OpenAI dissolves its dedicated robotics team to fold researchers into general AI groups

OpenAI announced that its independent robotics team has been dissolved and its researchers reassigned to other divisions, a move the company says will speed up the development of robots that are powered by its general‑purpose models. The shift reflects a growing industry belief that large language, vision and multimodal models can serve as the "brains" for embodied systems without needing separate, robot‑specific AI pipelines.

What organizational change did OpenAI make?

OpenAI confirmed to Fortune that the stand‑alone robotics team no longer exists as a distinct unit. Researchers formerly in that team are now part of the main research organization reporting to Chief Research Officer Mark Chen, while the applied robotics and hardware groups have been folded into the broader Applied organization under President Greg Brockman. The company stressed that the reorganization is intended to accelerate progress rather than abandon robotics altogether.

Why is the change important for the robotics field?

Historically, OpenAI maintained a dedicated robotics group in its early years, shut it down in 2020, and rebuilt a separate team in 2024 after a partnership with startup Figure. The latest restructure signals that OpenAI now treats robotics as a downstream application of its core AI research, rather than a niche problem requiring its own specialized team. By embedding robotics expertise into larger research and applied groups, OpenAI hopes to leverage breakthroughs in language, vision and multimodal models across all embodied AI use cases.

How are general‑purpose models being applied to robots?

OpenAI pointed to the performance of its GPT‑6 "Astra" model as the catalyst for the reorganization. In internal experiments, Astra received camera feeds, robot state information and a control interface, then issued commands directly to a simulated manipulator. An Nvidia study cited by Fortune reported a 74.8% success rate on a standard simulated manipulation benchmark, noticeably higher than the 51.0% achieved by the best specialized robotic models on the same leaderboard. Researchers also note that models like Astra can reason about task sequences, adapt to variations in object position or lighting, and diagnose where a step went wrong—capabilities previously reserved for models trained on large robot‑specific datasets.

What do experts think about the shift toward "embodied AI"?

Academics and industry leaders see both opportunity and caution. Guanya Shi, an assistant professor at Carnegie Mellon University, said the move suggests OpenAI views robotics as no longer a "special problem" but as a natural extension of general intelligence. Conversely, MIT professor Russ Tedrake highlighted that companies with robots already deployed in factories have an advantage: real‑world data collected on the shop floor can be combined with powerful foundation models to accelerate task‑specific learning. He also warned that many robots still need on‑device, low‑latency models because large language models remain cloud‑bound and relatively slow.

How can creators, developers, or researchers explore this trend with Grafy?

Grafy users can experiment with embodied AI concepts without building a physical robot. The Grafy Chat app lets you run multiple models side‑by‑side, including large language and multimodal models, and feed them simulated sensor data (e.g., images or state vectors) via a custom prompt. By chaining model outputs in Grafy Graph, you can prototype a loop where a model receives a visual description, decides on an action, and produces a control command—mirroring the Astra‑to‑robot pipeline described in the Fortune report. For deeper dives into the research papers behind these experiments, the Grafy Reader app can ingest PDFs, index the text, and answer questions about model architectures, benchmark results, or deployment strategies. This workflow lets developers prototype embodied‑AI reasoning before committing to hardware.

What does the restructuring mean for OpenAI’s roadmap and finances?

The change arrives amid a period of cost scrutiny at OpenAI. The company reported operating losses of $10.9 billion in its most recent quarter and has been trimming "side quests" such as the Sora video model to focus on core, revenue‑generating products. By consolidating robotics talent into existing research and applied teams, OpenAI can reuse compute resources for both language and embodied tasks, potentially reducing duplication. CEO Sam Altman still aims to showcase a publicly demonstrated humanoid robot by 2027, indicating that the long‑term vision remains unchanged even as the short‑term execution pivots toward integrated AI development.

---

Frequently asked questions

Did OpenAI shut down all robotics work?

No. OpenAI dissolved the separate robotics unit but reassigned its staff to the main research and Applied groups. The company says robotics remains a core part of its long‑term strategy and that the move is meant to speed up progress.

Are GPT‑6 Astra’s robot results limited to simulation?

The publicly cited results come from a simulated manipulation benchmark run by Nvidia, where Astra achieved a 74.8% success rate. OpenAI has not yet released comparable numbers from real‑world robot trials, though the company plans a public humanoid demo in 2027.

How can I try a similar embodied‑AI loop in Grafy?

Use Grafy Chat to run a large language model, feed it a textual description of a scene (or an image via a prompt), and ask it to output a control command. Connect that output to a downstream step in Grafy Graph that interprets the command as a simulated robot action. For paper‑level details, upload the relevant PDF to Grafy Reader and query the model about the experiment’s methodology.

---

Source: Hacker News — front page — OpenAI disbands its robotics team as its general models start running robots