Meeting on a narrow road

Topic: 

Jean-Louis Gassée, while a respected computer entrepreneur, wrote a critical post on robocars recently which matches a very common pattern of critical articles:

The pattern is as follows:

  • The author has been hearing about robocars for a while, and is interested
  • While out driving, or sometimes just while thinking, they encounter a situation which seems challenging
  • They can't figure out what a robocar would do in that situation
  • They conclude that thus the technology is very far in the future.

His scenario is the very narrow road, so narrow that it really should be one-way but it isn't. In most of the road, two cars can't pass one another. Humans resolve this through various human dynamics, discussion and experience.

In most of these examples, the situation is not one that is new to robocar developers. They've been thinking about all the problems they might encounter in driving for over a decade in many cases. It's extremely rare for a newcomer to come up with a scenario they have not thought of. In addition, developers are putting cars on the road, with over a million miles in Google's case, to find the situations that they didn't think of just by thinking and driving themselves. It is not impossible for novices to come up with something new -- in fact a fresh eye can often be very valuable -- but the fresh eyes should check to see what prior thinking may exist.

Some of the problems are indeed hard, and developers have put them later on the roadmap. They will not release their cars to operate on roads where the unsolved situations may occur. If snow is hard, the first cars will be released in places where it does not snow, or they will not drive on their own if it's snowing. In the meantime, the problems will be solved, in a priority order based on how often they happen and how important they are.

The "two cars meet" situation involves very rare roads in the USA, so it's not a high priority problem there, but it would not be a surprise problem. That's because current plans have cars only drive with a map of the road they are driving. No map, they don't drive the road.

That means they know the road well, and exactly how wide it is at every spot, and what its rules are (one-way vs. two-way and so on.) They will know their own width and the width of oncoming vehicles accurately. If they can't safely drive a road, they won't drive it. If it's a rare road, the cost of that will be accepted. Driving every road everywhere is a nice dream, but not necessary to have a highly useful product. While Google's ideal prototype is planned to be released for urban situations without a wheel, cars that need to go places where they can't drive will continue to offer wheels or other interfaces (joysticks, tablet apps) that let a human guide them to get through problems.

The two-cars meeting problem is interesting because it's actually one where the cars can far outperform humans. It's also one of the rare times that communication between cars turns out to be useful. (Typically car to server to server to car, not direct v2v, but that's another matter.)

The reason is that super narrow roads, including country roads and urban back-alleys have occasional wide-spots and turn outs where people can pass. They have to, to be two-way. And these will all be on the map. Cars on such a road would desire traffic data about other cars on the road. They will be able to make predictions about when they might encounter another car coming the other way. Most interestingly, one or both of the cars can adjust their speed so that they will encounter one another precisely at one of the wider spots where passing can take place.

In fact, if they do this well, they might drive a one-lane road at a nice fast speed, barely slowing down in these wider passing zones, in part because by knowing the width of the vehicles they will be able to confidently pass quite closely. If a robocar is meeting a human driven car, it would leave some slop, picking the right passing zone, arriving early in case the other car is faster than expected, waiting if it is slower.

This remarkable ability would allow us to build low-traffic roads and alleys which are mostly only one lane wide, but which could carry traffic fairly quickly and safely in both directions. Gassée's problem is far from a problem -- it's actually a great opportunity to vastly decrease the cost and land requirements of road construction. I wrote about this a couple of years ago, in fact.

Even without communication, a robocar would do pretty well here. Its map would tell it, should it encounter another vehicle on the road it can't pass, just where the closest passing spot is. It could back up if need be, or if the other car should back up, it could nudge in that direction, or even display instructions to a human driver on a screen. It would be able to do this far better than humans could because of its accurate measurements and driving ability. Generally, any human car should defer to the robocar's superior knowledge and superior ability to manage a close pass-by. The car would figure it out the moment it sensed the other car, and immediately adjust speed to meet at a passing point, or possibly to back up. Unlike humans, they will be able to drive in reverse at high speed if they have 360 degree sensors.

Human drivers could actually play a role in this. Those running a mobile app like WAZE could know about other cars running the app, or robocars. The app could give them advice to speed up or slow down to encounter the other car at a wide spot. Of course, if there are cars not using the app, they would just fall back to the old fashioned human approach. One could imagine a sign at the entry to a narrow road saying, "We recommend running the XYZ app for a smoother trip down this road."

Not all these problems that people put forward were as easily resolved as this one, so I am not calling for people to "shut up and let the experts get to work." There are many problems yet to be solved. Most of them can be be solved by punting because you don't need to drive everywhere. Though Google has shown that having a steering wheel that can be grabbed while moving is a bad idea, I do expect most cars to have some form of control that can be activated when a car is stopped. If a road needs the human touch, it will be available.

Comments

People keep saying "what would a robocar do in (situation)" as though it's a criticism.

I always say "it would do the same thing that a regular car does when it has to drive through water: Nothing, because someone built a bridge there". There's nothing unfamiliar about restricting vehicle operation to specific areas. Few modern cars can drive very fast on surfaces that aren't graded and prepared as roads (and even then they need a proper paved and smooth road to move at maximum speed.) Maybe the solution to "what do robocars do about narrow roads" is "don't let them go down narrow roads".

Absolutely this will be the answer at the beginning. If there is a situation the cars can't handle safely, they won't do it. You will either need to get a means to steer (such as a pop-out steering wheel) and drive those manually or semi-auto, or you will need to take a taxi on those roads, or you will use a manual rental car if you need those roads.

People do act like it's a deal-breaker for the tech if it can't drive everywhere. It's annoying of course but hardly a deal breaker.

It's ok to ban robocars from using narrow roads if they're operating in the U.S.

However, the vast majority of European villages, towns and cities pre-date the invention of the car. In the UK, pretty much every journey contains several sections of narrow awkward roads. My 18 mile commute contains at least 10 separate sections of single lane driving. It's just the norm here.

i think it's going to be hard to create algorithms to handle these trickle sections. Often there is a local protocol for each section. It's really jarring when you get a driver who doesn't know the procedure. On my commute, this is pretty rare because everyone is a regular driver of that route.

Obviously nothing's impossible but if Google was based in the rural UK then I reckon their moonshot would be personalised helicopters rather than robocars.

So why is it going to be hard to handle these tricky sections?

Today's self-driving cars are all about having a map of the road before driving it. Knowing exactly how wide it is at every spot, and yes, what the local protocol is if there is one. (Should the team not know the local protocol, it will only take one incident before locals inform them of it. It's not like people in the old cities don't have e-mail.)

Do you have some examples of the local protocols? The computer will be able to calculate the most rational protocol without being told, the one that minimizes backing up by vehicles that can't go as easily in reverse as forward the way robocars can. If the local protocol is not a rational one, it could still be taught.

But again, the approach planned is not to have a computer figure this out at all, but to have humans figure it out, and record that in the map. The computer only has to figure it out if the road changes shape or rules and makes the map wrong. Which happens, but pretty rarely on the grand scheme of things.

I fully appreciate all those points and agree with you. My more general point was that these problems occur with a far greater frequency in places built before the invention of the car.

The West coast of the US has incorporated the car into it's urban planning from the very start. It's a fundamentally different built environment to my locale.

You said yourself that Google or whoever 'will not release their cars to operate on roads where the unsolved situations may occur.' Sadly, I think that means that much of Europe will be waiting significantly longer than the US for the mass adoption of fully autonomous cars.

The link is a notorious local bottleneck. It has 2 restricted entry points at the top and bottom of the hill and there is also a road feeding into the middle of the narrow section between the 2 narrow entrances. I would guess that this is a harder problem than the drive through Taco restaurant.

https://www.google.co.uk/maps/place/Banwell,+North+Somerset/@51.3277808,-2.8647506,3a,75y,142.37h,68.87t/data=!3m7!1e1!3m5!1sHeeO6paklHbrsruiBrZtVw!2e0!6s%2F%2Fgeo3.ggpht.com%2Fcbk%3Fpanoid%3DHeeO6paklHbrsruiBrZtVw%26output%3Dthumbnail%26cb_client%3Dmaps_sv.tactile.gps%26thumb%3D2%26w%3D203%26h%3D100%26yaw%3D27.190214%26pitch%3D0!7i13312!8i6656!4m2!3m1!1s0x4871f781beb85a07:0x6e051b873cfdae67!6m1!1e1

There certainly are more roads in Europe that are harder to drive. But what fraction of all the roads are these? This is mostly the older roads, particularly roads in the old cities. Human drivers avoid these roads already -- many old cities restrict what cars can come into them. This has not prevented 82% of European ground travel from happening in private cars.

Robocars don't have to drive every road to be highly useful. It's not a big deal if there are roads they don't handle. Those can be handled with human drivers, or human driven taxis, or in some cases walking or cycling or scootering the last few 100m.

The road you show in the streetview is one I think is ideal for solution with self driving cars and some web resource. I think they could do a lot better than people at it. This is one of those things, like doing a square root, that computers should do a lot better than people. Actually, I think a road like this would also be well managed if the human drivers ran a smartphone app to manage such roads. Not that humans are likely to be good about that unless there is a fine for not doing so. The robots would always do it.

But even if they don't like this road, all that means is that people who live on this road or go to this road are less likely to be robocar users for those trips. This is hardly a dealbreaker for all European use!

Brad,

Do you think that there will be remote human operators that could wire into robocars if necessary? What do you think would be the technical hurdles to get that type of system to work?

Thanks!

Yes, most cars will, if they don't have a capable human in them, want the ability to accept advice and commands from a remote operator. In fact, to make sure they can always get that, they may minimize time they spend in cell phone dead spots -- which they will know the locations of in advance.

The remote operators will not "drive" the cars like an RC car. Rather, they will issue commands, like clicking on a photo or video sent by the car and saying, "go there," letting the car worry about how to steer and stopping for obstacles. If a car is so disabled that this is impossible, they might, very slowly, move it with more specific commands like "turn 20 degrees, drive 10 meters" but that should be extremely rare, and be done quite slowly just to get a car off the road.

Brad,

I have been thinking about the improvement rate of manual intervention in autonomous car tests. I am trying to string together a timeline for the Google car's total miles driven over time and total autonomous miles driven over time. I have the following timeline put together from public announcements before Google started publishing monthly reports:

Oct-10 140,000 miles driven
Mar-12 200,000 miles driven
Aug-12 300,000 miles driven
Mar-13 500,000 miles driven
Apr-14 700,000 miles driven

My question is: are each of these milestones total miles tested? (So some % was autonomous mode and some % was manual mode of this total) Or are these all data points for miles in autonomous mode, or a mix? From an initial look it appears like the last point (700,000 miles) may be the miles driven in autonomous mode.

It looks like Google’s cars are currently running in manual mode about 35% of the time. Where do you think that % will sit in 2020?

Google's announcements have sometimes given just one number (and not always saying which but you can probably figure it out from your spreadsheet) and sometimes they have given both numbers -- total miles driven and total miles in autodrive mode.

The miles in autodrive mode is the more important of the two numbers, but the total miles is also important, it reveals how much data collection has been done to train the cars to handle every situation that comes up in those miles.

The most important number, revealed only once, I think, is "autodrive miles between necessary interventions." The last time they gave this number it was 83,000 miles -- not as good as a human, but approaching it.

Add new comment