OpenAI dissolves robotics team, targets 2027 humanoid demo

By Rebecca Stone

OpenAI has disbanded its standalone robotics team, Fortune reported on Friday, Nov. 13, while a study run by Nvidia put a GPT-6 Astra agent at 74.8% on a simulated manipulation benchmark.

The reorganization places the robotics research group inside OpenAI’s main research team under chief research officer Mark Chen. Its applied robotics and hardware teams now sit within the Applied organization under president Greg Brockman. Fortune did not report how many people moved.

Why did OpenAI restructure its robotics operation?

OpenAI told Fortune that its researchers first viewed video models such as Sora as the best route into robotics. The company later found that its general-purpose frontier models could be applied directly to robots, a shift it cited to explain the reorganization.

A spokesperson said robotics “remains a core part of OpenAI’s long-term vision.” OpenAI is building its own humanoid, and CEO Sam Altman aims to demonstrate it publicly in 2027. The 2027 date is a roadmap target for a demonstration, not a reported commercial launch.

How did GPT-6 Astra perform?

In experiments described by Fortune, researchers gave an Astra-based agent camera images, information about the robot and a control interface for commanding its actions. The Astra-powered agent then reached 74.8% success on a simulated manipulation benchmark.

Nvidia’s study recorded 51.0% for the best specialized robotics models on the same leaderboard. That lead measures simulated manipulation rather than performance on deployed robots, a distinction that matters to specialists working with physical systems.

Does one general model solve the robotics problem?

Guanya Shi, an assistant professor at Carnegie Mellon University’s Robotics Institute, said the move suggests robotics is “no longer a special problem” but a downstream application of general intelligence. He nevertheless sees a role for small onboard models because cloud-hosted language models are slow.

Russ Tedrake, an MIT professor and co-founder of Walden Robotics, said robot makers retain an edge because they collect real-world deployment data. Pete Florence, co-founder of Generalist, said it remains unclear whether any existing model will dominate.

Google DeepMind took the specialized route in the same week, releasing its Gemini Robotics 2.5 vision-language-action model to partners. The release put a purpose-built robotics model against OpenAI’s broader approach just as OpenAI reorganized its own teams.

What does OpenAI’s second robotics reset mean?

This is the second time OpenAI has broken up a robotics group. The company wound down its first team in 2020; co-founder Wojciech Zaremba later cited a shortage of robot data. OpenAI rebuilt a standalone team in 2024 alongside a partnership with Figure.

Figure ended that partnership in February 2025 to develop its own robot AI. Former members of OpenAI’s disbanded team appeared at the company’s booth at the CoRL robot-learning conference in Austin that week without presenting a demo, Fortune reported.

The latest reorganization follows the original team’s dissolution, the Figure partnership and OpenAI’s broader move toward general models. It places research under Chen while assigning applied robotics and hardware work to Brockman’s organization, concentrating development around a model that can operate without a dedicated robotics architecture.

How much financial pressure surrounds the change?

OpenAI faces cost pressure ahead of a possible initial public offering in early 2027, Fortune reported. The report also put the company’s operating loss at $10.9 billion for the previous quarter, bringing financial constraints into the same frame as the robotics restructuring.

The commercial contest still turns on issues that a simulated benchmark cannot settle. Robot companies retain deployment data unavailable to OpenAI, while slow cloud-hosted inference gives specialists a reason to retain small onboard models. Through 2027, the competitive balance will depend on whether OpenAI can translate the Astra agent’s 74.8% simulated score into reliable physical-robot control at its planned public demonstration.