OpenAI folds its robotics team: the general model is the robot brain now
OpenAI's booth at CoRL in Austin, number 208, sat empty this week. On Friday Fortune explained why: OpenAI has disbanded its stand-alone robotics team. https://fortune.com/2026/11/13/openai-disbands-robotics-team-as-its-general-models-start-running-robots/ Robotics researchers move into main research under Chief Research Officer Mark Chen. Applied robotics and hardware move into Applied under President Greg Brockman. A spokesperson called it a commitment, not a retreat, and said the company still targets a first public demo of its own humanoid in 2027.
The reason is the interesting part. According to Fortune, researchers have seen surprising performance from simply connecting GPT-6 Astra to robots. At the start of the year OpenAI thought building on Sora was the path. Now the bet is that a general model drives the robot directly. CMU's Guanya Shi summed up the shift: robotics is no longer a special problem, just a downstream application of general intelligence. This is the third version of OpenAI robotics: wound down in 2020, rebuilt in 2024 alongside Figure, Figure deal ended in February 2025, and now absorbed.
A paper on this week's Hugging Face board shows what that looks like in practice. JustRUA (https://arxiv.org/abs/2611.16009, code https://github.com/ZefanW/JustRUA) puts GPT-6 Astra in charge of a robot arm through a fixed harness and then takes the harness apart piece by piece. A lean setup that keeps the full episode history and stops each motion early goes a long way. Scene geometry helps most, even when depth is predicted from plain RGB. The RoboDojo sim score goes from 29.7 to 50.3 with no learned policy, no self-improvement loop and no task-specific tuning.
Put the two together and the org chart follows the harness. If a better wrapper around a frontier model gets you most of the gain, a separate robotics team building separate models looks like overhead. It is the same story coding went through: one general model, many harnesses, and the specialist work moves into the plumbing. The risk is obvious too. Sim scores are not a humanoid walking into a kitchen, and 2027 is a promise, not a demo.
← Back to all articles
The reason is the interesting part. According to Fortune, researchers have seen surprising performance from simply connecting GPT-6 Astra to robots. At the start of the year OpenAI thought building on Sora was the path. Now the bet is that a general model drives the robot directly. CMU's Guanya Shi summed up the shift: robotics is no longer a special problem, just a downstream application of general intelligence. This is the third version of OpenAI robotics: wound down in 2020, rebuilt in 2024 alongside Figure, Figure deal ended in February 2025, and now absorbed.
A paper on this week's Hugging Face board shows what that looks like in practice. JustRUA (https://arxiv.org/abs/2611.16009, code https://github.com/ZefanW/JustRUA) puts GPT-6 Astra in charge of a robot arm through a fixed harness and then takes the harness apart piece by piece. A lean setup that keeps the full episode history and stops each motion early goes a long way. Scene geometry helps most, even when depth is predicted from plain RGB. The RoboDojo sim score goes from 29.7 to 50.3 with no learned policy, no self-improvement loop and no task-specific tuning.
Put the two together and the org chart follows the harness. If a better wrapper around a frontier model gets you most of the gain, a separate robotics team building separate models looks like overhead. It is the same story coding went through: one general model, many harnesses, and the specialist work moves into the plumbing. The risk is obvious too. Sim scores are not a humanoid walking into a kitchen, and 2027 is a promise, not a demo.
Comments