OpenAI folds robotics team as GPT-6 Astra controls robots directly

OpenAI's first attempt at robotics ended in 2020 when co-founder Wojciech Zaremba said the company lacked necessary training data and saw large language models as a faster route to AGI. The team was rebuilt in 2024, but this second disbanding comes with a different rationale: OpenAI now appears to believe that general-purpose models like GPT-6 Astra can control robots directly, reducing the need for specialized robotics AI.

The evidence cited includes an Nvidia study where an Astra-powered agent achieved a 74.8 percent success rate on simulated tasks versus 51 percent for the best specialized models, and a new benchmark called StationeryBench where Astra completed 7 out of 100 tasks while Ai2's MolmoAct2 completed none. Astra also navigated and cleaned up a kitchen on a humanoid robot. However, the tests were simulated for the Nvidia study, and it remains unclear how much robotics data OpenAI used in training.

Cornell University and Google DeepMind researcher Yoav Artzi sees Astra as a major leap in spatial understanding, noting near-human accuracy on the still-unpublished REMAP benchmark and suspecting training on large amounts of 3D data such as Blender scenes. Even so, specialized robotics models may retain advantages in cost, reliability, speed, and energy consumption. OpenAI faces financial pressure to focus on its core AI business, where it has lost ground to Anthropic, while CEO Sam Altman reportedly targets a first public robot demo for 2027.