DEV Community

News
News

Posted on

OpenAI folds its robotics researchers into research and applied teams as its general models start running robots

Infographic cover

๐Ÿ“… November 14, 2026 โ€ข Fortune

OpenAI confirmed to Fortune that its stand-alone robotics team is being folded into the main organization: robotics researchers move to the research org under CRO Mark Chen, while applied robotics and hardware join the Applied organization under president Greg Brockman. Roles and the roadmap are preserved โ€” reassignments inside the company, not departures โ€” and the reported goal stands: a first public robot demo targeted for 2027 (per Fortune, November 13, 2026). The company frames the move as a bet that general models โ€” GPT-6 Astra โ€” now apply to robotics directly, not a step away from them.

Why the bet โ€” earlier this year the robotics plan was built around Sora, the argument now is that general models caught up. In NVIDIA research, an agent on GPT-6 Astra reached 74.8% success on simulated robotics tasks versus 51% for specialized models โ€” simulated DROID tasks, not a physical robot (per the study quoted by the-decoder). On StationeryBench from GitHub robocurve, Astra faced Ai2 MolmoAct2 on 5 tabletop-object tasks on a dual-arm YAM across 200 trials: Astra fully completed 7 of 100 tasks, MolmoAct2 0 of 100, average progress 46 vs 12 (per the-decoder, November 14, 2026).

The balance matters: specialized robotics models keep genuine advantages in cost, reliability, speed and energy consumption (per the papers quoted by the-decoder), so folding teams into the general stack is a bet, not a settled outcome. The race is wider than one company โ€” Anthropic advances Claude robotics research and Google pushes Gemini 4 Argon robots. A Fortune detail shows how fast the call was made: at CoRL in Austin OpenAI still held a booked but empty booth #208, and CMU professor Guanya Shi puts it plainly โ€” robotics is no longer a special problem.

OpenAI has restructured here before: the first robotics team closed in 2020 โ€” not enough training data, and LLMs looked like the faster path to AGI per co-founder Wojciech Zaremba โ€” and the effort returned in 2024. The difference now is the evidence base: general models have measurable robotics results to point at, from the NVIDIA study to StationeryBench, which makes folding the team a bet on acceleration rather than a retreat from the field.

Key figures

IMPACT: for automation teams the signal is directional โ€” embodied work is being pulled into the same general-model stack you already build on, one model family for text, vision and robot control, with the reported 2027 demo as the checkpoint. Treat simulator numbers with care: success on simulated DROID tasks does not equal a robot on your factory floor, and the cost, speed and energy advantages of specialized models remain the counterargument to track. If you plan robotics integrations, watch whether OpenAI ships developer-facing interfaces as the research org absorbs the team.

Sources: fortune.com โ€” the-decoder.com โ€” tech.yahoo.com โ€” seekingalpha.com

IMPACT checklist


Sources:


Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky

AI Automation

Top comments (0)