OpenAI's Secret Robotics Lab: Building the Future of Household Robots

In an unassuming San Francisco building, a team of approximately 100 workers is teaching robots how to perform everyday household tasks. This isn't a flashy demonstration or a public showcase—it's OpenAI's quietly expanding robotics lab, operating around the clock to collect the data needed to build functional humanoid robots.

The company has been scaling this effort over the past year, building on its earlier robotics work that was discontinued in 2020. Now, with renewed focus and a different approach, OpenAI is positioning itself to make the humanoid robot moment a reality.

The robotics lab operates out of the same building as OpenAI's finance team in San Francisco, a location that underscores the project's stealthy development. Insiders with knowledge of the program reveal that the lab has more than quadrupled in size since its launch in February 2025.

Workers in the lab use 3-D printed controllers, known as GELLOs, to operate two Franka robots—metal arms with pincers at the end. These robots, manufactured by German robotics research company Franka, perform a range of household tasks:

• Putting bread in a toaster
• Folding laundry
• Placing objects like rubber ducks in cups

The program began with teaching the Franka robot to put a rubber duck in a cup and has since shifted to increasingly sophisticated tasks. The lab runs three shifts and a few dozen workstations that collect data continuously, with cameras recording both the operator and the robot performing tasks.

OpenAI's strategy diverges significantly from competitors like Tesla and Figure, which often use motion capture suits and virtual reality headsets to train full-sized humanoid robots. Instead, OpenAI is taking a quieter, more scalable path focused on contractor-driven data collection.

Everyone is fighting for a way to develop large data sets. 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.

This approach mirrors a 2023 study from University of California, Berkeley researchers that described a low-cost system for collecting robotics data using teleoperated arms. One of those researchers joined OpenAI in August 2024 and now works on "Building the Robot Brain."

Experts suggest this method could offer advantages. Jonathan Aitken, a robotics expert with the University of Sheffield, notes that OpenAI's GELLO strategy is cheaper than motion capture suits and allows the robot to more easily learn how specific human movements translate into its own motions.

The lab's growth reflects OpenAI's commitment to scaling its robotics ambitions. In December, a project supervisor emphasized the need to increase productivity and efficiency to generate more hours of functional data. Over recent months, the lab has nearly doubled its expectations for data collection.

OpenAI has at least a dozen engineers working on the robotics project, according to LinkedIn profiles. The company has also invested in other robotics companies, including Figure, 1X, and Physical Intelligence. However, its 2024 partnership with Figure—designed to build "next generation AI models for humanoid" robots—ended in February 2025 when Figure exited the deal.

Workers are rated on how many "good hours" of functional training data they can generate, a performance metric that mirrors how AI companies have historically scaled data labeling for large language models.

OpenAI's robotics efforts align with CEO Sam Altman's vision. Last year, he 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.

The company is setting up new robot stations with robotic arms that more closely mimic how a human moves. Some of the data collected is used to train robots in computer simulations, and the robot arms are regularly tested to see how well they perform.

While the lab has a humanoid robot on display—described by multiple people as "iRobot-like"—it mostly collects dust, and few have seen it in operation. The vast majority of work remains focused on teleoperating robotic arms.

It does seem to be very early in the process. From a technical standpoint, it's a really beautiful and configurable interface to lots of different types of robots.

OpenAI's robotics push is part of a broader expansion into hardware. Last week, the company put out a request for proposals from U.S. manufacturing companies that could act as partners for its push into consumer devices, robotics, and cloud data centers.

The company has not specified how much it intends to spend or provided a timeframe for the work. A representative for OpenAI declined to comment on the program.

In December, the company told employees it plans to open a second lab in Richmond, California, with a job posting for a "robotics operator" listing Richmond as the location.

OpenAI's approach to robotics represents a fundamental question facing the industry: Can massive data collection from human operators create the foundation for functional household robots? The company is betting that a low-cost, scalable approach will eventually lead to a "ChatGPT moment" for robotics.

However, experts caution that this remains unproven. Alan Fern, an AI and robotics expert at Oregon State University, notes that while many companies hope collecting enough data will translate into robot motions, "that's something that hasn't been proven out yet."

As OpenAI continues to scale its data collection operations and expand into new facilities, the robotics world will be watching closely to see if this quiet, contractor-driven approach can deliver the breakthrough that has eluded even the most well-funded competitors.