OpenAI's Astra beats specialized robotics models, Will DePue says

OpenAI researcher Will DePue says on X that Astra "decimated" specialized robotics models in an OpenAI robotics result, which he calls deeply surprising.

He says the result could mean transfer improves with scale, that there is a fundamental algorithmic shift, or something else, and that he does not know which.

In a follow-up post, DePue offers untested gut theories. He doubts Astra used more FLOPs than the specialized models on multimodal data. He says chain-of-thought improves robotics tasks even at low reasoning budgets, and that transfer may rise with scale through optimization and data effects. He says CAD and computer-use training seem unlikely to be the main cause.

Neither post reports benchmarks, metrics, or the models compared.