用 GPT-6 Astra 做 Agentic Robotics：一小时建出机器人仿真
伯克利与 NVIDIA 用 Astra 演示 Graph-as-Policy：从一段擦金属杆视频生成物理仿真，再写技能图、导出代码上真机，失败视频还能反哺改进。
Agentic Robotics is a new paradigm in robot autonomy. For the past few years, robotics has been split between two camps. Classical autonomy breaks each task into hand-built modules for perception, grasping, and control. It’s fast and reliable, but every new task needs a team of engineers to build and tune it. End-to-end VLAs learn whole tasks from demonstrations. They promise generalization, but they need tons of robot data and haven't been widely deployed in production. Agentic Robotics hands the engineering to frontier coding agents. For a new task, the agent writes a modular robot program, tests it in simulation, diagnoses what broke, and fixes it, all offline. The output is fast, readable code for the robot to execute (the frontier model never runs on the robot). This paradigm centers on a harness called Graph-as-Policy. Instead of asking one model to write all the code, it breaks the task into a graph of small skills: find the part, grasp it, move it, check the result. Separate agents each own a node, test it in simulation, and revise it until the whole graph works. A VLA might even be a node inside the graph! @Ken_Goldberg's lab at UC Berkeley and @nvidia recently showed this with GPT-6 Astra: -> Build the sim: From a short video of a robot wiping a metal bar with a sponge, Astra built a physics simulation that reproduced it in under an hour. -> Write the policy: Using Graph-as-Policy, Astra built the skill graph, then improved its success rate and speed in simulation. -> Deploy: The exported code ran on the real robot and worked. When a trial fails, the agent learns from the failure video. With Agentic Robotics: -> No demonstration data is required -> Engineers can still read, test, and fix the code (not a black box) -> Runs at industrial speeds -> Scaling gains in frontier coding models flow straight into robotics It’s already in production. Goldberg's company, Ambi Robotics, used the approach in production for their package sorting robots. This approach is still early. The big question is whether the Physics models can get good enough to handle contact-rich tasks, like inserting cables into a GPU rack. And broadly, it's unclear how this could scale + generalize. But the implication is exciting. If frontier models can do the heavy lifting of writing robot code, they greatly reduce the bottlenecks in data collection and task-specific engineering. Robotics is now just another domain that superintelligence will RL on!