Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHacker News AI→ original

OpenAI Disbands Robotics Team: GPT-6 Astra Solved 74.8% of Tasks in Simulation vs. 51%

OpenAI has disbanded its standalone robotics team because its general-purpose GPT-6 Astra turned out to be capable of controlling robots without specialized training. In simulation, an agent built on Astra completed 74.8% of object-manipulation tasks, while the best specialized models on the leaderboard managed 51.0%. Note, though, that this was a virtual robot, not a real one. Teams building physical robots should test how far an off-the-shelf model can take them.

AI-processed from Hacker News AI; edited by Hamidun News
OpenAI Disbands Robotics Team: GPT-6 Astra Solved 74.8% of Tasks in Simulation vs. 51%
Source: Hacker News AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHacker News AI ↗
Editor-in-Chief
Key metric #1Researchers moved to Mark Chen's team, and the applied team moved to Brockman's organization
Key metric #2Altman is promising the first public demonstration of a humanoid in 2027
Key metric #3According to media reports, OpenAI's quarterly loss was $10.9 billion

OpenAI has dissolved its standalone robotics team, and the reason is simple: the company's general-purpose model, GPT-6 Astra, which was trained on text, code and images, turned out to be capable of controlling robots directly. In the test, researchers connected it to a virtual robot. Astra received camera images, the robot's state data and an interface for commands. The agent completed 74.8% of object-manipulation tasks (picking up, moving and placing objects), while the best specialized robotics models on this leaderboard score 51.0%. If a general-purpose model outperforms a "brain" built only for robots, a separate team looks unnecessary. For anyone building robots or investing in them, this is an important signal.

What exactly happened

The company confirmed the reorganization to journalists and stressed that robotics remains part of its long-term strategy and is not a retreat. The researchers were moved into the core research team led by chief research officer Mark Chen. Applied engineers and hardware specialists were sent to the larger Applied organization, which reports to president Greg Brockman.

The reshuffle was so recent that OpenAI had already registered booth No. 208 at the CoRL conference in Austin. The booth was empty and had no logo, but former members of the team still walked the hall and gathered people around them. The main question of the conference is how valuable robotics experience will remain as general-purpose models become stronger.

The direction has a difficult history. OpenAI shut down its first robotics team back in 2020: there was too little robot data, and the path to AGI through large language models seemed more efficient. In 2024 the team was reassembled, and the company announced a partnership with the startup Figure on models for humanoids. In February 2025, Figure said it would build its own AI for its robots. OpenAI now says it is developing its own humanoid, and Sam Altman is aiming for the first public demonstration in 2027.

Why a general-purpose model suddenly became useful

According to an OpenAI spokesperson, early in the year researchers believed the best path to robots ran through models like Sora. To generate plausible video, such a model must implicitly understand the physics of the world. But it then became clear that ordinary frontier models, including GPT-6 Astra, can be applied to robotics directly. That realization pushed the company toward the reorganization.

Carnegie Mellon University professor Guanya Shi says general-purpose models already handle tasks that once required dedicated training. The bottle is in a different place, the lighting has changed, the model needs to work out the order of steps or find the step where things went wrong, and they can manage this. In his view, OpenAI now sees robotics as "no longer a special problem" but as one more application of general intelligence, like legal, financial or software tools built on the same base models.

There is also a financial context. Altman last year ordered the company to drop "side quests," and OpenAI has already shut down several projects, including Sora, to redirect computing resources to its core products. According to media reports, the loss for the last quarter was $10.9 billion, and the company is looking for a new funding round and preparing for a possible IPO in early 2027. The decision therefore looks both like an engineering choice and an economic one.

What this means for startups and where the limits are

This approach challenges the idea on which robotics startups raised billions: a robot needs its own "brain." MIT professor Russ Tedrake, co-founder and CEO of Walden Robotics, disagrees: robotics companies have something labs do not, namely experience operating in real environments such as factory floors. Data from those settings helps adapt systems to a specific facility, and the better the base model, the faster they can learn from that data.

The limitations are serious. The 74.8% result was achieved in simulation, not on a real robot. Large models are slow and run from the cloud, so a robot needs a constant network connection. Shi still sees a role for compact specialized models that run on an onboard computer without internet access. Pete Florence of the startup Generalist believes it is still unclear which kind of model will win in robotics.

The practical takeaway for teams building robots: before assembling your own control model, check how far you can get with an off-the-shelf general-purpose one. The value is shifting toward real deployment data, reliability and operation without the cloud.

Story

Key Questions Answered

What exactly happened?

The company confirmed the reorganization to journalists and stressed that robotics remains part of its long-term strategy and is not a retreat. The researchers were moved into the core research team led by chief research officer Mark Chen. Applied engineers and hardware specialists were sent to the larger Applied organization, which reports to president Greg Brockman. The reshuffle was so recent that OpenAI had already registered booth No. 208 at the CoRL conference in Austin.

Why a general-purpose model suddenly became useful?

According to an OpenAI spokesperson, early in the year researchers believed the best path to robots ran through models like Sora. To generate plausible video, such a model must implicitly understand the physics of the world. But it then became clear that ordinary frontier models, including GPT-6 Astra, can be applied to robotics directly. That realization pushed the company toward the reorganization.

What this means for startups and where the limits are?

This approach challenges the idea on which robotics startups raised billions: a robot needs its own "brain." MIT professor Russ Tedrake, co-founder and CEO of Walden Robotics, disagrees: robotics companies have something labs do not, namely experience operating in real environments such as factory floors. Data from those settings helps adapt systems to a specific facility, and the better the base model, the faster they can learn from that data. The limitations are serious.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…