Sam Altman, CEO of OpenAI, stated last year that the world has yet to experience its "humanoid robot moment," although he assured that milestone is near. Meanwhile, the company has been quietly advancing in that direction. According to sources cited by Business Insider, OpenAI has discreetly established a humanoid robotics lab in San Francisco, in the same building as its finance team. There, nearly 100 data collectors are working to teach a robotic arm to perform household tasks as part of a broader project to develop humanoid robots. OpenAI had previously ventured into robotics in its early days, achieving a robotic hand that solved a Rubik's cube, but closed that project in 2020 to focus resources on other areas. The new lab, so far little known, represents a renewed interest in hardware and robotics, although these projects are not yet central to the company's mission. Recently, OpenAI requested proposals from U.S. manufacturers to partner in the development of consumer devices, robotics, and cloud data centers, without detailing budget or timelines. Additionally, the company plans to open a second lab in Richmond, California, according to a job posting from December. In the San Francisco lab, there is a humanoid robot described as "similar to iRobot," although it is rarely used. The main focus is on the teleoperation of Franka robotic arms, controlled by 3D-printed devices called GELLOs. These arms perform tasks such as putting bread in a toaster or folding clothes. The program began in February by teaching the robot to place a rubber duck in a cup and has progressed to more complex tasks. Unlike companies like Tesla, which bet on spectacular demonstrations with complete humanoid robots and motion capture suits, OpenAI opts for a more discreet strategy, prioritizing large-scale data collection through contractors to train robotic arms in everyday tasks. OpenAI has invested in other robotics companies, such as Figure, 1X, and Physical Intelligence. In 2024, it announced a collaboration with Figure to develop "next-generation AI models for humanoid robots," although this partnership dissolved in early 2025. OpenAI's current approach is based on massive data collection, in contrast to the reinforcement learning it employed in the past. According to experts cited by Business Insider, the key is to obtain large datasets to train AI algorithms capable of performing physical tasks. OpenAI's method, which uses GELLO controllers assigned directly to each robotic arm, could offer cost and efficiency advantages over motion capture systems. The lab operates in three shifts and has several active workstations 24 hours a day. Cameras record both operators and robots, and productivity is measured by the amount of useful data hours generated. This model resembles the data labeling process that OpenAI and other companies have used to train large language models. OpenAI also uses the collected data to train robots in virtual simulations and regularly evaluates the performance of the robotic arms. It is unclear when the company will be able to translate these advances into a complete humanoid robot or how its approach will compare to companies investing in comprehensive humanoid systems. According to experts, the project is still in an early phase, but the infrastructure developed by OpenAI could be adapted to different types of robots and represent a significant advance in the integration of AI and robotics. (Source: businessinsider.es)