Push cumulative high-quality, generalizable data up the x-axis. The red curve — how much fresh human data each new task costs — falls; the teal curve — how much the loop improves itself from its own deployment experience — rises. Where they cross, the flywheel produces more usable data than each new task consumes: the loop goes self-sustaining. I call that crossing the data escape velocity (机器人学习的第一宇宙速度): past it, marginal human cost slides toward zero-shot, and somewhere just beyond, success rates clear the bar where deployment pays for itself — the commercial-viability band. Generalist's GEN-0→GEN-1 jump (64%→99%) and π's data flywheel are early sightings of this curve.