Loading page…
The company says the reorganization will accelerate robotics work. Simulation results support its bet, but physical tests and deployment constraints leave room for specialists.
Listen to this story
OpenAI is folding robotics researchers into its broader research and Applied organizations, betting that general-purpose models can control robots without a standalone robotics unit. The evidence is promising in simulation—an Astra agent scored 74.8% on a manipulation benchmark—but less decisive on physical hardware, where it fully completed only seven of 100 tasks. The shift puts more weight on foundation models and deployment data, while real-world reliability and latency remain constraints; OpenAI says it is still developing a humanoid robot, with a public demonstration targeted for 2027.
Research moves under Chief Research Officer Mark Chen, while applied robotics and hardware teams join the Applied organization led by Greg Brockman.
An Nvidia study reported Astra at 74.8% on a simulated manipulation benchmark, versus 51.0% for the top specialized models; the result does not establish physical-robot performance.
On StationeryBench, Astra fully completed seven of 100 tasks; its average progress score was 46, compared with 12 for Ai2's MolmoAct2.
OpenAI is putting its robot researchers inside the teams building its broader AI systems, rather than keeping a separate robotics unit. The company confirmed the reorganization to Fortune on November 13, saying it wants to accelerate robotics—not leave it behind—because its general-purpose models have shown they can be applied directly to robots.
At the start of the year, OpenAI's researchers expected robotics progress to build on work such as Sora. The reasoning, a company spokesperson told Fortune, was that generating realistic video required some understanding of the physical world. Since then, OpenAI says its general-purpose frontier models, including GPT-6 Astra, have demonstrated a more direct route to robotics.
That route gives the model information about a robot and its surroundings, then a way to issue commands. In experiments described by Fortune, researchers supplied Astra with camera images, robot information and a control interface. The model could then select actions, rather than merely describe what it saw.
Carnegie Mellon robotics professor Guanya Shi interpreted the reorganization as a sign that OpenAI sees robotics as “no longer a special problem.” He told Fortune the company likely views it as an application of general intelligence, much as shared frontier models already support specialized coding, legal and financial tools.
One result helps explain the enthusiasm. An Nvidia study cited by Fortune found that an Astra-powered agent achieved 74.8% success on a simulated manipulation benchmark. The best specialized robotics models on that leaderboard had reported 51.0%. Those were tasks performed by a simulated robot, not a physical machine—a material limit on what the comparison demonstrates.
Physical testing offers a more restrained picture. The Decoder described StationeryBench, a benchmark involving five desk-object tasks and dual-arm YAM robotic arms. Across 200 trials, Astra fully completed seven of its 100 tasks, while Ai2's specialized MolmoAct2 completed none. Astra's average progress score was 46, against 12 for MolmoAct2: a relative advantage, but few fully completed tasks.
Shi said general models can already cope with changes in object position or lighting and reason through a task's sequence or a failed step. But he also sees a continuing role for smaller, specialized models. Large models can be slow and depend on cloud hosting, requiring a network connection. Smaller models can run on the robot's own computer.
Robotics companies also retain something AI labs lack, according to MIT professor and Walden Robotics cofounder and CEO Russ Tedrake: experience operating robots in actual workplaces. The data gathered on factory floors and at other sites helps machines adapt to particular jobs. Better foundational models make that local data more valuable, he told Fortune, because robots can learn from it faster.
OpenAI has changed course on robotics before. It wound down an earlier team in 2020; cofounder Wojciech Zaremba later cited limited robot data and a more efficient route to artificial general intelligence through other methods. The company rebuilt a stand-alone robotics team in 2024. This time, it says robotics remains part of its long-term vision.
The hardware ambition remains: OpenAI says it is developing its own humanoid robot. Fortune reported that CEO Sam Altman is targeting a first public demonstration in 2027.
Loading discussion...
Join the conversation
Explain which jobs would make that tradeoff worthwhile.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.