On November 13, 2026, according to Fortune, OpenAI has disbanded its independently operated robotics team, folded robotics researchers into the main research team led by Chief Research Officer Mark Chen, and merged the applied robotics and hardware teams into the Applied organization led by President Greg Brockman. A company spokesperson said the restructuring reflects its commitment to the robotics field.
According to NVIDIA research data, a robot agent with GPT-6 Astra as its brain scored 74.8% on a simulated manipulation benchmark, while the best specialized robotics model scored 51%.
Three Times in Six Years: A History of Organizational Change
OpenAI had a robotics team in its early years and cut it in 2020. Co-founder Wojciech Zaremba attributed the decision to a scarcity of robot training data and the greater efficiency of pursuing AGI through the large language model path.
In 2024, OpenAI rebuilt an independent robotics team and partnered with Figure to develop robot AI models. The partnership ended in February 2025, and Figure developed models on its own. OpenAI then announced it would build its own humanoid robot, with CEO Sam Altman setting a goal of a first public demonstration in 2027.
In 2026, the independent team was disbanded again. This is the third structural adjustment in six years: from nothing to existence, from existence to disbandment, and from rebuilding to integration. This trajectory shows that OpenAI’s answer to the proposition that “robots need a dedicated AI team” continues to be rewritten as general model capabilities evolve.
The Core Bet: Will General Intelligence Eliminate the Specialness of Robotics?
A company spokesperson told Fortune that earlier this year the team believed building robot AI on top of video models such as Sora was the optimal path, because video models implicitly capture an understanding of continuous temporal sequences in the physical world. As general frontier models such as GPT-6 Astra surged in capability, researchers found they could directly connect general models to robot control interfaces: by providing camera images, robot state information, and an action control interface, the model can issue commands.
Guanya Shi, an assistant professor at Carnegie Mellon University’s Robotics Institute, interpreted the adjustment as a signal that “robotics is no longer a special problem.” OpenAI may believe that robot control is essentially a downstream application of general intelligence, requiring only that a general model be aimed at a specific task.
Shi also noted that cloud-based large language models have response-latency issues, and robot scenarios requiring real-time reactions still need lightweight local small models to work alongside them. A general large model can serve as a “strategic brain,” but it may not alone take on the function of “neural reflexes.”
DeepMind’s Opposite Bet: The Specialized Model Route
Three days before OpenAI announced the disbandment of its robotics team, Google DeepMind released Gemini Robotics 2.5 on November 10, a vision-language-action model designed for robotics scenarios. In contrast to OpenAI’s route, DeepMind chose to build a dedicated architecture for the physical world.
DeepMind’s announcement showed that in a proof of concept with Mercedes-Benz, Gemini Robotics 2.5 drove the Apptronik Apollo 2 humanoid robot to complete auto-parts sorting and packaging tasks, achieving a 96% success rate in a real factory environment.
The 74.8% versus 51% figures cited by OpenAI come from NVIDIA’s simulated manipulation benchmark, while DeepMind’s 96% comes from a real factory scenario. The two numbers point to different dimensions. Russ Tedrake, an MIT professor and co-founder of Walden Robotics, believes that real deployment data is the core moat for robotics companies.
Disbanding the Organization While Continuing the Direction
From an organizational logic perspective, disbanding the independent robotics team is an expression of technical confidence: when general model capabilities are strong enough, separately maintaining a team to “make robots smarter” becomes duplicate investment; it is better to fold personnel into the core research flow.
The challenges facing robotics include perception calibration, safety redundancy, and mechanical reliability, and these will not automatically disappear as general models become stronger. After the independent team is disbanded, there is currently no public answer as to who will take on these engineering problems.
An OpenAI spokesperson said the plan for a public humanoid robot demonstration in 2027 remains unchanged.
Two Routes and Unresolved Questions
The AI robotics field is in a contest over approaches: OpenAI is betting on the emergent capabilities of general frontier models and disbanding its dedicated organization to reduce redundancy; DeepMind is betting on a specialized architecture customized for the physical world and validating its effectiveness with real factory data. Both have data support and researcher endorsement.
Pete Florence, co-founder of robotics startup Generalist, judged that it is currently unclear whether any existing model can ultimately dominate.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接