OpenAI is restructuring its robotics division by reassigning researchers to core teams under Chief Research Officer Mark Chen and applied teams under Greg Brockman, a shift driven by the belief that general-purpose models like GPT-6 Astra can directly control robots, potentially obviating specialized robotics AI. Evidence cited includes an Nvidia study where an Astra-powered agent achieved a 74.8% success rate on simulated DROID robot tasks versus 51% for the best specialized models. On the new StationeryBench benchmark involving dual-arm YAM robotic arms manipulating desk objects across 200 trials, Astra fully completed 7 of 100 tasks with an average progress score of 46, while Ai2's MolmoAct2 completed none and scored 12. Astra also demonstrated navigation and cleanup on a humanoid robot in a kitchen setting. Cornell and Google DeepMind researcher Yoav Artzi notes Astra achieves near-human accuracy on the unpublished REMAP benchmark, suggesting extensive 3D training data such as Blender scenes may have been used. While general-purpose models show improved spatial understanding and generalization, specialized models may retain advantages in cost, reliability, speed, and energy consumption. This marks OpenAI's second robotics pivot; the division was shuttered in 2020 due to data scarcity, rebuilt in 2024, and now faces pressure to prioritize core AI products amid competition from Anthropic, though CEO Sam Altman reportedly targets a public robot demo by 2027.
Read original
thedecoder