Hacker News Reader

Comment by taneq | Hacker News Reader

This is why efforts based on exhaustive mapping make me nervous. Things change, sometimes rapidly. The vehicle should be using its maps as a general guide, not as some kind of ground truth.

View on HN

The vehicle should be using its maps as a general guide, not as some kind of ground truth.

Not to sound like a jerk, but I would be shocked if there were anyone working on this that was not pursuing things in this way. It just seems kind of obvious that you can't have some big hunk of metal rolling around, following some abstracted track from a map, without "looking" where it is going.

Cars obviously have sensors for real-time data capture. However, the question is how well they handle new signage/lights, road changes, construction, flooded lanes, etc. when those aren't in the database. It's fair to say that autonomous driving today is some point between "we run 100% off maps" and "we can handle things like a human could if we lost access to our maps." (And much closer to the former.)

And yet we have Waymo building 2cm-resolution maps and not operating outside of mapped areas, and (more scarily) Tesla geotagging false positives where Autopilot misidentifies some roadside feature and panic brakes.

If all the cars are always mapping and collecting data will maps ever be out of date?

Yes. Cast your mind to lightly-traveled rural roads. If you are driving one of the many unpaved roads in Arizona during flash flood season, and the last car through was a couple hours ago, will you bet your life that the road is still there?

But hey, that's not a realistic danger in the city, right? Well, look up the sinkhole named "Steve" that opened up in the middle of an Oakland, CA freeway.

Then there's Highway 1, which is known for landslides. Not to mention that little incident with the Bay Bridge back in '89...

Think of all the many times a software developer says "X will never happen."

Then think of what percentage of the time they're right.

By definition, any car that makes an update to a map has experienced an out of date map.

Are those approaches mutually incompatible? It seems to me that it would be easier to build a system which can drive fast when what it's seeing matches what it's expecting to see (most of the time), and which falls back to a much more conservative stance in the case of surprisal (i.e. expectation not matching observation).

Reply on news.ycombinator.com