On January 23rd, last year, Sam Altman suggested that the world has not yet come to its ownHumanoid RobotAt the same time, however, he said that “this moment is in the making”. And behind the scenes, his artificial intelligence company is poking hands, driving this moment into reality。
According to the source, Business Insider is a very important source of informationIn the past yearOpenAI A human robot lab was set upI don't know. The laboratory is co-located with the company ' s finance team in a building in San Francisco and has about 100 data collectors. As part of an ambitious programme to develop human-shaped robots, they are beginning to train robotic arms to perform various household tasks。
According to 1AI, OpenAI had been involved in robotics in the early years, and the robots that had been developed at the other time could even complete the Cube's restoration. However, the project was suspended in 2020, when the corporate spokesperson stated that the initiative was “to refocus the team on other projects”。
The details of the internal operation of this new robotic laboratory were never disclosed in the media。
One person with knowledge of OpenAI strategic planning said that the company was promoting a number of new hardware projects at an early stage of R & D, of which robotics were one, although none of those projects was currently part of the company ' s core mission。
Last week, OpenAI sent a call for proposals to local American manufacturing companies seeking partners to work together on three main areas: consumer-grade equipment, robotics and cloud data centres. However, the company did not disclose the magnitude of the planned investment, nor did it give a time-line for the project。
According to informed sources, the laboratory has expanded more than four times since its inception in February 2025。
Last December, the company informed its staff that a second laboratory was planned for Richmond, California. During the same period, an “robots operator” call for OpenAI issued by a manpower outsourcing company revealed that the duty station was Richmond。
There was a human robot in the lab, described by a number of knowledgeers as a "like iRobot under the flag" product, but the robot was mostly left idle, and few people had seen it run. Most of the research and development efforts in the laboratory are concentrated on remote-controlled mechanical arms。
OpenAI allows data collectors to manipulate two Franka robots with a 3D print controller called GELLO. These metal metallic arms are equipped with a clamp to carry out household tasks such as putting bread in a toaster and folding clothes。
Following the start of the data acquisition project in February, the focus of research and development was on training Frank robots to put rubber duck in the cup, the training mission gradually shifted to more technical and complex operations。
The exposure of this laboratory has given the world a rare glimpse: The world's most influential artificial intelligence giant, which path is setting the robotic field。
Competing parties such as Tesla are keen to stage a luxurious technical demonstration, often with a training model that uses action to capture clothing and virtual reality to manipulate full-sized human robots. OpenAI has chosen a more low-key route - training a mechanical arm to perform basic tasks by scaling up data collection by outsourcing teams。
These two R & D pathways show precisely that top artificial intelligence enterprises still have a long way to go to build practical domestic robotics, and that development in this area remains highly dependent on human inputs。
Connected magazine reported last fall that OpenAI has started recruiting robot engineers. A search of information on the British Home Page revealed that at least 12 engineers were dedicated to the project。
According to informed sources, in December last year, a project manager suggested that laboratories needed to improve production efficiency in order to obtain effective training data for longer periods. Laboratory data collection targets have nearly doubled over the past months。
OpenAI has previously invested in a number of robotic enterprises, including Figure, 1X and Physical Industries. In 2024, the company entered into a partnership with Figure aimed at developing a “mutual human robotic artificial intelligence model”, but Chief Executive Officer Figure Bret Adcock announced in February 2025 that the two parties had terminated their cooperation。
When the first robot project ended in 2020, it was widely believed that OpenAI had shifted its focus to the development of ChatGPT. Today, companies clearly release signals in the area of expanding hardware and equipment, while their accumulated ChatGPT knowledge base may be used in the future to train robots to understand and interact with the real world。
Early projects are based on an enhanced learning model, which is a way of enabling robots to learn on their own through a “test-and-test” mechanism. Today, companies have turned to large-scale data acquisition-driven robotic training paths。
In an interview with Jonathan Aitken, a robot expert at Sheffield University, said: “All companies are fighting for ways to build large-scale data sets. We already have artificial intelligence algorithms that can do our job through big data trainingAnd the core challenge is how to access high-quality data sets. "
OpenAI’s data-collection strategy is very different from the robotics development model of Tesla, Figure and others. The technical route of the latter is for staff to record their entire body movements and operate full-sized human-shaped robots through action catch and virtual reality。
OpenAI's data-collection programme coincided with a study published in 2023 by researchers at the University of California in Berkeley. The study proposed a low-cost, scalable, long-range system for collecting data on mechanical arms. One of the researchers, according to leading British sources, joined OpenAI in August 2024 and is responsible for building a robotic intelligence hub。
Alan Finn, an artificial intelligence and robot expert at Oregon State University, said that the model was industry-standard configuration, and robots could learn by imitating operators. For its part, Aitken points out that OpenAI ' s GELLO controller strategy has a significant advantage over an artificial intelligence enterprise using action capture: not only is it less costly, but also because each controller is directly mapped with a mechanical arm, robots can more intuitively transform specific human actions into their own mechanical movements。
According to the source, OpenAI has dozens of workstations in San Francisco's lab, with three shiftsData collection is ongoing around the clockI don't know. The cameras synchronize the operations of operators and robots, and the performance assessment criteria for staff members are the time frame for the effective training data they produce。
This model, which relies on outsourced staff and is performance-based, is the same as the early expansion of the large language model to map operations by artificial intelligence enterprises, including OpenAI。
According to Finn: “Many companies are looking to be able to turn them into robotic action instructions with enough data to achieve a scale-up effect like ChatGPT.” He added, however, that “the idea has not yet been proven in practice”
According to the source, OpenAI is working on a new robotic workstation, equipped with a mechanical arm that will be more in tune with the movement patterns of human limbs。
In addition, the company will use part of the collected data for robotic training in a computer simulation environment and will regularly test the actual performance of the arm。
It is not clear how fast OpenAI plans to transform these data results into complete human-shaped robots; it remains to be seen whether this research and development path, which relies on a mechanical arm and low-cost input, can emerge from a large group of cash-backed all-human-shaped robotic systems。
“It is clear that this technology is at a very early stage. At the technical level, however, the system is an operational interface capable of adapting to a variety of robots and of being well designed and flexible.”