OpenAI disbands standalone robotics team to integrate research into general models

A shift in internal structure

OpenAI has officially disbanded its standalone robotics team, reassigning its researchers to other parts of the company. The robotics research team will now report to Chief Research Officer Mark Chen within the main research group. Meanwhile, the applied robotics and hardware teams have been moved to the larger Applied organization, which reports to President Greg Brockman.

This reorganization was confirmed by a company spokesperson who stated that robotics remains a core part of the company's long-term vision. The move is framed as a commitment to the field rather than a retreat from it. It marks a significant change for a company that had previously wound down its robotics efforts in 2020 due to limited data.

The transition was noticeable at the CoRL academic robotics conference in Austin this week. Although OpenAI had registered a booth at the event, the space sat empty without any branding or displays. This physical absence highlights the internal shift from a dedicated robotics unit to a more integrated research approach.

The bet on general intelligence

The reorganization reflects a growing industry belief that general-purpose models can power 'embodied AI.' This refers to AI models that control devices acting in the physical world, such as drones, driverless cars, or robots. By using general models, the company aims to avoid building separate, specialized AI models just to serve as the digital brains for different machines.

OpenAI previously assumed that progress in robotics would require building on models like Sora, a video generation tool. They believed Sora's training gave it an implicit understanding of physics. However, the company recently found that its general-purpose frontier models, including GPT-6 Astra, can be applied directly to robotic tasks without that extra step.

This discovery suggests that robotics may no longer be viewed as a 'special problem' by the company. Instead, it is being treated as a downstream application of general intelligence. This is similar to how the same underlying models now power specialized tools for legal work, financial analysis, or software coding.

Performance in simulation

Recent experiments have tested how well a general-purpose model can function without specific training on robotic data. In a study by Nvidia, an agent powered by Astra was given camera images and a control interface to command a robot in a digital simulation. The agent achieved a 74.8% success rate on a manipulation benchmark.

This performance significantly outperformed the best specialized robotic models on the same leaderboard, which reported a 51.0% success rate. Researchers noted that general models can now handle tasks that previously required dedicated training. They can reason about sequences of tasks and adapt when lighting changes or objects are moved to different positions.

Despite these gains, some experts point out the limitations of large language models. Because they are powerful but slow, they often need to be hosted in the cloud, requiring a constant internet connection for the robot. Some believe smaller, specialized models are still necessary to run on a robot's onboard computer for offline tasks.

Cost pressures and future goals

This move comes as OpenAI faces pressure to reduce costs and focus on core products. Last year, CEO Sam Altman instructed the company to abandon 'side quests' to reallocate computing resources. The company has already jettisoned some efforts, including Sora, to focus on more potentially profitable endeavors.

The company is currently navigating heavy operating losses, which reportedly reached $10.9 billion in the most recent quarter. These financial pressures are high as the company seeks new funding and prepares for a potential IPO in early 2027. Streamlining the robotics team into general research helps consolidate these resources.

While the partnership with the startup Figure ended in February 2025, OpenAI is still developing its own humanoid robot. Sam Altman has set a target for a first public demonstration of this robot in 2027. The company maintains that it remains committed to bringing general-purpose robots into the world through its new integrated structure.