[ad_1]
If artificial intelligence robots help us around the house, they will need a lot of experience navigating human environments. Simulators, virtual worlds that look and behave like in real life, are the best place to learn, and Facebook has created one of the most advanced systems of its kind to date.
Called Habitat, the new Facebook simulator was briefly mentioned a few months ago, but today it received the full treatment needed to accompany a document on the system presented to the CVPR.
Teaching a robot to navigate a realistic world and accomplish simple tasks is a process that takes a considerable amount of time. It is therefore unrealistic to do it in a physical space with a real robot. It can take hundreds of hours, even years of real time, to learn, time and time again, the best way to get from one place to another, or how to grab and pull a drawer.
Instead, the robot's AI can be placed in a virtual environment that approximates the real, and the basics can be broken down as quickly as the computer can perform the calculations that govern this 3D world. This means that you can do hundreds or even thousands of training hours in just a few minutes of intense computing time.
Habitat is not in itself a virtual world, but rather a platform on which such simulated environments can run. It is compatible with several existing systems and environments (SUNCG, MatterPort3D, Gibson and others) and is optimized for optimal efficiency, so that researchers can run it hundreds of times faster than reality.
But Facebook also wanted to advance advanced technologies in virtual worlds and therefore created Replica, a database for Habitat including a number of photorealistic rooms organized in a whole house: a kitchen, a bathroom, doors, a living room with sofas, everything. It was created by Reality's Facebook Labs, and is the result of careful photography and in-depth mapping of real world environments.
The details with which they are recreated are high, but you may notice artifacts, especially along inaccessible ceilings and edges. These areas are probably not a concern because AI vision officers do not rely on the details of distant ceilings and angles for navigation – shapes such as chairs and tables, or how the walls define a corridor are much more important.
The innumerable annotations that the team has made on 3D data are even more important. It's not enough to simply capture the 3D environment: objects and surfaces must be labeled in a consistent and comprehensive way. It's not just a sofa, it's a gray sofa with blue pillows. And according to the logic of the agent, he may know whether or not the couch is "soft", that he is "on a carpet", that he is "near the TV", and so on.
Habitat and replica represented with a single color by semantic tag
But including labels of this type increases the flexibility of the environment and a comprehensive API and task language allow agents to perform complex problems in several steps, such as "go into the kitchen and tell What color is the vase on the table?
After all, if these assistants are supposed to help, for example, a disabled person who can not move easily home, she will need a certain level of knowledge. Habitat and Replica are designed to help create that meaning and give agents the practice they need.
Despite its advances, Habitat is just a small step on the road to truly realistic simulator environments. On the one hand, the agents themselves are not rendered realistically: a robot can be big or small, have wheels or legs, use depth cameras or RGB. Some logic will not change – your size does not influence the distance between the sofa and the kitchen – but some will do; A little robot can either walk under a table or be unable to see what's on it.
Habitat seen through a variety of virtualized vision systems
In addition, although Replica and many other similar 3D worlds are rendered visually realistically, they are almost completely non-functional in terms of physics and interactivity. You can ask an agent to go into the bedroom and locate the second drawer of the closet – but there is no way to open it. In fact, there is no drawer, but just a piece of landscape labeled as such, which can not be moved or touched.
Other simulators focus more on the physical than visual aspect, like THOR, a simulator designed to allow AIs to become familiar with tasks such as opening this drawer, an incredibly difficult task to learn from scratch. I interviewed two developers of THOR on Habitat. They both praised the platform that offered AI a perfectly designed place to learn navigation and observation tasks, but pointed out that lack of interactivity meant that Habitat was limited in its teaching.
However, it is obvious that both are necessary, and it seems that for now, one can not be the other – the simulators can be either physically or visually realistic, not both. There is no doubt that Facebook and other researchers in artificial intelligence are working hard to create one that can.
[ad_2]
Source link