OpenAI disbands robotics team, folds work into core research

OpenAI disbands robotics team, folds work into core research

OpenAI has disbanded its dedicated robotics team and reassigned its staff across its research and applied organizations, the company confirmed to Fortune. OpenAI says the restructuring is intended to accelerate robotics development. According to Fortune, the company is developing its own humanoid robot, with CEO Sam Altman targeting a first public demonstration in 2027.

The OpenAI robotics team’s research staff now sit within the main research organization reporting to Chief Research Officer Mark Chen. Applied robotics and hardware teams have been assigned to the larger Applied organization, which reports to President Greg Brockman. The change preserves those functions but removes their separate organizational home.

OpenAI told Fortune that its general purpose frontier models, including GPT-6 Astra, had shown they could be applied directly to robotics. The company said those results prompted the reorganization, replacing an earlier research approach built around models such as its video generator Sora.

A different route to humanoid intelligence

This is another restructuring of OpenAI’s robotics effort. Fortune reports that the company wound down its original robotics team in 2020, then rebuilt a dedicated team in 2024 and announced a partnership with Figure to develop models for humanoid robots. Figure ended that partnership in February 2025, saying it would develop its own robot AI.

The current decision challenges the premise that robot intelligence necessarily requires a separate model development organization. Carnegie Mellon University robotics researcher Guanya Shi told Fortune he interpreted it as OpenAI treating robotics as a downstream application of general intelligence rather than a distinct research problem.

Fortune cited an Nvidia study in which an Astra powered agent achieved a 74.8% success rate on a simulated manipulation benchmark, compared with 51.0% reported by the best specialized robotic models on that leaderboard. Those results concern simulated manipulation, not operation of a physical humanoid. In the experiments described by Fortune, the model received camera images, robot information and an interface through which it could command actions.

Shi nevertheless sees a role for smaller specialized models running onboard robots, pointing to the latency and network dependence of large models hosted in the cloud.

For humanoid developers collecting data in workplaces, stronger general models could complement rather than replace deployment experience. Russ Tedrake, an MIT professor and cofounder and CEO of Walden Robotics, told Fortune that robots working in real facilities gather data that helps them adapt to particular tasks and sites. Better foundational models, he argued, make that experience more valuable by allowing robots to learn from it faster.

Source: Fortune

Similar Posts