OpenAI Disbands Its Robotics Team as Two Tests of GPT-6 Astra Score 74.8% and 22.48%

OpenAI has dissolved its standalone robotics team and moved its researchers into the main research group, a company spokesperson confirmed to Fortune on Nov. 13. The spokesperson said robotics remains central to OpenAI's long-term plans.

Fortune, which labelled the story an exclusive by Wen Shao, published the story at 12:05 p.m. Eastern on Friday. The only account of the change traces to Fortune and that spokesperson.

Where the robotics team went

Robotics researchers now sit in the main research team, which reports to Chief Research Officer Mark Chen. The applied robotics and hardware teams moved into OpenAI's Applied organisation, which reports to President Greg Brockman.

Group	New home	Reports to
Robotics research	Main research team	Mark Chen, Chief Research Officer
Applied robotics and hardware	Applied organisation	Greg Brockman, President

Fortune said OpenAI had booth #208 at CoRL, the academic robotics conference held in Austin this week. When Fortune visited, the booth was empty and carried no OpenAI branding. A few former robotics team members were there without a demo.

OpenAI did not say how many people were affected.

The 2020 shutdown and the 2024 rebuild

This is the second time OpenAI has closed a robotics team. It wound the first one down in 2020, when co-founder Wojciech Zaremba cited a shortage of robot data and a more efficient route to AGI, per Fortune.

OpenAI rebuilt the group in 2024 alongside a partnership with Figure. That partnership ended in February 2025, when Figure said it would build its own robot AI.

Fortune said OpenAI is developing its own humanoid robot, and that CEO Sam Altman is targeting a first public demonstration in 2027.

Two benchmarks, two very different Astra scores

The technical case for the move is GPT-6 Astra, OpenAI's general model, driving robots directly. Fortune cited an NVIDIA study in which an Astra-powered agent scored 74.8% on a simulated manipulation benchmark. The best specialised robotics models on that leaderboard scored 51.0%.

1. Astra-powered agent74.8 %
2. Best specialised robotics model51 %

Source: Fortune, Nov. 13, 2026

A separate paper gives a colder number. Twelve authors posted "An Unexpected Robot Policy" to arXiv on Sept. 21 (2609.24170), using Astra as the low-level controller with no task-specific fine-tuning on all 42 RoboDojo tasks. It averaged 22.48% success over 2,100 trials, which the paper says ranks above 40 public policies. GPT-5.5 averaged 0.88% and DeepSeek-Flash 1.92%.

The abstract page does not list the authors' affiliations. The paper says Astra does well on tasks needing semantic understanding and poorly on tasks needing high precision, dynamic control or two-arm coordination. It calls the results early evaluations.

The two numbers come from different benchmarks, so they do not contradict each other. They show how far the result depends on the test. Fortune's account does not name the benchmark the NVIDIA study used.

What outside roboticists said

Guanya Shi, an assistant professor at Carnegie Mellon University's Robotics Institute, said, per Fortune, that the move suggests OpenAI sees robotics as "no longer a special problem". He still sees a need for smaller models that run on the robot without a constant connection.

Russ Tedrake, an MIT professor and chief executive of Walden Robotics, said robotics companies keep an edge through deployment data, such as from factory floors. Pete Florence, cofounder of Generalist, said it is unclear which model will dominate robotics.

For a view from the hardware side, see Agility's debut on $1.78 million of revenue. For Nvidia's startup investing, see our report on its 121 deals in 2026.

Altman's 2027 target for a first public humanoid demonstration is the next date on record.