OpenAI has disbanded its standalone robotics team and reassigned its researchers to other divisions, marking a significant change in how the company plans to develop artificial intelligence for physical machines. The restructuring was confirmed by OpenAI to Fortune on November 13, 2026. The company’s robotics researchers are moving into its main research organisation, while applied robotics and hardware employees are joining the larger Applied division. OpenAI says the move is intended to accelerate robotics development rather than signal an abandonment of its ambitions in the field.
The decision reflects a growing belief within OpenAI that its general-purpose frontier AI models can increasingly perform tasks previously associated with specialised robotics systems. The company says models such as GPT-6 Astra have demonstrated capabilities that can be applied directly to robot control. Instead of maintaining a separate organisation dedicated to developing robotic intelligence, OpenAI wants its broader AI research teams to build models that can also operate machines in the physical world.
OpenAI Reorganises Its Robotics Division
Under the restructuring, OpenAI has distributed its robotics employees across two major parts of the organisation.
The robotics research team has been moved into OpenAI’s main research organisation, reporting to Chief Research Officer Mark Chen. Meanwhile, the applied robotics and hardware teams have joined the company’s larger Applied division, which reports to OpenAI President Greg Brockman.
The change effectively ends the standalone robotics team’s existence, although employees will continue working on robotics-related projects within their new departments.
OpenAI told Fortune that robotics remains an important part of its long-term vision. The company described the restructuring as a way to accelerate progress towards general-purpose robots by connecting robotics researchers more closely with teams developing its core AI models.
The timing of the reorganisation also attracted attention at the Conference on Robot Learning (CoRL) in Austin, Texas. OpenAI had registered a booth at the event, but Fortune reported that the designated space was empty when its reporter visited.
Why OpenAI Is Moving Away From a Separate Robotics Team
The central reason behind the restructuring is a shift in OpenAI’s approach to robot intelligence.
At the beginning of 2026, the company’s researchers believed that progress in robotics would benefit from technologies developed through projects such as Sora. The video-generation model requires an understanding of how objects, movement and physical environments behave, capabilities that could potentially support robotic systems.
However, OpenAI says its general-purpose frontier models have since demonstrated that they can be applied directly to robotics tasks.
This has encouraged the company to reconsider whether robotics requires a separate model-development organisation.
Traditionally, robotics companies have developed specialised AI systems trained on robot-specific data. These models learn how to recognise objects, understand physical environments and control mechanical movements.
OpenAI increasingly believes that improvements in general-purpose models could reduce the need for some of this specialised development.
The company’s strategy is to improve the underlying AI systems and use those capabilities across multiple applications, including coding, research, enterprise software and robotics.
GPT-6 Astra Demonstrates Potential in Robot Control
One development influencing OpenAI’s thinking is the reported performance of GPT-6 Astra in robotic experiments.
According to Fortune, researchers have tested the model by connecting it to robotic systems through camera images, robot-state information and a control interface. This allows the AI model to interpret its environment and issue instructions for physical actions.
In a study involving NVIDIA’s simulated robotics environment, an Astra-powered agent achieved a 74.8% success rate on a simulated manipulation benchmark.
That compared with 51% for the best specialised robotics model reported on the same leaderboard.
| Benchmark result | Success rate |
|---|---|
| GPT-6 Astra-powered agent | 74.8% |
| Best specialised model on the reported leaderboard | 51.0% |
The results suggest that general-purpose models may be capable of handling certain robotic tasks without requiring the same level of specialised training traditionally used in robotics.
However, an important distinction remains: the reported experiment took place in a digital simulation rather than on a physical robot operating in the real world.
Physical robots must deal with unpredictable environments, mechanical limitations, changing lighting conditions, sensor errors and safety requirements. Demonstrating strong performance in simulation does not automatically establish equivalent performance in real-world deployments.
OpenAI’s Robotics Journey: From Shutdown to Relaunch
The latest restructuring is not the first time OpenAI has changed its approach to robotics.
The company originally operated a dedicated robotics research team that explored developing general-purpose robots. In 2020, OpenAI wound down that effort.
At the time, co-founder Wojciech Zaremba cited difficulties obtaining sufficient robot training data and explained that other approaches to AI development offered a more efficient path towards artificial general intelligence.
The company subsequently concentrated heavily on large language models trained on extensive datasets, an approach that eventually produced ChatGPT in 2022.
However, OpenAI returned to robotics in 2024, rebuilding a standalone team and exploring ways to provide AI models to companies developing humanoid robots.
Its renewed interest included a partnership with Figure, a humanoid robotics startup. The collaboration focused on developing AI models capable of powering Figure’s robots.
That partnership ended in February 2025 after Figure announced that it would develop its own AI systems.
OpenAI continued investing in robotics research and began exploring the possibility of developing its own humanoid robot.
The November 2026 restructuring represents another major adjustment in the company’s strategy.
OpenAI Still Plans to Develop Humanoid Robots
Despite eliminating the standalone team, OpenAI has not announced an exit from robotics.
The company has previously indicated that it is developing its own humanoid robot, with CEO Sam Altman targeting a first public demonstration in 2027.
The project reflects OpenAI’s broader ambition to build AI systems capable of interacting with the physical world rather than remaining limited to software applications.
Humanoid robots could eventually perform tasks in factories, warehouses, commercial facilities and other environments designed around human movement.
However, building useful robots requires more than advanced AI models. Companies must also develop reliable hardware, sensors, actuators, battery systems and safety mechanisms.
OpenAI’s decision to move applied robotics and hardware employees into the Applied division suggests that these activities will continue within the company’s broader product and engineering organisation.
A Different Approach to the Future of Robotics
OpenAI’s restructuring reflects a wider debate within the AI industry about whether general-purpose models can become the intelligence layer for robots.
Companies developing specialised robotics systems have traditionally relied on models trained using large quantities of robot-generated data. These systems are designed to perform specific physical tasks and operate under particular environmental conditions.
General-purpose models offer a different approach. They can combine visual understanding, reasoning and planning capabilities to respond to changing situations.
However, specialised robotics companies retain an important advantage: real-world deployment experience.
Robots operating in factories and warehouses collect valuable data about physical environments, mechanical performance and practical tasks. This information can help specialised systems adapt to specific workplaces.
Researchers also point to technical limitations involving response speed and connectivity. Large AI models may require cloud infrastructure and can be slower than smaller models designed to run directly on a robot’s onboard computer.
As a result, general-purpose models and specialised robotics systems may continue to complement each other rather than one completely replacing the other.
The Bigger Picture
OpenAI’s decision to dissolve its standalone robotics team highlights a broader shift in the AI industry’s development strategy. Instead of building separate intelligence systems for every application, leading AI companies are increasingly exploring whether one general-purpose model can support multiple types of work, including software development, visual reasoning and physical machine control.
The strategy could reduce duplication in AI research and allow improvements in general models to benefit robotics applications more quickly. However, physical intelligence remains a difficult engineering challenge. Real-world robots need reliable hardware, fast responses, specialised control systems and extensive experience operating in unpredictable environments. Strong benchmark results in simulation are an important development, but they do not eliminate these requirements.
Looking Ahead
OpenAI will now pursue robotics through its main research and Applied organisations rather than a dedicated standalone division. The company’s progress will depend on whether general-purpose models can consistently control physical machines outside controlled simulations. Its planned humanoid robot demonstration in 2027 could provide a more concrete indication of how far the technology has advanced.
For the broader robotics industry, the decision raises questions about the future role of specialised robot intelligence. If general-purpose AI models become sufficiently capable, robotics companies may increasingly focus on hardware, deployment data and real-world engineering while relying on external foundation models for reasoning and planning. However, specialised models are likely to remain important for tasks requiring low latency, reliable control and operation without continuous cloud connectivity.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
