The reorganization and its rationale
OpenAI has reorganized its robotics team: its robotics researchers are moving to the main research team led by Mark Chen, chief research officer, while the applied robotics and hardware teams are joining Greg Brockman's Applied organization. The Decoder reports this, noting that the news was initially presented as a fresh disbanding of the unit and that OpenAI later clarified that it was a reorganization, with work on robotics continuing.
According to the outlet, the decision was made with such short notice that the company had already booked a booth at the Conference on Robot Learning in Austin, which was then left empty. It is the second time OpenAI has dismantled a standalone robotics team: it had already happened in 2020, when, according to co-founder Wojciech Zaremba, the necessary training data was lacking and large language models appeared to be a faster path to AGI. The team had been rebuilt in 2024.
Generalist models versus specialized models
Behind the choice, The Decoder writes, is the conviction that generalist models such as GPT-6 Astra can directly control robots, reducing the need for dedicated models. In support of this argument, the outlet cites an Nvidia study in which an agent based on Astra achieved a success rate of 74.8 percent on simulated robotic tasks, compared with 51 percent for the best specialized models; the tests were conducted on a simulated DROID robot, not a physical one.
A second reference is StationeryBench, a new benchmark that compared GPT-6 Astra with Ai2's MolmoAct2 on five tasks involving desk objects and dual-arm YAM robotic arms: across 200 trials, Astra fully completed 7 of 100 tasks, while MolmoAct2 completed none, with an average progress score of 46 versus 12. In another test, also reported by the outlet, Astra on a humanoid robot navigated and tidied a kitchen. The Decoder notes, however, that it is unclear how much robotic data was used in training and that other large generalist models, from Anthropic and Google, have also shown recent progress on robotic tasks.
What remains to be verified
The picture is not definitive. According to The Decoder, specialized models could retain advantages in cost, reliability, speed and energy consumption, even without the same generalization capability. Yoav Artzi, a researcher at Cornell University and Google DeepMind, sees in Astra a significant leap in spatial understanding: on the REMAP benchmark, not yet published, the model would achieve accuracy close to human levels, and Artzi hypothesizes training on large amounts of 3D data such as Blender scenes. However, these are reported assessments, not independently verified results.
The timing and industrial objectives also remain to be clarified: according to the outlet, CEO Sam Altman is reportedly aiming for a first public demonstration of a robot in 2027, and the company would like eventually to build everyday robots. In the background is financial pressure to focus on the core AI business, where OpenAI has lost ground to Anthropic. The reorganization, therefore, should be read as a bet on generalist models rather than an abandonment of robotics, but the available data comes largely from simulations and benchmarks that have not yet undergone external review.
