Counter take is that it's the perfect environment from a learning perspective: if you succeed there, you can succeed in easier conditions.

Not really, hectic traffic and nonexistent roads aren't a superset of organized traffic and good roads. If you take a human driver from a country with hectic traffic and put them in one with organized traffic they'll struggle and vice versa, the same would hold here. Being skilled in one doesn't really help you out in the other.

They're both a superset of the problem space: safely navigate from A to B.

And broadly speaking, with more varied training data you can achieve better out of sample results.

This sounds like your opinion and doesn’t seem accurate to me at all.

I think the parent is right. Every country has different traffic laws anyway - but driving in a developing country is often fundamentally different.

In Vietnam the effective priority of vehicles is "biggest first". A neural net trained in that environment won't be very useful in the UK, where priority is defined by road markings and street signs.

Absolutely, when they need to test on hard-mode.

I agree. The hard part of autonomous driving is when people and things behave unpredictably and there aren't many context clues.

However, as someone who regularly visits and drives in a 3rd world country, there's a "vibe" to it if you will, and once you learn the vibe it becomes natural. (It seems to be that people respect the laws of physics over the laws of the land, if you will) If you train the cars on a 3rd world country's driving patterns, further refinement will be necessary to fit in better in a more developed country. But early on it will deal with stuff like driving on a road with no lines at the edges or between lanes, one lane bridges, sharing the road with pedestrians, cyclists, motorcyclists, and animals, etc.