What changed
OpenAI has reorganized its robotics team, moving researchers to other departments such as the main research team under Chief Research Officer Mark Chen and the Applied organization led by Greg Brockman. This reassignment is part of a broader strategy to integrate robotics work with general-purpose models, which the company believes could control robots directly, reducing the need for specialized robotics AI.
What this means in practice
The decision was made with short notice, leading to an empty booth at the Conference on Robot Learning (CoRL) in Austin. OpenAI claims that general-purpose models like GPT-6 Astra can perform robotics tasks effectively, as demonstrated in a simulated environment where an Astra-powered agent achieved a 74.8 percent success rate on simulated robotics tasks, compared to 51 percent for the best specialized models. A new benchmark called StationeryBench showed Astra completed 7 out of 100 tasks, while another model did not complete any. Astra also demonstrated navigation and cleanup capabilities on a humanoid robot.
Limitations and availability
The supplied announcement does not specify how much robotics data OpenAI incorporated during training. Specialized robotics models may still offer advantages in cost, reliability, speed, and energy consumption. While some researchers see Astra as a major leap in spatial understanding, the exact training data and methods remain unclear. OpenAI's long-term goal includes building everyday robots, but the company faces financial pressures to focus on its core AI business.
