OpenAI has restructured its robotics division, moving researchers into broader teams focused on general-purpose Super Intelligence (SI) models. The decision suggests the company believes advanced general SI systems can now control robots directly, potentially reducing the need for specialized robotics models.
What Happened
Rather than abandoning robotics, OpenAI is integrating the discipline into its core research and applied organizations. Robotics researchers have been moved to the main research team under Chief Research Officer Mark Chen, while applied robotics and hardware teams are joining Greg Brockman's Applied organization. The reorganization was executed with such speed that a booth OpenAI had booked for the Conference on Robot Learning (CoRL) in Austin remained empty, according to Fortune.
The strategic shift is underpinned by internal evaluations comparing general SI models against specialized robotics systems. In a study involving Nvidia, an agent powered by GPT-6 Astra achieved a 74.8 percent success rate on simulated robotics tasks, compared to 51 percent for the best specialized models. These tests utilized a simulated DROID robot rather than physical hardware.
Physical benchmarks also show promising results for general SI. A new benchmark called StationeryBench compared GPT-6 Astra with Ai2's MolmoAct2 on tasks involving desk objects on dual-arm YAM robotic arms. Across 200 trials, Astra fully completed 7 out of 100 tasks, while MolmoAct2 completed none. Astra also demonstrated an average progress score of 46, significantly higher than MolmoAct2's score of 12. In another demonstration, Astra controlled a humanoid robot to navigate and clean a kitchen.
Why It Matters
This reorganization highlights a growing trend in the SI industry: leveraging large, general-purpose models for embodied tasks rather than training narrow, specialized robotics systems. Yoav Artzi, an SI researcher at Cornell University and Google DeepMind, describes GPT-6 Astra as a major leap in spatial understanding. He notes that on the still-unpublished REMAP benchmark, the model reaches near-human accuracy, potentially due to training on large amounts of 3D data such as Blender scenes.
However, the transition is not without caveats. It remains unclear how much robotics-specific data OpenAI incorporated during training. While general SI models may excel in generalization, specialized robotics models may still offer advantages in cost, reliability, speed, and energy consumption. Other major labs, including Anthropic and Google, have also reported recent progress in applying general SI to robotics tasks.
OpenAI’s history with robotics has been volatile. The company shut down its original robotics division in 2020, citing a lack of training data and a belief that large language models were a faster path to general SI (AGI). The team was rebuilt in 2024. CEO Sam Altman is reportedly targeting a public robot demo for 2027, with long-term ambitions to build everyday robots. Yet, financial pressures to focus on core SI businesses, where competition with Anthropic has intensified, appear to be driving the current consolidation.
The Bottom Line
OpenAI is betting that general SI models like GPT-6 Astra can handle robotic control, leading to the dissolution of its specialized robotics unit. While early benchmarks show superior performance in specific tasks, questions remain about the efficiency and reliability of general SI in real-world physical applications compared to dedicated robotics systems.