Hacker News Reader

Comment by darkerside | Hacker News Reader

I wonder if this was affected by the fact that the pedestrian was in another lane up until the very last second. Perhaps the car detected the pedestrian but failed to consider it an obstacle since it wasn't in its direct path. It could be unexpectedly difficult to account for pedestrian crossing speed if it caused automated cars to stop when a car in the next lane happened to "wobble" towards the automated car's lane.

View on HN

I dunno, if it decided to ignore the pedestrian because she was in the other lane that's extremely troubling. The pedestrian was moving laterally across the road. If the car has detected that, it should infer that she might become an obstacle very shortly.

Driving is all about predicting the future. Think of every time you've been able to tell that someone is going to change lanes even though their blinker is off, or slowed down when a ball bounces into the street because you know there's going to be a kid following it. If the car isn't capable of that, it's not ready for public streets.

When combined with her walking a bicycle, I suspect this is an edge case they never tested. It may have tripped up their pedestrian detection or path prediction (or both) if they got inconsistent hits on the bike.

There in lies what might be core to the issue though. "An edge case we hadn't tested" Perhaps how we DO AI right now just isn't quite ready for situations this complex yet? Maybe AI based on "oh I've experienced something like this before" is not quite enough? Are we able to do better in getting a system that can infer what's about to happen and make decisions without experiencing it directly... or even indirectly? I know some AI systems are able to do this now with limited cases... but it does make me still wonder if 0.05% of the time the complexity of this task is still just a little outside the capabilities of our learning systems. And maybe we only see the results when something bad happens in that tiny window of time when the AI is unsure.

I am obviously no AI expert, just what I know from loosely following the field. But things like this cross my mind from time to time.

What's most troubling is it doesn't even matter if the uber car thought the slow object in the other lane was a pedestrian, a car, a tree, whatever. Even if it thought the object was say, the most "normal" thing it could be, another car, this would still be a special situation requiring action. Without knowing the objects classification the estimated speed is enough to decide. Why would a car be stopped in the middle of a lane on a fast road? It should be treated as an obstacle that could grow to the side. After all it could be a police, tow, construction, or disabled vehicle, and a cop or tow worker might be about to walk to the side.

Actually many states now require by law that you get as far as possible from a lane with a disabled vehicle, as many human accidents have happened.

I am convinced uber has been basically pretending to do the mountains of careful and sophisticated crap waymo actually has gone to great lengths to do, and is just racing to put anything out so they can keep stringing investors along as far as they can before the jig is up. Well the jig is up now.

I think target fixation also contributes to accidents involving emergency vehicles on the shoulder.

Back to Uber, the number I hear is that they are striving to go better than 13 miles/intervention to prevent an incident. For Waymo this is over 5000 miles. I'm convinced too.

There is not only continuing and stopping, there is also slowing down, as many human drivers do as well in unclear situations.

This is the exact thing I've been thinking all the time: Something's not totally right here: Let's at least slow down to better assess the situation. Everything works better in slo-mo. My human brain would then have much better time to evaluate the scene. This would definitely also hold for the sensors and processing systems for an autonomous vehicle.

And even if it doesn't, and the system still concludes it's noise/an obstacle that isn't going to move etc., a low-speed collision is preferable to a high-speed one. Unless you can be completely confident you aren't going to hit something, slowing down is not a bad default action.

Even if you imagine it thought it was another vehicle stopped in the other lane of traffic, the SDV should have slowed. There's no way you should assume a stopped vehicle isn't hiding another danger and blowing through while speeding is what stupid humans do and what SDVs shouldn't.

In my high school drivers' education class, they specifically taught us to watch for feet that could be stepping into the roadway behind parked cars.

If the car cannot predict "there's something in the next lane, could be in mine the next moment, better slow down just a bit", it has absolutely NO BUSINESS AT ALL driving around on public roads. At each and every occasion that I drove in a city, I have needed to take at least one such minor evasive maneuver due to someone suddenly getting into my lane, be it a car, a bike or a pedestrian. This is not even driving 101; if the car is unable to handle that, it is quite literally unfit for the road.

I have a Forester as well, and it detects things in much wider angles. It even sees cars 160 degrees to the side of me and 30' away when backing up.

Reply on news.ycombinator.com