In a few decades, we may be debating human-robot equality … But before worrying about the demands of machines, many households will be able to watch Euro matches (oops, taboo) with their toes. fan-shaped. Facebook AI Research, the basic research laboratory of Mark Zuckerberg’s firm, unveiled several works at the Facebook AI Innovation Summit on Wednesday.
The future made in Facebook could well solve the problem of the mental load and, who knows, eliminate the arguments of couples around the thankless tasks of everyday life. Habitat 2.0, the 3D simulation platform developed by the Facebook laboratory in partnership with the 3D specialist Matterport, opens the doors to a future where machines could take over a large part of domestic tasks: filling the fridge, taking out the trash cans, help find car keys …
Fairies of the home
“We are working on artificial intelligence having a physical form [embodied AI] such as robots that influence the world and interact through language, explains Dhruv Batra, researcher at FAIR. Among these AIs, there are also virtual agents, such as assistants, who also have language and who can act in the virtual world. In both cases, the agent perceives the world from an “egocentric” point of view. He must, like the human being, build a memorial representation of what he has seen, remember the places where he has been, where he has seen certain objects and where he has met people. The idea is to achieve a behavior similar to that of humans who, in view of their past experience, are able to anticipate in an apartment where they have never set foot that, for example, a fridge is found in the kitchen.
Ultimately, we can ask the robot to check that our laptop is on our bedside table and bring it to us, but this requires developing an AI capable of fully understanding the request and then identifying the object. And to achieve such a result, it must be trained using large databases. This is where the Habitat 2.0 simulator plays a key role. It allows, thanks to a database of 1000 models of apartments, offices or commercial spaces designed in three dimensions, to train artificial intelligence to become real fairies of the house in an extremely realistic context.
These 3D models are dynamic, that is, objects and furniture can be moved. You can open the cupboards and put the fridge away. Researchers have the ability to easily change the environment and observe how the robot reacts to an object it has never seen before. Training the machine in the virtual world makes it possible to solve the problem of infrastructures, to put it in various contexts before considering deploying it in the real world.
“There is still a lot of work”
“We have developed a task for storage, continues the scientist, the AI moves in the different spaces of the accommodation, picks up the objects to put them back in the right place. There is a task for errands where the robot stores food in the fridge, on the counter or in the cupboards. Likewise, we are working on a scenario where the machine sets the table. And we can say that our greatest accomplishment from a robotics point of view is that our 3D simulator is much faster than any other simulator available in open source today ”.
Between the Roomba-type robot vacuum cleaners we see today and those who will spontaneously put our socks in the dirty laundry basket, the road still seems long. We are clearly not going to see them next month. The robot’s navigation in dynamic spaces, where several people are moving at the same time, is not ready. Likewise, controlling the robot arm to grab objects in a closet or grabbing soft objects are still far from the mark.
“There is still a lot of work to do,” concedes Dhruv Batra. If we don’t have a date before we consider handing off all the daily gruesome chores to our metal friends, the Habitat 2.0 simulator could kick-start.