> why not do RL and make a penalty for bad design?

But this is clearly explained in the article. We can tell whether the design was good or bad only weeks or even months later, which is too long to wait for artificial trajectories. Tests, by contrast, provide validation in seconds or minutes.