OpenAI has restructured its dedicated robotics division, reassigning its researchers to various internal teams. The company stated that this move is intended to accelerate progress in robotics rather than indicating a withdrawal from the field. The primary goal is enabling its broader, general-purpose artificial intelligence models to operate and control physical machinery, including humanoid robots.
Organizational Changes and Strategy
The company confirmed the reorganization through a company spokesperson, who emphasized that robotics remains a central component of OpenAI’s long-term vision. The shift was characterized as a sign of commitment rather than a retreat. OpenAI has moved its robotics research staff into the main research team, which reports to Chief Research Officer Mark Chen. Meanwhile, the teams focused on applied robotics and hardware have been transferred to the larger Applied organization, overseen by OpenAI President Greg Brockman.
This strategic realignment is predicated on the belief that general-purpose models, which are trained on data covering language, code, and vision, can be directly applied to “embodied AI.” This capability allows AI to manage physical devices, such as autonomous vehicles, drones, or robots, potentially eliminating the need for highly specialized AI models designed solely as the “brains” for specific machines.
Technical Advancement and Industry Analysis
The shift reflects a growing trend across the AI sector: the integration of massive, general AI models into the physical world. Previously, OpenAI had maintained a standalone robotics team but wound it down in 2020. The company later rebuilt this team in 2024 and formed a partnership with the startup Figure, a collaboration that concluded in February 2025 when Figure announced it would develop its own robot AI. Despite these changes, OpenAI stated its continued commitment to developing general-purpose robots, with CEO Sam Altman targeting a public demonstration in 2027.
Industry experts interpret the move as a significant philosophical pivot. Guanya Shi, an assistant professor at Carnegie Mellon University’s Robotics Institute, viewed the reorganization as an indicator that OpenAI views robotics “no longer a special problem.” Instead, he suggested the challenge can be addressed as a downstream application of general intelligence.
This view was supported by OpenAI’s internal logic. The company noted that initially, progress was thought to require building upon models like Sora, which needed an implicit understanding of physical world mechanics to generate realistic videos. However, the company reported that its general-purpose frontier models, including GPT-6 Astra, demonstrated direct applicability to robotics, motivating the disbandment of the dedicated robotics team.
Testing these general models has shown promising results. In an Nvidia study, an agent powered by Astra achieved a 74.8% success rate on a simulated manipulation benchmark. This performance figure was compared to 51.0%, which was reported by the best specialized robotic models on the same leaderboard.
While the general capability of models is impressive, experts also highlighted remaining challenges. Russ Tedrake, an MIT professor and cofounder/CEO of Walden Robotics, pointed out that robotics firms retain a distinct advantage: robots gain valuable, real-world work experience and site-specific data on factory floors. This field data, Tedrake explained, enhances the value of foundational models, allowing faster adaptation to specific tasks and environments.
Financial Pressures and Future Outlook
The organizational overhaul also occurs amid financial pressures. OpenAI has been under pressure to reduce costs and focus on core, profitable products. The company reportedly recorded heavy operating losses totaling $10.9 billion in the most recent quarter. As OpenAI prepares for a possible Initial Public Offering (IPO) in early 2027 and seeks a new funding round, streamlining its efforts has become critical.
While general models are powerful, they also present logistical hurdles. As Shi noted, large language models are resource-intensive, requiring constant cloud connectivity. This maintains a perceived need for smaller, specialized models capable of running locally on a robot’s onboard computer without reliance on the internet.
Despite the internal restructuring, former members of the robotics team at the recent CoRL conference offered a compelling answer to one of the industry’s hottest questions: the value of specialized expertise as general-purpose AI models continue to advance.
