OpenAI is disbanding its standalone robotics team and moving its members into broader research, applied robotics and hardware organizations, according to reporting summarized by The Decoder. The company presents the reorganization as an acceleration of robotics rather than a retreat from the field.
The strategy reflects a belief that general-purpose models such as GPT-6 Astra may control robots without a separate model family. An Nvidia simulation study reported a 74.8% task success rate for an Astra-powered agent, compared with 51% for the strongest specialized model tested. That result came from a simulated DROID robot, not physical hardware.
A separate StationeryBench test was more sobering. On dual-arm robots manipulating desk objects, Astra fully completed seven of 100 tasks, while Ai2’s MolmoAct2 completed none. Astra made more partial progress, but the low completion rate illustrates the distance between benchmark gains and dependable real-world operation. Specialized systems may retain advantages in speed, energy use, reliability and cost.
OpenAI closed an earlier robotics division in 2020, rebuilt the team in 2024 and reportedly aims for a public robot demonstration in 2027. Moving researchers closer to the main model and applied teams may reduce organizational boundaries, but it does not remove the need for physical testing and robot-specific data.