Makes you wonder how autonomous self driving systems are going to handle cases like these. How far could a self driving car "see" to realize the navigation data it has is incomplete and should instead improvise?

Back when I worked on self driving cars we called it the cliff problem. It's pretty straightforward to spot things that are in your way (trees, cones, broken down car) using LIDAR. It's really hard to notice when a road abruptly ends. From the perspective of the cars' LIDAR a roll over and a sheer cliff look pretty similar. Either way you just won't hit any targets at all with the laser beam. There was a silly obstacle on many self driving car courses that basically had the car drive up and over a relatively steep bump. At the crest of the hill the cliff detection would usually trigger and a lot of cars would just pause there, the roll over looked like a sheer cliff when the car was at a 15 degree up angle.

I assume we can make it work with computer vision for much the same reason humans don't routinely (though sometimes) drive off cliffs.

I'm also waiting for people to exploit these errors and DDoS the systems. Know a particular route will get an autonomous car into a dead end or off a cliff? Order 100 pizzas to the address with stolen credit cards...

An order of 100 pizzas might require human interaction (by policy, not by technological necessity) and depending on how the vehicles are configured, may only result in one vehicle being sent anyway. Order just a few pizzas, and space them apart in time - long enough that they won't be on the same vehicle, but short enough that the problem won't be discovered before the next vehicle is sent.

The system could switch to manual or more cautious autonomous driving when a confidence score is not met. The confidence score could be a score based on how many other cars have driven the road, number of accidents, satellite footage, etc.

If they have medium-range laser rangefinders, then they may be able to sense obstacles like this just at the edge of braking distance. Would work fine at night but maybe not so well in a sandstorm.

Yeah, I have an older car with not that great headlights. I imagine a self-driving car would have been able to "see" the hazard better than I could at night.

This happened to me in the U.S. (Virginia), not long ago. Google maps tried to have me drive straight through a river. It may technically be true that the road crosses the river, but the crossing is impossible for most vehicles unless the river is VERY low. Reported this issue to the Google Maps team months ago, but it's still there... https://www.google.com/maps/place/37%C2%B029'34.7%22N+79%C2%...

This happened to me, in Guatemala. Was using a rental car Garmin GPS whose map had some serious errors. The worst was, driving at night, thanks to my brights being on, I saw the road drop off ahead and slowed us to a stop. It was a sheer cliff into a ravine. The actual road had been knocked out and a replacement road had been built 100 feet away. We had to backtrack to find it.

Ay, what part of Guatemala? That's my home country.

Here's another story: Driving across Guatemala on the Panamerican highway during hurricane Stan[0] (although we didn't know how devastating it was going to be at the time), we saw half of the highway that we had just passed collapse behind us down the side of mountain. We were one of the last 2 or 3 cars to drive west-bound that day. On that drive, we also saw full grown trees sliding down the mountainside, still rooted and standing vertically.

[0] https://en.wikipedia.org/wiki/Hurricane_Stan#Guatemala