His face has no eyes, nose, or mouth. Only a screen, a digital banner veiling his face horizontally; almost a grating allusion to our overconnected age. When turned on, it warns “operate using OpenAI”, the creator of ChatGPT. This will quickly become apparent. The robot “One” from the company Figure was the subject ofa new demonstration in mid-March. “Hey Figure One, what do you see now?” an engineer asks him. “I see a red apple on a plate in the center of the table, a drainer…”. Well planted on its two legs, the robot perfectly describes the space in front of it, in a slow and metallic voice, as if coming from its bowels. “Can I have something to eat?” the engineer retorts. “Of course”. “One” then extends his long arms and grabs the apple from his hands. “Can you explain to me why you did that?”, questions the human, handing him a box and ordering him to throw some trash into it. The robot simultaneously easily executes the cognitive task – explaining its action – and the physical task – clearing up.
Figure is the hot new robotics start-up. The American company has just raised $675 million from prestigious investors: OpenAI, of course, as well as Microsoft and Amazon. The big names in “Big Tech”. She currently embodies the dazzling progress of the discipline. As the demonstration proved, the advent of generative AI has something to do with it. Technology is currently astonishing the whole world. For the first time, by speaking to it in natural language, a machine responds to a question or an order on which it has not been specifically trained. While we thought these AIs were incapable of performing well in manual tasks, technological developments are already proving the opposite.
In the tech sector, the same idea is spreading: humanoid robots with arms and legs, like those in Figure One, would be a good way to take advantage of this potential. They could fetch items, assist in warehouses or hospitals, patrol oil platforms or sewers into which we always hope to send fewer mortals. “Decline many more actions, be more versatile,” adds Matthieu Cord, professor of robotics at the Sorbonne. Why not even, one day, make ourselves useful in our homes. “Generative AI will allow robots to evolve with humans,” insisted Vincent Vanhoucke, the boss of Google DeepMind, in an interview with L’Express last summer. Jensen Huang, iconic boss of computer processor designer Nvidia, and also a Figure investor, recently admitted at an event that “the ChatGPT moment for robots may be around the corner.”
Dexterity, vision…
The idea of giving brains to robots resembling humans is not new: science fiction was the first to imagine it, from Terminator to C3PO. The tests so far had not been conclusive. Not very talkative, often not very maneuverable, robots have mainly spread across industry in the form of articulated arms or goods conveyors. Like tools, with very codified rules. LLMs – the basis of generative applications – have changed the rules of the game. They function as giant databases for robots. Enough to allow them to “manipulate objects that they have never seen, to move in diverse environments that have never been visited”, Vincent Vanhoucke also explained to us. LLMs, with their mastery of language, are also used for the decomposition and planning of tasks, as well as their translation into movement.
Robotics is riding on other technological advances. “The quality of vision: now, robots see as well as humans,” says Matthieu Cord. Maybe even better. The result, in particular, of the numerous works carried out to create autonomous vehicles capable of identifying pedestrians, road markings or signs at high speed. Nvidia announced this week the publication of a new learning model exploiting these qualities, called Gr00t. “Today, these robots are evolving into self-learning machines that feed on various data sources such as sensor readings, camera images, voice/text commands and even human demonstrations,” explains Deepu Talla, vice president of the robotics section at Nvidia.
The sensory capabilities of machines are another area of this work. Nicolas Mansard, director of robotics research at LAAS-CNRS, and developer of Talos, a French humanoid robot, discusses current studies on the “skin” of robots. In reality, a system of sensors, placed on the arms and torso, in order to gauge the weight and material of the elements. “Robots in factories, like the arms that assemble cars, do not feel the efforts they exert,” recalls Nicolas Mansard. Asking them to carry fragile objects is therefore risky. This also poses a problem “if we wish to accomplish specific tasks, for example, engraving or grinding. It is then necessary to quantify the efforts required of the robot,” specifies the scientist. Those who do not feel their strength are, finally, more dangerous for humans.
Sports robots
The robots’ movements, meanwhile, have improved. Apptronik, with its humanoid Apollo, can move with 25 kilos of load on its arms. Phoenix, from the company Sanctuary AI, knows how to scan a barcode using a scanner. With even greater dexterity, a Tesla Optimus knows how to handle an egg without breaking it. In January, Figure showed his robot serving himself a cup of coffee using a capsule machine.
Robots even demonstrate phenomenal agility, like the quadrupeds developed by Boston Dynamics. “These devices are now very mature. They work well on all types of terrain with a lot of autonomy. They are able to fall and get up easily, to venture into nature or sites that have not been designed for robots”, distinguishes Nicolas Mansard. Quite the opposite of humanoids, who sometimes need heavy motors at their ankles to lift their legs, and are overall very unstable. “The humanoid rests on a very small support base and it doesn’t take much to make it fall,” recalls Serena Ivaldi, research director at Inria. “Besides, Figure, in his videos, performs rather classic walks, which we know work in certain conditions.”
This is what makes humanoid robots still a hard sell. “Seeing it as a product seems presumptuous to me,” says Nicolas Mansard. The object is still expensive – Tesla had mentioned a cost of around $20,000 for its Optimus. “Industrialists will wonder if the cost is justified by the application,” underlines Serena Ivaldi. You should also keep in mind that robots have a battery, which generally lasts one to four hours. And in the event of breakdown or breakage, which frequently happens in the event of a fall, they must be repaired.
There remains room for progress for humanoid robots, particularly in mechatronics, movement, but also in intelligence. With LLMs, the success rate of experiments by Google DeepMind engineers fluctuated between less than 50% – reasoning – and more than 75% – symbol recognition. Reliability is not yet there. The same mystery surrounds Figure. “As long as there is no scientific article, we cannot really know,” summarizes Serena Ivaldi, who recalls that the videos published by companies like Figure or DeepMind are only selected pieces, among hours and hours of testing. They are made to impress. Let’s face it, it’s a success.
.