fortune.com web signal

OpenAI folds robotics team into main research and applied orgs

5 sources tracking this story
OpenAI Robotics Agents ai-business

TL;DR

  • GPT-6 Astra scored 74.8% on simulated robotics tasks versus 51% for specialized models, the benchmark OpenAI used to justify the team elimination.
  • On physical hardware, Astra completed only 7 of 100 bimanual desk tasks versus zero for specialized rival MolmoAct2, leaving commercial viability unproven.
  • Each Astra robotics attempt costs $36.40 at 18.56-second latency, making real-time physical robot control commercially unviable at current economics.

OpenAI has dissolved its stand-alone robotics team for the second time in six years. The research staff now reports to Chief Research Officer Mark Chen, and the applied robotics and hardware staff report to President Greg Brockman, Fortune reported.

The first robotics effort was wound down in 2020, with cofounder Wojciech Zaremba citing "limited robot data" at the time. OpenAI rebuilt a stand-alone team in 2024 and partnered with humanoid startup Figure. That partnership ended in February 2025.

The reason for the reorganization this time is different. Researchers believe general models like GPT-6 Astra can control robots directly, without a dedicated stack. Astra reportedly achieved "74.8% success rate on a simulated manipulation benchmark" against 51.0% for specialized robotics models.

A company spokesperson said robotics "remains a core part of OpenAI's long-term vision" and framed the move as acceleration rather than retreat. CEO Sam Altman is targeting a first public demonstration of the company's own humanoid robot in 2027.

Outside the company, the read is mixed. Guanya Shi, of Carnegie Mellon University's Robotics Institute, told Fortune the move suggests OpenAI sees robotics as "no longer a special problem." MIT's Russ Tedrake, who runs Walden Robotics, argued robotics startups still hold the advantage of real-world deployment experience. Generalist cofounder Pete Florence said it is "unclear that any existing model would dominate robotics."

One detail from the week, among a dense run of robotics stories on our tracker: at the CoRL robotics conference in Austin, OpenAI's booth sat empty despite being registered.

What others are reporting

Coverage cluster as of 24h after publish

  1. The Decoder Read →

    Adds CoRL empty booth corroboration, Astra vs. VLA benchmark breakdown, and frames the pivot against OpenAI's financial pressure from losing ground to Anthropic.

    General-purpose models like GPT-6 Astra can control robots directly, reducing the need for specialized robotics models.
  2. Business Upturn Read →

    Frames the move as a recurring pattern, citing the 2020 shutdown and 2024 rebuild, with the CoRL empty booth as field corroboration of an abrupt decision.

    The confirmed news is a team reshuffle, not an exit from robotics.
  3. Superpower Daily Read →

    Focuses on the simulation-to-physical gap: Astra's 74.8% sim score versus 7 of 100 completed real tasks, arguing specialized on-device models stay necessary until cost and latency improve.

    general-purpose models have shown they can be applied directly to robots.
  4. DEV Community Read →

    Skeptical take: cites CMU researchers disputing simulation benchmarks as proof of embodied AI, and flags this as OpenAI's second robotics disbandment in six years.

    A general model beat a dedicated model on one simulated test, at a price and latency no real robot can afford yet.

Shared on Bluesky by 1 AI expert