OpenAI has quietly established a humanoid robotics laboratory in San Francisco over the past year, marking a significant return to robotics development after abandoning similar efforts in 2020. The lab, which operates out of the same building as the company’s finance team, now employs approximately 100 data collectors who are training robotic arms to perform household tasks as part of a broader initiative to build functional humanoid robots.
The facility has more than quadrupled in size since launching in February 2025, with plans to open a second location in Richmond, California announced in December. Unlike competitors such as Tesla and Figure, which showcase flashy demonstrations using full humanoid robots controlled by motion capture suits and VR headsets, OpenAI is taking a quieter, more methodical approach focused on contractor-driven data collection.
Data collectors use 3D-printed controllers called GELLOs to teleoperate two Franka robots—metal arms with pincers that perform household tasks like putting bread in a toaster or folding laundry. When the program began in February, work focused on simple tasks like placing a rubber duck in a cup, but has since progressed to increasingly sophisticated activities. The lab runs three shifts with dozens of workstations collecting data around the clock, with workers rated on how many “good hours” of functional training data they generate.
Last week, OpenAI issued a request for proposals from US manufacturers to partner on consumer devices, robotics, and cloud data centers, though the company didn’t specify budget or timeline. The company has at least a dozen engineers working on the project and has previously invested in robotics companies including Figure, 1X, and Physical Intelligence. However, Figure CEO Brett Adcock revealed in February 2025 that his company was exiting its 2024 partnership with OpenAI.
This data collection strategy mirrors a 2023 study from UC Berkeley researchers describing a low-cost, scalable system for collecting robotics data using teleoperated arms. One of those researchers joined OpenAI in August 2024. Experts note that OpenAI’s GELLO approach may offer advantages over motion capture suits—it’s cheaper and allows robots to more easily learn how human movements translate into their own motions. However, robotics experts caution that the company appears “very early in the process” and it remains unclear how quickly this arm-based data collection will translate into a functional humanoid robot.
Key Quotes
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.
Jonathan Aitken, a robotics expert with the University of Sheffield, explains the fundamental challenge facing AI robotics companies. This highlights why OpenAI’s data collection strategy is central to their approach and why the company employs 100 data collectors working around the clock.
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. That’s something that hasn’t been proven out yet.
Alan Fern, an AI and robotics expert at Oregon State University, captures the industry’s optimism and uncertainty. His comment reveals that OpenAI and competitors are betting on unproven assumptions that the data-driven approach that revolutionized language models will work equally well for physical robotics.
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.
Jonathan Aitken’s assessment of OpenAI’s GELLO system provides both encouragement and caution. While praising the technical approach, he emphasizes that OpenAI remains far from producing functional humanoid robots, tempering expectations about near-term commercial applications.
Our Take
OpenAI’s secretive robotics lab reveals a fascinating strategic pivot that could define the next phase of AI development. By applying the data-intensive methodology that made ChatGPT successful to physical robotics, the company is essentially betting that embodied AI will follow the same scaling laws as language models. However, the physical world presents fundamentally different challenges than text generation.
What’s particularly telling is OpenAI’s restraint compared to competitors. While Tesla hosts flashy robot demonstrations, OpenAI is grinding through unglamorous data collection with contract workers teaching arms to handle rubber ducks. This suggests either greater realism about the timeline to functional humanoids, or a different strategic vision focused on building foundational datasets rather than near-term products.
The reliance on human teleoperation to train robots also underscores an irony: the path to automated household robots runs through massive amounts of human labor. Whether this approach yields the transformative breakthrough the industry anticipates remains the trillion-dollar question.
Why This Matters
This development represents OpenAI’s strategic expansion beyond large language models into physical AI systems, potentially reshaping the competitive landscape in humanoid robotics. While companies like Tesla and Figure pursue high-profile approaches with full humanoid systems, OpenAI’s methodical, data-driven strategy leverages its core competency in training AI models with massive datasets.
The robotics push signals that leading AI companies believe the “ChatGPT moment” for robotics is approaching, though experts remain skeptical about whether data collection alone will produce the scaling effects that transformed language models. The reliance on contract workers and performance metrics mirrors how OpenAI scaled data labeling for ChatGPT, suggesting the company is applying lessons learned from its LLM success.
For the broader AI industry, this reveals how far even the most advanced companies remain from functional household robots and how much human labor still underpins AI development. OpenAI’s quiet approach contrasts sharply with competitors’ flashy demonstrations, potentially indicating a more realistic timeline for practical robotics applications. The race for robotics data could become as critical as the race for training data was for language models, with significant implications for manufacturing, household automation, and the future of work.
Related Stories
- OpenAI Lost Nearly Half of Its AI Safety Team, Ex-Researcher Says
- Sam Altman and Jony Ive Partner on AI Device, Target $1B Funding
- Elon Musk’s $97B OpenAI Bid: Strategic Move or Elaborate Troll?
- OpenAI’s Competition Sparks Investor Anxiety Over Talent Retention at Microsoft, Meta, and Google
Source: https://www.businessinsider.com/open-ai-robotics-lab-humanoid-robots-2026-1