Hacker News Reader

Comment by nmjohn | Hacker News Reader

Your artwork is great! I'm not sure #1 would really be a problem as that situation never should occur and is totally preventable (there always should be traffic cones/barricades surrounding the cover, something the car easily could detect). In the rare case it does happen, the utility company would/should be at fault for leaving a manhole open like that.

I can also think of a few ways to prevent number 2 (basically, a combination of GPS + knowing where all intersections occur + road data from thousands of other connected cars = knowledge of where every stop sign is. Certainly not foolproof, but I think ultimately it is a problem that has possible solutions)

The rest of your examples are fantastic though, #3 being quite terrifying actually. They are all very intriguing thought-experiments and I look forward to seeing your future additions!

View on HN

> In the rare case it does happen, the utility company would/should be at fault for leaving a manhole open like that.

They're who's at fault, but you're the one who's dead.

I'd rather my car avoids holes in the road even if they haven't been marked [yet].

> They're who's at fault, but you're the one who's dead.

This is what my dad taught me too and why I've since always been cautious both as a driver, bicyclist and pedestrian even when I'm in the right to drive/go. Being right isn't as fun when someone's on their way to the hospital/morgue.

An open manhole will not cause you to veer out of control and die in a horrific accident. At worst, you might have a blowout. But in general, unless you have a very small car, most cars tires are larger in diameter than the diameter of the pothole.

Provided that your tires are normal (and not the thin-sidewall don't-my-ride-look-cool tires with big rims), and in relatively normal balance - and you are going a normal speed - it most likely will drop slightly then bounce off the lip, jarring car forcefully, but likely not doing much damage.

If the tire doesn't blow (it may), then you might get a cracked rim (if alloy) or bent rim (if steel and depending on the force and deformation), and you would probably want to have the tire inspected (because the plys may be compromised in the area from the sharp impact, which could cause premature interior delamination of the tire in the future).

Yes - ideally you want you or your car to avoid an open manhole or large pothole, but in general, its unlikely to be the cause of a serious accident.

LIDAR imaging already scans the road for it's topography. An open manhole would be easily detected (its equivalent to even a particularly deep pothole - un-navigable).

I don't know how Tesla's video only solution would cope, but they're not Level 5 yet.

How would it show up different than a regular pot hole? What if its almost filled with water and the only thing giving away its true nature is its roundness? The point is to explore the limits. If these aren't the limits, then they are likely very close. I'd like people to think of those cases before they jump into their car and start playing cards (like i one of the videos in this thread).

Basic geometry. A pothole of greater then a certain depth, at a known angle from sensor will produce a different distance profile (namely - rather then an expected arc, you'd see a sudden increase in distance, corresponding to a non-planar surface). Depth is inferred from the deviation from the plane of the road-surface around it.

This is not actually a particularly interesting edge case.

The water-case is more interesting, but I would counter by posing the question of how a human would know whether a water-filled pot hole is safe to navigate? A clear water pothole would not impinge the LIDAR, a murky water one might and could fool LIDAR I suppose (I can not find any solid info on this interface).

Of course car RADAR and a visual camera are likely to also be fitted - both of which could identify a water-filled pothole which LIDAR might struggle with (RADAR can penetrate water and find the unknown backscatter, a camera can simply see the murky-water pothole and choose to navigate around it or stop - in both cases a human can't make a better informed decision).

It depends on the context. Humans may make a better decision if there is something else providing the information. Perhaps the manhole cover sits at the curb triggering immediate attention. Perhaps there are cones, only in the wrong place because some drunk guy moved them for fun. Such is reality.

The problem is not so much availability of information (a sensor packed car has a lot more information than a human behind a windshield), but making sense of that information, particularly in those 0.0001% uncommon, strange cases.

Humans can connect evidence (a manhole cover sitting next to a water filled hole) because they generally know how things work (aka common sense knowledge, long standing problem in AI). Humans can infer what happened and predict what could happen. These things are not really available to AI which is these days more like see-react, not see-anticipate-react. I have a specific post on this on my blog:http://blog.piekniewski.info/2016/11/03/reactive-vs-predicti... The blog has many other posts on limits of todays approach and on some ideas to fix it in the future.

I agree that some of these cases may be challenging to some humans (particularly inattentive), but we want the self driving car to be (much) safer than inattentive driver, so we need to set the bar high.

Clearly if you consider any of these scenarios in isolation we can dream up a solution, but can you do that with literally any adverse scenario? That seems like the point to me.

Yes, that is exactly my point. Solving these particular situations is like typing all the possible Winograd Schema sentences into a chat bot. Certainly possible but does not solve a larger problem.

In reality there will always be a different situation. These situations sit in the statistical long tail (are too infrequent to reliably train, and too frequent to ignore) and vary enormously from case to case. These are just examples.

Except if the first scenario is solvable (fairly trivially) then why should I assume there must exist an unsolvable scenario?

An argument from the vaguery of "humans have context" with regards to driving is a poor one - humans that are driving do not do a good job of analysing context because of reaction times, and do not share a common context they react to similarly.

You haven't made a compelling case, because you've yet to present a compelling example (i.e. one definitely unsolvable by reasonably usable technology on an autonomous platform). The manhole scenario is based on assumption about the operation of LIDAR which simply aren't true and you've had to modify it in a number of ways to try and make it tricky (i.e. when has there ever been a manhole completely full of water?)

I think you're still confused. The point isn't that any of these scenarios don't have solutions. It's that there is a practically infinite number of scenarios like them and it's impossible to plan for each and every one of them.

A sinkhole (named "Steve"?) just opened up in a highway near Oakland, CA. Initial photos showed a hole a couple feet across. Manhole-sized. Repair crews later found that under the asphalt the sinkhole was large enough to swallow a car. Human drivers recognized that something was amiss and avoided the hole. Would a self-driving car?

Let's put "ability to detect solid road in front of car" near the top of the requirements list. Whether it is a pothole, sinkhole, missing manhole cover, or missing bridge, it needs to be a mandatory requirement. From what I've seen of LIDAR imaging it seems like it is possible.

> In the rare case it does happen, the utility company would/should be at fault for leaving a manhole open like that.

Sometimes, the utility wouldn't be at fault. It has happened in the past the people would steal manhole covers to sell them for scrap (you know, to get money for their next 'fix').

This is less of an issue now, as most scrap companies won't take manhole covers anymore, unless the seller can prove they represent the locality shown on the cover, and that they have such authorization...

...but there are less scrupulous scrap companies.

Hey, I just finished another one. A school bus with a swinging open stop sign in the middle of the freeway.

Anyway, just a food for though, autonomy really requires a lot of "intelligence" and our technology is not quite there to deal with all these bizarre corner cases. I'm glad you like it.

I think such things can be discounted by knowing that motorways never have stop signs (same with the prank stop signs). Of course, there are a lot of things that only cause a change in driving behaviour in very specific contexts. The bus' stop sign probably only applies when the bus is stopped. Similarly, in Germany at least a bus may turn on hazard lights on a bus stop, requiring everyone to drive past only at walking speed. A bus stopped on the shoulder of a motorway with hazard lights on is a different context again and thus should not trigger the same behaviour.

There are a lot of rules and laws and they change from country to country, or in the US' case, even from state to state. Self-driving cars must know these things and react accordingly. So I think the scenarios presented here are just a few (admittedly, more far-fetched than others) more contexts amidst the probably hundreds of others that already have to work correctly for switching safely between city and motorway driving, driving in a living street, observing right of way correctly in all circumstances (roundabouts, weird stuff like four-way stops, signs changing ROW for one intersection, or a stretch of road, lowered kerbs, people exiting a living street even though it's to the right, cars on an on-ramp and perhaps letting them in based on how far the on-ramp still continues, ...).

Stop signs are interesting in any case, since they have a characteristic shape. If we go full autonomous, then snow-covered signs must be correctly observed as well, at which point any octagon shape may be a stop sign (perhaps, again, depending on context). Same with signs that don't reflect well anymore at night.

The real sign SDVs are here will be when infrastructure starts accommodating their needs. Humans aren't really good at driving either, so we've invented a lot of ways to help them and direct their attention. Open manholes are supposed to be marked clearly, because people do miss that stuff. When there's snow on the road hiding lane markings, someone will come and clean it out. Signs are made to be retroreflective. Etc.

So at some point, I suppose the infrastructure (broadly understood - including laws) may be modified to reduce the dependence on cultural context and other things machines are weak at. So for instance, it won't be every sorta-octagonal shape that works as stop sign, it will be required by law to be clearly visible and also have some machine-friendly accommodations, and SDVs will be free to ignore signs without those accommodations.

(Doesn't solve the prank problem, but humans are equally vulnerable to a targeted prank anyway.)

> (Doesn't solve the prank problem, but humans are equally vulnerable to a targeted prank anyway.)

I would hesitate using the word "equally". People are actually quite robust. Particularly the second human in row will certainly not be tricked by the same prank that tricked the first one.

I want to emphasize the word "targeted" I used. Pranks involve an intelligent agent with malicious intent and an attacker's advantage - i.e. prankster is free to exploit any vulnerability of its victim. People have different vulnerabities than machines, but they still have them.

Sure they do, agree with that. But word "equally" suggest the susceptibility is the same. I would actually emphasise the difference. It is much easier to fool a machine than a human, particularly if we have a copy of the machine at hand and tinker with it (see adversarial examples for deep nets). Humans are all different, so we can never expect our "adversarial example" to be 100% certain to work.

Especially if it's a familiar road.

> When there's snow on the road hiding lane markings, someone will come and clean it out.

In Nordic countries, you don't see the lane markings for several months, they are under snow and ice. And sometimes, when it melts in April or May, you notice you have been driving on the roadside for a few months :-)

Oh, and you don't see anything at all if you are behind a truck or a bus.

In Nordic countries...

In some parts of some Nordic countries. In my part of my Nordic country we haven't had proper snow for probably about 6 or 7 years and even then it was only for a few weeks.

But more generally. If the snow clearing machines where also driver less then they might be able to run more or them more often and keep the roads clearer.

It's not just a matter of running them often. Where are you going to put all this excess snow? And operating the plows won't be free even without labor.

> Humans aren't really good at driving either,

I'm not sure where did that come from. One fatality per 200 million miles not good? Seriously, there are millions of people travelling every day. And true, there are accidents, but I think this "Humans aren't really good at driving" mantra is not really serious (but frequently repeated recently by PR of some companies).

Attentive humans are extremely good at driving, distracted humans are much worse, but perhaps the technology should focus first on the much easier task of making sure the driver attends. That would probably safe many lives, before we can have a real autonomous car.

>When there's snow on the road hiding lane markings, someone will come and clean it out.

In New England, snow can completely cover the road surface for days or even weeks at a time, and ever-changing piles of snow cover the curbs and parts of the lanes. Humans just choose a path without regard to where the lanes are in the summer. On some roads this turns a four lane road into a two lane road with a lane-width snow pile between the lanes. In a few spots it turns a two lane road into a one lane, with drivers from different directions taking turns.

Remember when Boston was about to start dumping snow into the sea because they ran out of places to put it?

Yeah, there were a few spots in my neighborhood that they gave up on trying to plow, but those of us with 4WD trucks were able to get through. Not sure how a self driving truck is going to know which snowbanks it can drive through/over and which it can't. Sometimes I couldn't tell until I tried. That was fun!

Now I really wish I had taken more photos that year specifically to illustrate this sort of thing.

But if it's that reliant on a special sensor it's quite susceptible to not only pranks but sabotage or plain old maintenance issues.

I know a few places where the stop sign is > 90% covered by bushes in summertime, and the continuous transversal white line > 90% covered by gravel (and the remaining 10% have worn off). Yet people do stop. Either because they know the place, or because they "feel" there is something suspicious (the comparative size of roads, the angle of the crossroad, the fact that there is a "pile" of gravel there, etc.). Not sure how a car could decide this by itself.

I wouldn't say never; sometimes you see them in really odd places.

You could discount that by noticing that stop sign generally shouldn't move relative to the ground. But then, you may want to notice when a police car wants to pull you over by waving that thing they wave out of the car's window...

I actually love trying to make computers deal with the real world - it quickly reveals just how goddamn complicated the real world is, and how many things we think as hard and fast are utterly arbitrary.

Right, there are cases where such a moving stop sign would actually be a real stop sign. The problem is, we try to program in the complexity that cannot be anticipated.

That's something what's easy to train. Each of these are a bit country specific like how a road sign is mounted and where, but this is far from a real problem.

What is the real problem then?

Reply on news.ycombinator.com