I think the training limits other applications like self driving cars a lot too. Training is frequently done in contrived scenarios. Now, imagine if GM or Ford put sensors and telemetry for future self driving car use on their models decades ago. You'd have practically every road mapped with ten thousand observations a piece by now. It wasn't for lack of forsight to not collect these data, it was really for lack of profit. Sensors and other data collecting systems are not new technology. Neither is machine learning, its also decades old. The only thing different these days is that the compute required to run these models on these data has gotten cheaper, but you could have been collecting data the whole time with anticipation of cheap compute to run these models in the future. Then you'd have a leg up on everyone in the business with your massive dataset of decades of real world, not test tracks and other such situations utilizing an employed driver in all likelihood.