A door handle. A staircase. A keyboard. A screwdriver. Look around you: almost everything around you was designed for hands with five fingers and two legs.
To work in such a world, a robot therefore has every interest in resembling us. That is Sam Altman's reasoning. And he has just announced that OpenAI is going to get on with it.
What Sam Altman said
Asked on the Sources podcast about the possibility of OpenAI building a humanoid robot, the company's boss answered in three words: "We will do it."
He explained that the human form almost imposed itself, since the world is made for humans. But according to him, the robot's appearance matters far less than its "brain": the system that will allow it to see with its cameras, understand an instruction, plan its movements and adapt to the unexpected in the real world.
To Fortune magazine, he gave a timeline: an impressive demonstration in 2027, but no robots in the streets for several years yet.
What this robot would be for
OpenAI sets out a two-stage objective. In the short term, versatile robots capable of helping skilled workers build infrastructure. In the longer term, a personal robot, capable of being useful at home.
The company already lists 27 robotics job openings. It has said neither what its robot will look like, nor what the 2027 demonstration will show.
This is not a first for OpenAI. In 2019, the company made a splash with a robotic hand capable of solving a Rubik's Cube, trained almost entirely in simulation. But in 2021, it shut down its robotics team, for lack of enough data to go further, and focused its efforts on language models, which gave rise to ChatGPT. OpenAI's return to robotics shows how much the landscape has changed: the models have become powerful enough to consider giving them a body.
A race already well under way
OpenAI is arriving in a very crowded competition. Tesla is developing its Optimus robot, companies such as Figure AI and Boston Dynamics are multiplying prototypes, and Chinese manufacturers are moving fast.
All of them run into the same obstacle: data. To teach a robot to fold a garment or tidy a kitchen, you need millions of hours of examples. That is why Indian workers film their movements, a camera on their heads, for a few dollars an hour.
What to keep in mind
A demonstration is not a product. Videos of impressive robots have been multiplying for years, and yet, as we explained, household robots still do not exist. Grasping a fragile object, opening a stuck door, avoiding a running child: the real world remains an immense challenge.
Sam Altman admits it himself: the body is the easy part.
What we take away
OpenAI conquered screens with ChatGPT. It now wants to enter the physical world. It is a logical ambition: an intelligence that cannot act on the world remains limited.
But it raises a question we are digging into tonight: why do we want our robots to resemble us so much?