Astra Code Policy Latency
I’ll check the Astra code-as-policy implementation specifically, including whether it generates a persistent controller or repeatedly calls the model during execution. Researched code policies Compared runtime behavior and clarified latency tradeoffs Searched 23 websites Yes, that's roughly the idea, but there are two important distinctions: the code doesn't automatically become perfect, and zero model latency doesn't mean zero control latency. Phase 1: Development (slow) Astra writes Python control code Runs experiments on the robot or in simulation Inspects failures and rewrites code Validates and freezes a controller Phase 2: Execution (fast) Camera + robot state Generated Python policy IK / motion control Robot acts; new observations feed back into code This is the architecture tested by SE3 Labs with GPT-6 Astra. Their code-as-policy controller was developed, tested, then frozen and executed without further model calls. SE3 Labs Approach Astra inference during execution Main limi