<rant>Self driving cars already train on simulated roads because that allows the creation of any scenario. Even human pilots use simulators to train, especially for those rare situations. And since robot training is expensive and slow in the real world, the only alternative is to do it in a sim.
Simulation allows the composition of any scene that might be very hard to record in the wild - for example, an octopus sitting as a hat on the head of an elephant... where would you be able to get that photo? but surely you imagined it in 0.1 seconds, using your imagination - a powerful simulator humans have in their heads. Such images are crucial in training AI.
I think AI will reach human level when it will be equipped with a sandbox where it can try out its ideas and concepts, similarly to how scientists use labs to test their theories. When AI gets its world simulator, it will be able to learn reasoning and meaning that is grounded in verification. Just like us. We have the world itself as our fundamental "simulator" and experiment on it to learn physics, biology and AI. AI needs a simulator too.</>
Also why commercial game developers usually don't whine about 3D APIs like FOSS indie devs do.
Adding support for yet another graphics backend is a trivial task, compared to the overall feature set of a game engine.