Autonomous Driving Corner Cases

[Skotty]

Thought it might be nice to have a thread where people can list out corner cases they run into on a daily basis that would likely cause trouble for an autonomous car or at least not be handled as well as a human driver typically would. This will serve a few purposes.

1) A reminder of how hard a problem this really is.
2) Perhaps as an informational resource for people working on autonomous driving technology who want some ideas on things they haven't thought of yet.
3) As evidence that a human is not as inferior as many take them to be.

I'll start. On my drive home today, I encountered 3 different situations worth mentioning.

1) A work truck pulled onto the highway in front of me dragging an enormous limb underneath of it. Now true I live in redneck country, and this wasn't even a surprising sight. Might be a rare thing in most places, but that's typical of a corner case. Since following the truck would have a high risk of having a close encounter with an enormous limb, the safe thing to do was to either back off or pass the truck.

2) A little more common. Well, a lot more common. I was driving on the highway and there was a car at the side of the road with people near it. Since the passing lane was clear, I pulled into the passing lane to give the people more space. This is also somewhat of a safety thing. It's also a judgement call that might be challenging for a computer to make, because you have to assess how crowded the passing lane is and whether or not pulling into the passing lane is ultimately safer for all parties.

3) Near home, I was on the highway and a service truck was driving slowly on the shoulder, but the shoulder was not wide enough for the truck, so it was hanging into the right most lane. There were 3 lanes. I was in the middle lane, and another car was to the right of me and slightly ahead of me approaching the service truck. The left most lane was not clear, so I slowed just enough to allow space for the car on my right to pull into my lane in order to drive around the service truck.

Now some of these are might partially just be courtesy, but they were all cases where either I was a better driver than an autonomous car, or the autonomous car would have to have some really advanced capability.

 

[JeffK]

Besides very rare corner cases, the network learns from human actions... so yeah. If most humans do what you do, then the car will too. It's pretty straightforward.

Give it enough encounters in enough situations and it'll learn.

 

[3Victoria]

I don't see a benefit in listing a bunch of hypothetical situations, its not going to help.

 

[kurdakov]

on comments here: there are different AI algorithms to solve corner cases.

It's correct that deep neural net will learn from repeated observations, but training neural net might take quite a time, so even with huge fleet it might take years to learn most of corner cases.

Google has devised an approach to learn from fewer observation by storing facts in memory ( so called differentiable neural computers ) but it is not clear if they will share algorithm with others (most probably they will file for patent ).

Now there is another approach in adjacent field - now used for air combat training and maybe used in combat drones, but might be extended to vehicle autopilot http://www.omicsgroup.org/journals/...ted-air-combat-missions-2167-0374-1000144.pdf - here listing corner cases might be beneficial, because genetic fuzzy tree can be algorithmically verified, that it actually solves particular situation, then fuzzy tree works relatively fast and can be used for autopilot to make correct decisions.

So - before neural nets are trained for real autonomous driving from real world behaviour - it might make sense to apply genetic fuzzy trees approach in autopilot and listing corner cases is a must in this case.

 

[Jhall118]

Two difficult things:

Turning left in some busy intersection. Understanding when to "go for it", when to go half way, and when to wait. Knowing when a human driver is going to let you in. People do this very differently in Seattle vs LA.

Second, is those neighborhood roads that aren't one way, but because of cars parking on both sides, can only fit one car going through at a time. Who pulls over and when?

Both these problems are hard because it involves humans interacting with computers.

I am pretty optimistic about Tesla developing these things, but sometimes I look at the state of the current navigation, and realize they still have a lot of work to do

 

[Skotty]

I don't see a benefit in listing a bunch of hypothetical situations, its not going to help.

They are not hypothetical. They are actual occurrences I have experienced in just the last week.

 

[GoTslaGo]

Two difficult things:

Turning left in some busy intersection. Understanding when to "go for it", when to go half way, and when to wait. Knowing when a human driver is going to let you in. People do this very differently in Seattle vs LA.

Second, is those neighborhood roads that aren't one way, but because of cars parking on both sides, can only fit one car going through at a time. Who pulls over and when?

Both these problems are hard because it involves humans interacting with computers.

I am pretty optimistic about Tesla developing these things, but sometimes I look at the state of the current navigation, and realize they still have a lot of work to do

Yes, there's the infamous "Pittsburgh left" as well.

 

[S4WRXTTCS]

Yesterday while driving home from Portland there as a cop aways behind me purposely swerving along all the lanes. I couldn't figure out what in the world he was doing. My best guess is he was testing traction on the road since it was raining quite a bit.

I have no idea how an automated car would have interpreted that had it happened in front of it.

On a more of a daily basis there are two things I can think off on the top of my head that will cause issues with automated driving.

1.) Prey and Go -> Even for the best of us there are some situations where we simply don't have the information we need to pull out into traffic. Where there is an obstruction in the way of our vision. Where we have to move forwards slowly, and we just hope that our front end doesn't get taken off.

2.) The Seattle "You go, no you go" ritual. I have no idea how an automated car is going to figure that out. The only way to win is not to play the game.

 

[ItsNotAboutTheMoney]

Thought it might be nice to have a thread where people can list out corner cases they run into on a daily basis that would likely cause trouble for an autonomous car or at least not be handled as well as a human driver typically would. This will serve a few purposes.

1) A reminder of how hard a problem this really is.
2) Perhaps as an informational resource for people working on autonomous driving technology who want some ideas on things they haven't thought of yet.
3) As evidence that a human is not as inferior as many take them to be.

I'll start. On my drive home today, I encountered 3 different situations worth mentioning.

1) A work truck pulled onto the highway in front of me dragging an enormous limb underneath of it. Now true I live in redneck country, and this wasn't even a surprising sight. Might be a rare thing in most places, but that's typical of a corner case. Since following the truck would have a high risk of having a close encounter with an enormous limb, the safe thing to do was to either back off or pass the truck.

Any autonomous system needs to be able to recognize loose loads.

2) A little more common. Well, a lot more common. I was driving on the highway and there was a car at the side of the road with people near it. Since the passing lane was clear, I pulled into the passing lane to give the people more space. This is also somewhat of a safety thing. It's also a judgement call that might be challenging for a computer to make, because you have to assess how crowded the passing lane is and whether or not pulling into the passing lane is ultimately safer for all parties.

A relatively easy challenge for autonomy, I think. Recognize people + calculate space.

3) Near home, I was on the highway and a service truck was driving slowly on the shoulder, but the shoulder was not wide enough for the truck, so it was hanging into the right most lane. There were 3 lanes. I was in the middle lane, and another car was to the right of me and slightly ahead of me approaching the service truck. The left most lane was not clear, so I slowed just enough to allow space for the car on my right to pull into my lane in order to drive around the service truck.

Another case of space requirements. Autonomous systems would have to have a 360* view of objects, and they'd have to try to figure out what other vehicles are going to do. In that situation if you didn't slow down, the other driver might accelerate and move in front of you.

But anyway, if autonomous vehicles aren't courteous, who's going to know?

 

[WannabeOwner]

there as a cop aways behind me purposely swerving along all the lanes. I couldn't figure out what in the world he was doing

I've seen a solo police car bring a 3-lane road full of traffic to either a stop, or a "follow me at slow speed" (e.g. to allow someone else ahead to use whole road for vehicle recovery before this log-jam arrives) by doing that kind of all-lanes-oscillation.

Dunno if that was what happened in your case though.

 

[JeffK]

I've seen a solo police car bring a 3-lane road full of traffic to either a stop, or a "follow me at slow speed" (e.g. to allow someone else ahead to use whole road for vehicle recovery before this log-jam arrives) by doing that kind of all-lanes-oscillation.

Dunno if that was what happened in your case though.

I've seen this behavior too, but in my case the cop was texting...