OpenAI reorganizes its robotics team and moves researchers into other units
According to The Decoder, OpenAI has restructured its robotics group and placed its researchers in other departments. Robotics scientists now belong to the core research division run by Mark Chen, who holds the title of chief research officer, while the hardware and applied robotics staff report into Greg Brockman's Applied organization. The Decoder first described the team as disbanded again, but it changed that wording after OpenAI said the aim was closer collaboration with research and product and that robotics efforts continue.
Fortune reported that the move came with little warning: OpenAI had reserved a stand at the Conference on Robot Learning in Austin, and the stand remained empty afterward. The Decoder's reading is that OpenAI believes broad models like GPT-6 Astra can drive robots without relying on specialized robotics models, which would make dedicated robotics models less necessary.
A study by Nvidia reported that an agent built on GPT-6 Astra succeeded in 74.8 percent of simulated robotics tasks, while the best specialized models managed 51 percent. All of those trials ran on a virtual DROID robot rather than real hardware.
A new benchmark called StationeryBench compared Astra with Ai2's MolmoAct2 on five tasks involving desk objects, run on dual-arm YAM robotic arms, according to the report. Across 200 trials, Astra fully completed 7 of 100 tasks, while MolmoAct2 completed none. Astra's average progress score was 46, compared with 12 for MolmoAct2. In a separate test, Astra running on a humanoid robot navigated a kitchen and cleaned it up.
The source says it is still unknown how much robotics material OpenAI and other labs used in training. Anthropic and Google have also shown recent gains on robot tasks with their large general models. Purpose-built robotics systems may still be cheaper, more dependable, quicker and less power-hungry, even if they adapt less broadly.
Cornell University and Google DeepMind AI researcher Yoav Artzi said he regards Astra's spatial reasoning as a large step forward. Artzi said Astra reaches near-human accuracy on the REMAP benchmark, which has not yet been published, and he thinks OpenAI could have used large volumes of three-dimensional training material, for example Blender scenes.
https://twitter.com/i/status/2098496429752713618
Co-founder Wojciech Zaremba, according to the report, explained the 2020 closure of the robotics division by pointing to a shortage of training data and to large language models being a faster route to artificial general intelligence (AGI). OpenAI assembled a new robotics group in 2024. CEO Sam Altman is reportedly aiming for a first public robot demo in 2027, and the longer-term goal is in-house robots for everyday use.
The company also faces heavy financial pressure to concentrate on its core AI business, where Anthropic has pulled ahead.