OpenAI reassigns robotics researchers due to the performance of general-purpose models like GPT-6 Astra
OpenAI is moving its robotics team into the core research division and the applied robotics division. The company believes the general-purpose GPT-6 Astra model handles robotic tasks better than specialized models.
OpenAI is reassigning robotics researchers to other teams. Some are moving to the core research team led by Chief Research Officer Mark Chen, while applied robotics and hardware teams are joining the Applied organization led by Greg Brockman. According to Fortune, the decision happened so quickly that the company had already reserved a booth at the Conference on Robot Learning in Austin, which then remained empty.
According to OpenAI, general-purpose models such as GPT-6 Astra can control robots directly, reducing the need for specialized robotic models. In a study by Nvidia, an agent powered by the Astra model achieved 74.8 percent success on simulated robotic tasks, compared to 51 percent for the best specialized models. The tests were run on the simulated DROID robot, not on physical hardware. On the new StationeryBench benchmark, which compared the GPT-6 Astra model with Ai2's MolmoAct2 model on five tasks involving objects on a table using dual-arm YAM robotic arms, the Astra model fully completed 7 out of 100 tasks across 200 attempts, while the MolmoAct2 model completed none. The average progress score was 46 for the Astra model and 12 for the MolmoAct2 model. In another test, the Astra model navigated and tidied up a kitchen on a humanoid robot.
According to researcher Yoav Artzi of Cornell University and Google DeepMind, the GPT-6 Astra model represents a significant leap in spatial understanding, achieving near-human accuracy on the as yet unpublished REMAP benchmark. Artzi believes OpenAI may have trained the model on a large amount of 3D data, such as scenes from Blender. It remains unclear how much robotic data was actually used in training, and according to the source, specialized models may have advantages in cost, reliability, speed, and energy efficiency, even though they generalize less well.
OpenAI disbanded its robotics team back in 2020 due to a lack of training data, then rebuilt it in 2024. According to available information, CEO Sam Altman is targeting the first public demonstration of a robot in 2027, and the company has long wanted to manufacture its own robots for everyday use. At the same time, it faces financial pressure to focus on its core AI business, where, according to the source, it has lost significant ground to Anthropic.
Why it matters
If it is confirmed that general-purpose models can replace specialized robotic systems, this changes the economics of robotics development, since companies would no longer need to build and maintain separate specialized models for every task. So far, however, these are results from simulations and limited benchmarks with low absolute success rates, so practical deployability outside the lab remains uncertain.
Relevant practical impact
What this means
For a business
Companies developing or deploying robotic AI solutions should monitor whether general-purpose models such as GPT-6 Astra actually replace specialized solutions before investing in narrowly focused robotic models. For now, these are preliminary benchmarks with low success rates on real-world tasks.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.