders of magnitude” quicker than existing 3D simulators on the market, allowing researchers to simulate more complicated operations like setting the table and filling the refrigerator. At the same time, Facebook teamed up with Matterport, a 3D space capture startup, to open-source what it says is the world’s largest library of indoor 3D scans.

The latest version, Habitat 2.0, improves on the original in several areas. It presents ReplicaCAD, a 3D collection of residences created by artists that match real rooms with furniture that can open and shut, such as cabinets and drawers. It also comes with the Home Assistant Benchmark, a collection of tasks for robots that include cleaning the house, storing groceries, and setting the table, among other things. For instance, a fully embodied robot might check to see if a door is locked or collect a smartphone that is ringing upstairs.

ReplicaCAD, according to Facebook, comprises 11 layouts over 92 distinct things, including furniture, cooking utensils, novels, and more. It took professional 3D designers employed by Facebook over 900 hours to build. In the case of the Home Assistant Benchmark, robots must choose and put things from receptacles like countertops, sinks, and sofas, as well as open and close containers like drawers and refrigerators as needed.

In addition to the new dataset and benchmark, Facebook claims that Habitat 2.0 is a lot faster than the original, with rates surpassing 26,000 simulation steps per second 850 times real-time (30 steps per second) on an 8-GPU machine. Rendering a single picture and modeling rigid-body dynamics for 1/30th of a second is referred to as a step. According to Facebook, this is 100 times faster than previous efforts, reducing testing time from 6 months to under 2 days. Existing simulators, for example, generally reach 10 to 400 steps per second. Furthermore, ReplicaCAD compromises the performance benefits “directly transfer” to faster training times and improved accuracy from training models with more experience notably for object rearrangement tasks.

Along with Habitat 2.0, Facebook is releasing the Habitat-Matterport 3D Research Dataset, a collection of 3D indoor scans developed in collaboration with Matterport (HM3D). It’s a collection of 1,000 Habitat-compatible scans that include “accurately sized” residential areas including apartments, multifamily dwellings, and single-family houses, as well as commercial spaces like offices and retail stores.

Facebook says it plans to extend the collection in the future to include scans from other nations, as well as annotations that might aid AI in tasks such as item retrieval and the firm wants to explore dynamic settings.

LEAVE A REPLY

Please enter your comment!
Please enter your name here