Last year, Sam Altman posited that the world hadn’t yet had its “humanoid robots moment” — but, he said, “it’s coming.” In the background, his AI company has been gearing up to make that happen.
OpenAI quietly built up a humanoid robotics lab over the past year, insiders with knowledge of the program told Business Insider. The lab operates out of the same building as the company’s finance team in San Francisco, and employs around 100 data collectors. They’re teaching a robotic arm how to perform household tasks as a part of a larger effort to build a humanoid robot.
OpenAI explored robotics during its early years and built a robotic hand capable of solving a Rubik’s Cube. The company closed the project in 2020; a company spokesperson said at the time that it had chosen to “refocus the team on other projects.”
The inner workings of the new robotics lab haven’t been previously reported.
One person with knowledge of OpenAI’s strategy said the company is working on several new hardware projects, including robotics, that are in the early phases of development, and so far, none are core to the company’s mission.
Last week, OpenAI put out a request for proposals from companies that manufacture in the US that could act as partners for the company’s push into consumer devices, robotics, and cloud data centers. The company did not specify how much it intends to spend or provide a timeframe for the work.
A representative for OpenAI declined to comment.
The lab has more than quadrupled in size since it launched in February 2025, insiders said.
In December, the company told employees it plans to open a second lab in Richmond, California. A December job posting for a “robotics operator” with the company’s contracting agency lists Richmond as the location.
The lab has a humanoid robot that multiple people described as “iRobot-like” on display, but the bot is mostly collecting dust, and few have seen it in operation. The vast majority of the work in the lab is focused on teleoperating robotic arms.
OpenAI has data collectors using 3-D printed controllers, called GELLOs, to operate two Franka robots. These metal arms have pincers at the end and perform household tasks like putting bread in a toaster or folding laundry.
When the data collection program began in February, work focused on teaching the Franka robot how to put a rubber duck in a cup. Since then, it has shifted to increasingly more sophisticated tasks.
OpenAI’s lab offers a rare glimpse into how one of the world’s most influential AI companies is approaching robotics.
Competitors like Tesla put on flashy demos and often train with full humanoid robots controlled by motion capture suits and virtual reality headsets. OpenAI is taking a quieter path, scaling contractor-driven data collection to train robotic arms on basic tasks.
Both approaches show how far leading AI companies remain from building functional household robots, and how much of that work still relies on human labor.
The race for data
Wired reported last fall that OpenAI had begun hiring robotics engineers. The company has at least a dozen engineers on the project, according to a review of LinkedIn profiles.
In December, a project supervisor said that the lab needed to increase productivity and efficiency to get more hours of functional data, people with knowledge of the program said. Over the past few months, the lab has nearly doubled expectations for data collection, they said.
OpenAI has previously invested in other robotics companies, including Figure, 1X, and Physical Intelligence. Its 2024 partnership with Figure was designed to build “next generation AI models for humanoid” robots, but Figure CEO Brett Adcock said in February 2025 that the company was exiting the deal.
When the first robotics project ended in 2020, it was widely believed that the company was focusing more intensely on ChatGPT. Now, OpenAI has indicated it plans to expand into devices, and it could use its ChatGPT knowledge base to teach a robot how to interpret and interact with the world.
The earlier program focused on reinforcement learning — a trial-and-error approach in which robots learn through a reward system. Now, the company is collecting large amounts of data to train the robots.
“Everyone is fighting for a way to develop large data sets,” Jonathan Aitken, a robotics expert with the University of Sheffield, told Business Insider. “We know we have AI algorithms that are capable of being trained to do stuff using big data sets. The issue has always been getting that data set.”
OpenAI’s data collection strategy differs from the robotics efforts at companies like Tesla or Figure, where workers record full-body movement and use motion capture suits and virtual reality headsets to operate full-sized humanoid robots.
OpenAI’s data collection efforts mirror a study published in 2023 by researchers from the University of California, Berkeley, that describes a low-cost and scalable system for collecting robotics data using teleoperated arms. One of the researchers joined OpenAI in August 2024 and works on “Building the Robot Brain,” according to LinkedIn.
Alan Fern, an AI and robotics expert at Oregon State University, told Business Insider it’s a standard setup that allows the robot to learn by mimicking its operator. Aitken said OpenAI’s GELLO strategy could present an advantage over AI companies that use motion capture suits. It’s cheaper, he said, and because each controller maps directly to a robot arm, the robot can more easily learn how specific human movements translate into its own motions.
In San Francisco, OpenAI runs three shifts and a few dozen workstations that collect data around the clock, people with knowledge of the lab said. Cameras record both the operator and the robot performing tasks, and workers are rated on how many “good hours” of functional training data they can generate.
‘Very early in the process’
The reliance on contract workers and performance metrics mirrors how AI companies, including OpenAI, have historically scaled data labeling for their large language models.
“A lot of companies are hoping if you collect enough of this data, you can translate it into robot motions, and they will get a scaling effect and have this ChatGPT moment,” Fern said.
“That’s something that hasn’t been proven out yet,” he added.
The company is setting up new robot stations with robotic arms that more closely mimic how a human moves, insiders said.
OpenAI also uses some of the data to train robots in computer simulations and regularly tests the robot arms to see how well they perform, the people said.
It’s unclear how quickly OpenAI plans to translate this data into a full humanoid robot, or how its low-cost, arm-based approach will fare against companies investing heavily in full humanoid systems.
“It does seem to be very early in the process,” Aitken said. “From a technical standpoint, it’s a really beautiful and configurable interface to lots of different types of robots.”
Do you work for OpenAI or have a tip? Contact this reporter via email at gkay@businessinsider.com or Signal at 248-894-6012. Use a personal email address, a nonwork device, and nonwork WiFi; here’s our guide to sharing information securely.
Read the full article here
