There’s been a lot of press recently about an article in Slate by Lee Gomes which paints a pessimistic picture of the future of robocars, and particularly Google’s project. The Slate article is a follow-on to a similar article in MIT Tech Review
Gomes and others seem to feel that they and the public were led to believe that current projects were almost finished and ready to be delivered any day, and they are disappointed to learn that these vehicles are still research projects and prototypes. In a classic expression of the Gartner Hype Cycle there are now predictions that the technology is very far away.
Both predictions are probably wrong. Fully functional robocars that can drive almost everywhere are not coming this decade, but nor are they many decades away. But more to the point, less-functional robocars are probably coming this decade — much sooner than these articles expect, and these vehicles are much more useful and commercially viable than people may expect.
There are many challenges facing developers, and those challenges will keep them busy refining products for a long time to come. Most of those challenges either already have a path to solution, or constrain a future vehicle only in modest ways that still allow it to be viable. Some of the problems are in the “unsolved” class. It is harder to predict when those solutions will come, of course, but at the same time one should remember that many of the systems in today’s research vehicles were in this class just a few years ago. Tackling hard problems is just what these teams are good at doing. This doesn’t guarantee success, but neither does it require you bet against it.
And very few of the problems seem to be in the “unsolvable without human-smart AI” class, at least none that bar highly useful operation.
Gomes’ articles have been the major trigger of press, so I will go over those issues in detail here first. Later, I will produce an article that has even more challenges than listed, and what people hope to do about them. Still, the critiques are written almost as though they expected Google and others, rather than make announcements like “Look at the new milestone we are pleased to have accomplished” to instead say, “Let’s tell you all the things we haven’t done yet.”
Gomes begins by comparing the car to the Apple Newton, but forgets that 9 years after the Newton fizzled we had the success of the Palm Pilot, and 10 years after that Apple came back with the world-changing iPhone. Today, the pace of change is much faster than in the 80s.
Here are the primary concerns raised:
Maps are too important, and too costly
Google’s car, and others, rely on a clever technique that revolutionized the DARPA challenges. Each road is driven manually a few times, and the scans are then processed to build a super-detailed “ultramap” of all the static features of the road. This is a big win because big server computers get to process the scans in as much time as they need, and see everything from different angles. Then humans can review and correct the maps and they can be tested. That’s hard to beat, and you will always drive better if you have such a map than if you don’t.
Any car that could drive without a map would effectively be a car that’s able to make an adequate map automatically. As things get closer to that, making maps will become cheaper and cheaper.
Naturally, if the road differs from the map, due to construction or other changes, the vehicle has to notice this. That turns out to be fairly easy. Harder is assuring it can drive safely in this situation. That’s still a much easier problem than being able to drive safely everywhere without a map, and in the worst case, the problem of the changed road can be “solved” by just the ability to come to a safe stop. You don’t want to do that super often, but it remains the fail-safe out. If there is a human in the car, they can guide the vehicle in this. Even if the vehicle can’t figure out where to go to be safe, the human can. Even a remote human able to look at transmitted pictures can help the car with that — not live steering, but strategic guidance.
This problem only happens to the first car to encounter the surprise construction. If that car is still able to navigate (perhaps with human help,) the map can be quickly rebuilt, and if the car had to stop, all unmanned cars can learn to avoid the zone. They are unmanned, and thus probably not in a hurry.
The cost of maps
In the interests of safety, a lot of work is put into today’s maps. It’s a cost that somebody like Google or Mercedes can afford if they need to, (after all, Google’s already scanned every road in many countries multiple times) but it would be high for smaller players.
Even today’s costs are a tiny fraction of the cost of building and even maintaining the roads. It doesn’t take many users sharing the cost to make it quite affordable. This cost is going to go down fast, on a Moore’s law curve, as the software gets better and better and less human oversight is needed. As I’ll discuss below, only mapping and driving the most popular of roads remains a viable option. Even so, it would be folly to think the cost of the current first-generation mapping approaches won’t drop a great deal as the need to map roads expands.
Unmapped traffic signals
One particular problem raised is the concept of an unmapped traffic signal, such as the temporary traffic signals that construction crews use. In the long term, I do believe that all road crews and city planners will come to never modify roads without recording it in a database first — the trend of cheap networking and sensors that marketers call the “internet of things” will make this automatic in time. One can even imagine car makers paying to put a transponder (ie. a used cell phone) on every temporary traffic signal out there. (Google happens to know a bit about phones.)
Until that happens, we are not without a solution. While computer vision is not as good as humans at recognizing the millions of things in our environment, it’s actually pretty good at recognizing specific things, especially static things, if you take the effort to train it. And as such, efforts are underway to make systems that will reliably identify troublesome items, like new road signs, new traffic signals, and even police officers and construction flaggers. By reliably, I mean “as well as humans are able to recognize them.” The systems will be doing this not just with cameras, but the fused result of 3-D LIDAR and cameras, which gives them a much better ability in this area.
Gomes is also worried about the sun blinding a camera looking for a traffic signal. While that happens to humans too, the systems take a very simple approach: “If you don’t confirm green, you don’t go.” If blinded, the systems will slowly stop. Even at a green light, which is annoying if it happens with any frequency, but is generally safe. (If there’s a person inside or available to help remotely, they might tell it to go.) I joke that if you stop for more than a few seconds at a green, an audible alert will sound from any cars behind you, but that’s probably not enough to go without confirming green!
Gomes suggests that the lack of parking effort presents an issue. I would guess the reason Google hasn’t done a lot on parking is because it’s actually pretty easy. There are already commercial products out there that have addressed many of the problems of parking.
Weather can be challenging
Nobody has worked really hard on driving in snow yet. There are lots of markets where snow is not an issue, of course, or where telling people they must take the wheel during snow is not a deal-breaker. Still, in the academic community there is promising research on techniques to handle snow, such as ground penetrating radar. In the event those techniques don’t pan out, one sure approach is to do some relatively inexpensive additions to the road in places with snow, such as low cost magnets punched into the road, or special poles at the side of the road — which are already used in many areas to let the snowplows figure out where the road is when it’s snow covered.
There are issues with very heavy rain and blizzards, but new generations of LIDAR are showing superior results there, and frankly rain that is so heavy it blinds the sensors of a car is probably not rain you should be driving in, though you could choose to take the wheel of a regular car in such situations.
More work is needed on identifying debris and other road hazards. Again, at first, the vehicles will just be conservative. They might foolishly slow or stop for a crumpled newspaper on the road. Annoying, but tolerable if it’s not too frequent, and in time the systems will rise to this challenge too. While the list of “rare” events is large, what matters is that they are rare. As long as there is a conservative solution and the events are truly rare, a product can be useful. Each problem will be solved, one by one, based on how important it is.
Promising the moon
Gomes’ largest mistake is he feels that Google and other vendors have promised the world a complete solution any day now. While that will come, it is certainly not something people should expect on day one. No technology works that way.
The first vehicles that operate without supervision will only drive limited areas. That means if they do go between your origin and destination, they work for you on that trip. If they don’t, you use all the existing methods — manually driven cars, taxis, Uber-like services or transit to name a few. The more roads they can cover, the more trips they can serve.
Some vehicles will still have steering wheels, so you can have them drive where they can and you drive where and when they can’t. This is what all the car companies are making. Google has said their experiments suggest that switching from human operation to automatic operation while underway is risky, so their latest prototype has no steering wheel. Vehicles without wheels won’t be driven by you but they can take you to the border of their service areas where you can switch to a seamless rental car — think Zipcar or Hertz — or your own car or any other transport mode.
At the same time, while Google is concerned about the transition from automatic to manual driving, it’s quite reasonable to build a vehicle that has a pop-out handlebar or wheel which only allows the transition from manual to automatic when the vehicle has stopped or is otherwise in an assuredly safe state. So it might take you to the edge of its service areas, stop and pop out the handlebars and let you take it from there until you enter the service area again. Not ideal, but still very, very good.
It is perhaps harder to understand as well that the goal with these vehicles is not perfection. Not even perfect safety. They will probably never be as safe as an elevator. They will have accidents. The real goal is to make vehicles that drive more safely than humans do, and thus reduce the number of accidents. It is thus an error to suggest you can’t have people use the vehicles until every safety situation can be resolved completely. Rather, you just must assure that those situations are rare enough, and what the system does is good enough, that it outperforms people. Like many other computer systems, it is far from necessary for the systems to have full human situational awareness to outperform us.
The courts might take a different tack, and heavily punish anything less than perfection. If they do, and the legislatures don’t fix that and make them look at the overall safety record, then the technology will just flourish outside the jurisdictions that punish it.
The challenge of predictions
The great challenge of a forecaster is to look at a nascent technology and judge where it’s going:
- It may be impossible to refine it to usability, and thus will not be a success until many major breakthroughs come.
- It may seem clunky and primitive, but this is deceptive and it will soon take the world by storm.
- It may truly be clunky, and need decades to really reach a level of major use.
All 3 of these things happen when we predict the future of new technologies based on the prototypes or first products. The Apple Newton might be viewed as having done all of these depending on how you look at the time scales. The Newton was flawed and failed, but inside was the germ of something world-changing, though that didn’t happen for 2 decades.
One common cause of error is the difference between linear extrapolation, which most people are used to, and exponential extrapolation, to which only people in the computer industry are truly accustomed. On an exponential path you double every interval. On the linear path of 10 intervals, halfway along the path you are halfway there. On an exponential path of 10 doublings, halfway along the path you are only 3% of the way there and can easily conclude failure is certain — but you would be very wrong.
That doesn’t mean every grand exponential prediction is right, of course, just that it’s very hard to make good predictions. Everybody knows that overly optimistic prediction is common, but the existence of massive underprediction is less well understood.
That doesn’t mean there isn’t a lot left to do, and plenty of uncertainty over when it will get done. But there is a reason for the optimism. More to the point, because the changes to be wrought by robocars are so grand, even a 10% chance means that everybody should start getting ready.
As always, I should note that while I have consulted for Google’s self-driving car team, I am not speaking on their behalf.