OpenAI Disbands Dedicated Robotics Team, General-Purpose Models Replace Specialized Ones
nashnova research
OpenAI has dissolved its standalone robotics team and folded staff into its core research and applications divisions — because general-purpose models like GPT-6 Astra can now control robots directly, eliminating the need for a dedicated model. The move could redefine competitive moats across the robotics startup landscape.
Why disband now — didn't they say robotics matters?
OpenAI insists robotics remains a long-term core vision. What's gone is the standalone team, not the ambition.
The trigger: earlier this year the team planned to build on Sora (a video-generation model), but then GPT-6 Astra proved it could drive robots directly.
This means → they thought they needed a purpose-built key, then discovered the master key already works. The dedicated team lost its reason to exist.
Where did the people go?
Researchers move into the core research group, reporting to Chief Research Officer Mark Chen.
Applied-robotics and hardware staff join the larger applications division under President Greg Brockman.
In plain terms = no layoffs — the "robotics squad" was split into two bigger armies, one for foundational research, one for shipping products.
Is the general-purpose model really better than specialized ones?
An Nvidia study showed an Astra-powered agent hit a 74.8% success rate on a simulated robot-manipulation benchmark — versus 51.0% for the best prior specialized model.
But this score comes from a digital simulation, not a real robot in a real factory.
This means → the general model already crushes specialized models in the "virtual exam," but whether it can do the job on a real shop floor is still unproven.
Should robotics startups be worried?
Carnegie Mellon assistant professor Guanya Shi argues OpenAI may no longer treat robotics as a "special problem" — instead viewing it as a downstream application of general intelligence, much like coding, legal, and finance tools already run on general models.
But MIT professor Russ Tedrake pushes back: robotics startups have accumulated real-world deployment experience and task data from factory floors that big labs cannot easily replicate. The stronger the foundation model, the more valuable real-scene data becomes.
Shi also flags a practical bottleneck: large general models run slowly and need cloud connectivity, so small specialized models that run locally still have a role.
Is OpenAI also cutting costs?
The restructuring comes as OpenAI reported an operating loss of $10.9 billion in its most recent quarter. CEO Sam Altman has pushed to drop "side quests" and concentrate resources.
The company already shelved Sora and other projects to reallocate compute — folding the robotics team follows the same logic.
This reflects a core tension at OpenAI: enormous ambition, but a burn rate that forces constant triage — doubling down on the bets with the highest conviction.
How many times has OpenAI flip-flopped on robotics?
2020: first robotics team dissolved, citing limited training data.
2024: standalone robotics team rebuilt; partnership launched with humanoid-robot startup Figure.
February 2025: Figure partnership terminated.
Now: standalone team dissolved again, staff absorbed into main divisions. Altman previously set 2027 as the target for unveiling OpenAI's own humanoid robot.
In plain terms = built, disbanded, rebuilt, disbanded — all within four years. The only constant is the sentence "we take robotics very seriously."
市场有风险，内容仅供研究参考，不构成投资建议。