From my limited understanding of the technology involved I would guess there's some risk of overfitting.
That is, it requires significant work to train a car to drive period, whether you train it in a relatively "orderly" environment (as we would think of it) or not. However, the training to drive in a chaotic city center should generalize to handle orderly environments, whereas the inverse may not be true.
The whole idea of driverless cars is not to be people movers on tracks, but to actually provide the versatility of a human driver in a car, so getting one that can do "everything a human driver can do" seems like the bar everyone needs to hit for it to be commercially viable.
That is, it requires significant work to train a car to drive period, whether you train it in a relatively "orderly" environment (as we would think of it) or not. However, the training to drive in a chaotic city center should generalize to handle orderly environments, whereas the inverse may not be true.
The whole idea of driverless cars is not to be people movers on tracks, but to actually provide the versatility of a human driver in a car, so getting one that can do "everything a human driver can do" seems like the bar everyone needs to hit for it to be commercially viable.