斯坦福用 GPT-6 Astra 当大脑，让机器人自主整理厨房
斯坦福与 Caltech 用 Unitree G1 加 GPT-6 Astra，在没见过的厨房里自主建图、规划动作并完成收拾与取物任务。
Stanford gave a humanoid robot OpenAI's smartest model as a brain and put it in a kitchen it had never seen. It worked out the room on its own. The robot is a Unitree G1, and the team from Stanford and Caltech used it straight out of the box, with no training for this kitchen and no new moves. It walked around, built a 3D copy of the room from what it saw, and remembered where everything was. GPT-6 Astra handled the thinking. The robot knows five moves: walk somewhere, pick something up, put it down, open a drawer, and take something out of a drawer. Astra picks the next move and the order. With that, it tidied a kitchen across the room and fetched an object from a vague request, because it remembered where it had seen it. Home robots have usually needed training for each new space. This one ran on a general AI model and a short list of moves. It pauses every time Astra thinks. Building the 3D copy of the room takes time and costs money in API calls. On long tasks, its fingers overheat. Its brain is also a paid cloud service. Every decision it makes in that kitchen goes out to OpenAI's servers and comes back. Would you let a robot in your home if its brain ran on someone else's computer?