I agree, but FWIW in the linked work they did randomize some of the engine parameters during training to avoid fitting too much to a specific set of assumptions. Certainly though more accurate simulations, as long as they were still fast to compute, would be very useful!
yeah, I think a big reason why they needed to do randomization in the first place is because of inaccuracies in the simulation compared to the real world. And the extra randomization required almost two orders of magnitude more training time!
