OpenAI disbands independent robotics team to merge with core research
The company is shifting its robotics staff into general foundation model divisions to leverage existing AI progress for physical tasks.
The short version
- OpenAI disbanded its independent robotics team and moved staff to core research and the Applied division.
- The move aims to let robotics research benefit from the latest progress in general foundation models.
- OpenAI previously recruited heavily for robotics, offering some roles base salaries up to $500,000.
- The company continues to build data collection labs and prepare a hardware supply chain for 2027.
- Competitors like Google and NVIDIA are still developing dedicated models alongside general reasoning systems.
Why it matters
This shift indicates that the industry may be moving away from building separate 'robot brains' in favor of using general AI to power physical actions. For the reader, it means the path to functional robots might be faster if it relies on the same technology powering ChatGPT.
A sudden internal reorganization
OpenAI recently made the abrupt decision to disband its independent robotics research team. While the company had been aggressively recruiting for these roles, it is now moving the staff into different departments. Robotics researchers are joining the core research department led by Mark Chen, while hardware and application personnel are moving to the Applied division.
This restructuring follows a period of intense investment in the field of embodied intelligence. In September, the company was hiring on a large scale, with some positions offering a base annual salary of as high as 500,000 US dollars. The company also operates a data collection lab in San Francisco with around 100 staff members working in three shifts.
The move comes despite OpenAI's public commitment to robotics as a long-term goal. Sam Altman previously stated that the company expects to present an impressive robotics demonstration in 2027. By integrating the team, OpenAI aims to accelerate R&D progress by aligning robotics more closely with its primary AI models.
The shift toward general models
The reorganization stems from a change in technical strategy regarding how robots learn. Researchers previously thought using video generation models like Sora was the best way to teach robots the laws of physics. However, newer models like GPT-6 Astra have shown the ability to directly control robots, leading to the current merger.
By merging robotics into the general foundation model team, OpenAI can allow robotics research to share the training results of its core models. This approach suggests that robotics is being treated as a downstream application of general intelligence. It allows the company to re-evaluate the return on investment for developing entirely separate robotics foundation models.
Other companies are observing similar trends in the industry. For example, Unitree Robotics reported that connecting robots to GPT-6 allowed them to perform tasks like painting without complex motion programs. These general models can now handle visual understanding, spatial reasoning, and task planning, which previously required specialized training.
Data collection and hardware hurdles
Despite the organizational shift, OpenAI is still heavily invested in the physical infrastructure of robotics. The company is building a second data collection lab in Richmond, California, to gather the information needed for robot training. Unlike text models, robots require massive amounts of data on physical operations like grasping, touching, and moving objects.
The San Francisco lab uses 3D-printed teleoperation controllers to allow humans to operate robotic arms. These systems record scenarios like folding clothes or putting bread into toasters to create training data. This process is expensive and requires significant equipment, personnel, and on-site management to maintain a steady flow of information.
OpenAI is also preparing for the hardware side of the business by seeking manufacturing suppliers. The company recently issued a call for U.S. suppliers for components like motors, gearboxes, and power electronics. While they plan to select these suppliers in 2027, the specific amount of procurement contracts has not been disclosed yet.
Challenges of latency and cost
Using general models for robotics is not without significant technical challenges, particularly regarding speed. Research from NVIDIA showed that a direct control scheme using GPT-6 Astra took an average of 18.56 seconds per inference. In a real factory setting, conveyor belts and equipment cannot wait for a model to finish thinking for that long.
High costs also present a barrier to using general models for every single robot movement. One simulation test showed that a single round of successful tasks consumed about 30.9 million Tokens, which equates to an estimated API cost of 36.40 US dollars. These costs are currently for simulations and do not reflect the actual operating costs of physical robots in a production line.
Because of these issues, some competitors are still focusing on dedicated models for specific tasks. Google DeepMind, for instance, uses a multi-model approach where one model understands the production process while another is specifically trained for robot motions. This hybrid approach aims to balance the reasoning power of general AI with the speed and reliability of dedicated systems.
What to watch
- Whether OpenAI will eventually spin off its robotics and hardware businesses into a separate company.
- How the 2027 robotics demonstration will be structured given the new internal team layout.
- The success rate of using general models for real-time factory tasks compared to dedicated motion models.
Try it
Look up the official OpenAI careers page to see which robotics-related engineering positions are still being publicly listed.
Sources and links
- Reportingeu.36kr.com: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by eu.36kr.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
