Executive Summary: A rundown of different approaches for validation of self-driving and driver assist systems, and a recommendation to Tesla and others to have countermeasures to detect drivers not watching the road, and permanently disable their Autopilot if they show a pattern of inattention.

The recent fatality for a man who was allowing his car to be driven by the Tesla "autopilot" system has ignited debate on whether it was appropriate for Tesla to allow their system to be used as it was.

Tesla's autopilot is a driver assist system, and Tesla tells customers it must always be supervised by an alert driver ready to take the controls at any time. The autopilot is not a working self-driving car system, and it's not rated for all sorts of driving conditions, and there are huge numbers of situations that it is not designed to handle and can't handle. Tesla knows that, but the public, press and Tesla customers forget that, and there are many Tesla users who are treating the autopilot like a real self-driving car system, and who are not paying attention to the road -- and Tesla is aware of that as well. Press made this mistake as well, regularly writing fanciful stories about how Tesla was ahead of Google and other teams.

Brown, the driver killed in the crash, was very likely one of those people, and if so, he paid for it with his life. In spite of all the warnings Tesla may give about the system, some users do get a sense of false security. There is debate if that means driver assist systems are a bad idea.

There have been partial self-driving systems that require supervision since the arrival of the cruise control. Adaptive cruise control is even better, and other car companies have released autopilot like systems which combine adaptive cruise control with lane-keeping and forward collision avoidance, which hits the brakes if you're about to rear end another car. Mercedes has sold a "traffic jam assist" like the Telsa autopilot since 2014 that only runs at low speeds in the USA. You can even go back to a Honda demo in 2005 of an autopilot like system.

With cruise control, you might relax a bit but you know you have to pay attention. You're steering and for a long time even the adaptive cruise controls did not slow down for stopped cars. The problem with Tesla's autopilot is that it was more comprehensive and better performing than earlier systems, and even though it had tons of things it could not handle, people started to trust it with their lives.

Tesla's plan can be viewed in several ways. One view is that Tesla was using customers as "beta testers," as guinea pigs for a primitive self-drive system which is not production ready, and that this is too much of a risk. Another is that Tesla built (and tested) a superior driver assist system with known and warned limitations, and customers should have listened to those warnings.

Neither is quite right. While Tesla has been clear about the latter stance, with the knowledge that people will over-trust it, we must face the fact that it is not only the daring drivers who are putting themselves at risk, it's also others on the road who are put at risk by the over-trusting drivers -- or perhaps by Tesla. What if the errant car had not gone under a truck, but instead hit another car, or even plowed into a pedestrian when it careened off the road after the crash?

At the same time, Tesla's early deployment approach is a powerful tool for the development and quality assurance of self-drive systems. I have written before about how testing is the big unsolved problem in self-driving cars. Companies like Google have spent many millions to use a staff of paid drivers to test their cars for 1.6 million miles. This is massively expensive and time consuming, and even Google's money can't easily generate the billions of miles of testing that some feel might be needed. Human drivers will have about 12 fatalities in a billion miles, and we want our self-driving cars to do much better. Just how we'll get enough verification and testing done to bring this technology to the world is not a solved problem.

A Tesla blog post describes the first fatality involving a self drive system. A Tesla was driving on autopilot down a divided highway. A truck made a left turn and crossed the Tesla's lanes. A white truck body against a bright sky is not something the MobilEye camera system in the Tesla perceives well, and it is not designed for cross traffic.

With Mobility on Demand, you don't buy a car, you buy rides. That's certainly Uber's plan, and is a plan that makes sense for Google, Apple and other no-car companies. But even Daimler, with Car2Go/Car2Come, BMW with DriveNow and GM with Lyft plan to sell you a ride rather than a car, because it's the more lucrative thing to do.

So what does that car of the future look like? There is no one answer, because in this world, the car that is sent to pick you up is tailored to your trip. The more people traveling, the bigger the car is. If your trip does not involve a highway, it may not be a car capable of the highway. If your trip is up to a mountain cabin, it's more like an SUV, but you never use an SUV to go get a bottle of milk the way we do today. If it's for a cruise to the beach on a sunny day, the roof may have been removed at the depot. If it's for an overnight trip to a country home, it may be just beds.

I outlined many of these changes in this article on design changes in cars but today I will focus on the incredibly cheap and simple design of what should become the most common vehicle made, namely the car designed for a short urban trip by one person. That's 80% of trips and around 45% of miles, so this should be a large fraction of the fleet. I predict a lot of these cars will be made every year -- more than all the cars made today, even though they are used as taxis and shared among many passengers. What does it look like?

Small

A car for 1-2 people will be small. It will probably be around 1.5m wide, narrow enough that you can fit two in a lane, and have it park very efficiently when it has to wait. If it's for just one person, it won't be very long either. For two people, there will be a "face to face" configuration which is longer and an "tandem" configuration which is a bit shorter. The 2 person vehicles aren't a lot bigger or heavier than the one person, so they might be the most common cars, since you can serve a solo rider fairly efficiently with one, even if not perfectly efficient.

A car that is so narrow can't corner very fast. A wide stance is much more stable. There are a few solutions to that, including combinations of these:

• The wheels bank independently, allowing the vehicle to lean like a motorcycle when in corners. This is the best solution, but it costs some money.
• Alternately it's a two wheeler, which is also able to lean, but has other tricks like the LIT Motors C-1 to stay upright.
• It's electric, and has all the batteries in the floor, giving it a very low center of gravity. (One extreme example of this is the Tango, which uses lead batteries deliberately to give it that stability.)
• It never goes on fast roads, so it never needs to corner very fast, and its precision robot driving assures it never corners so fast as to become unstable, and it plans its route accordingly.

Not super aerodynamic

The car already has a big win when it comes to aerodynamic drag by only being half-width. The non-highway version probably gives back a bit of that because you don't need to worry as much about that if you are not going fast. Energy lost to drag goes up with the square of velocity. So a 30mph car has 1/4 the drag of a 60mph car, and 1/8th the drag of a similar car of full width. The highway car needs to be shaped as close to a "teardrop" as you can, but the city car can get away with being a bit taller for more comfortable seating and entry/exit.

When I give talks on robocars, the most common question, asked almost all the time, is the one known as the "trolley problem" question, "What will the car do if it has to choose between killing one person or another" or other related dilemmas. I have written frequently about how this is a very low priority question in reality, much more interesting to philosophy classes than it is important. It is a super-rare event and there are much more important everyday ethical questions that self-driving car developers have to solve long before they will tackle this one.

In spite of this, the question persists in the public mind. We are fascinated and afraid of the idea of machines making life or death decisions. The tiny number of humans faced with such dilemmas don't have a detailed ethical debate in their minds; they can only go with their "gut" or very simple and quick reasoning. We are troubled because machines don't have a difference between instant and carefully pondered reactions. The one time in billions of miles(*) that a machine faces such a question it would presumably make a calculated decision based on its programming. That's foreign to our nature, and indeed not a task desired by programmers or vendors of robocars.

There have been calls to come up with "ethical calculus" algorithms and put them in the cars. As a programmer, I could imagine coding such an algorithm, but I certainly would not want to, nor would I want to be held accountable for what it does, because by definition, it's going to do something bad. The programmer's job is to make driving safer. On their own, I think most builders of robocars would try to punt the decision elsewhere if possible. The simplest way to punt the decision is to program the car to follow the law, which generally means to stay in its right-of-way. Yes, that means running over 3 toddlers who ran into the road instead of veering onto the sidewalk to run over Hitler. Staying in our lane is what the law says to do, and you are not punished for doing it. The law strongly forbids going onto the sidewalk or another lane to deliberately hit something, no matter who you might be saving.

We might not like the law, but we do have the ability to change it.

Thus I propose the following: Driving regulators should create a special panel which can rule on driving ethics questions. If a robocar developer sees a question which requires some sort of ethical calculation whose answer is unclear, they can submit that question to the panel. The panel can deliberate and provide an answer. If the developer conforms to the ruling, they are absolved of responsibility. They did the right thing.

The panel would of course have people with technical skill on it, to make sure rulings are reasonable and can be implemented. Petitioners could also appeal rulings that would impede development, though they would probably suggest answers and describe their difficulty to the panel in any petition.

The panel would not simply be presented with questions like, "How do you choose between hitting 2 adults or one child?" It might make more sense to propose formulae for evaluating multiple different situations. In the end, it would need to be reduced to something you can do with code.

Very important to the rulings would be an understanding of how certain requirements could slow down robocar development or raise costs. For example, a ruling that car must make a decision based on the number of pedestrians it might hit demands it be able to count pedestrians. Today's robocars may often be unsure whether a blob is 2 or 3 pedestrians, and nobody cares because generally the result is the same -- you don't want to hit any number of pedestrians. Likeways, requirements to know the age of people on the road demands a great deal more of the car's perception system than anybody would normally develop, particularly if you imagine you will ask it to tell a dwarf adult from a child.

Reports from Tesla suggest they are gathering huge amounts of driving data from logs in their cars -- 780 million miles of driving, and as much as 100 million miles in autopilot mode. This contrasts with the 1.6 million miles of test operations at Google. Huge numbers, but what do they mean now, and in the future?

Executive summary: Can our emotional fear of machines and the call for premature regulation be mollified by a temporary increase in liability which takes the place of specific regulations to keep people safe?

Here is the first report of a real Tesla autopilot crash. To be fair to Tesla, their owner warnings specify fairly clearly that the autopilot could crash in just this situation -- there is a stalled car partly in the lane, and the car in front of you swerves around it, revealing it with little time for you or the autopilot to react.

The deeper issue is the way that the improving quality of the Tesla Autopilot and systems like it are lulling drivers into a false sense of security. I have heard reports of people who now are trusting the Tesla system enough to work while being driven, and indeed, most people will get away with this. And as people get away with it more and more, we will see more people driving like this driver, not really prepared to react. This is one of the reasons Google decided not to make a system which requires driver takeover ever. As the system gets better, does it get more dangerous?

Some technical notes:

• This is one of the things LIDAR is much more reliable at seeing than cameras. Of course, whether you can swerve once the LIDAR sees it is another matter.
• On the other hand, this is where radar fails. I mean the stalled car is clear on radar, but it's stationary, so you can't tell it from the road or guardrail which are also stationary.
• This is one of the classic V2V value propositions, but it's not a good one. You don't need 10ms latency to have a stalled car tell you it is stalled. Far better that car report to a server that it's stalled and for everybody coming down that road to learn it, whether they have line of sight radio to the stall, or V2V at all. Waze already reports this just with human manual reporting and that's a really primitive way to do it.

Declaration of Amsterdam

Last month, various EU officials gathered in Amsterdam and signed the Declaration of Amsterdam which outlines a plan for normalizing EU laws around self-driving cars. The meeting also included a truck automation demo in the Netherlands and a self-drive transit shuttle demonstration. It's a fairly bland document, more an expression of the times, and it sadly spends a lot of time on the red herring of "connected" vehicles and V2V/V2I, which governments seem to love, and self-driving car developers care very little about.

Let's hope the regulatory touch is light. The reality is that even the people building these vehicles can't make firm pronouncements on their final form or development needs, so governments certainly can't do that, and we must be careful of attempts to "help" that hinder. We already have a number of examples of that happening in draft and real regulations, and we've barely gotten started. For now, government statements should be limited to, "let's get out of the way until people start figuring out how this will actually work, unless we see somebody doing something demonstrably dangerous that can't be stopped except through regulations." Sadly, too many regulators and commentators imagine it should be, "let's use our limited current knowledge to imagine what might go wrong and write rules to ban it before it happens."

Speech from the Throne

It was a sign of the times when her Majesty the Queen, giving the speech from the throne in the UK parliament, laid out some elements of self-driving car plans. The Queen drove jeeps during her military days, and so routinely drives herself at her country estates, otherwise she would be among the set of people most used to never driving.

The UK has 4 pilot projects in planning. Milton Keynes is underway, and later this year, a variation of the Ultra PRT pods in use at T5 of Heathrow airport -- they run on private tracks to the car park -- will go out on the open road in Greenwich. They are already signing up people for rides.

Car companies thinking differently

In deciding which car companies are going to survive the transition to robocars, one thing I look for is willingness to stop thinking like a traditional car company which makes cars and sells them to customers. Most car company CEOs have said they don't plan to keep thinking that way, but what they do is more important than what they say.

Uber has announced the official start of self-driving tests in Pittsburgh. Uber has been running their lab for over a year, and had various vehicles out there mapping and gathering data, but their new vehicle is sleeker and loaded with sensors - more than on Google's cars or most of the other research cars I have seen. You can see several lidars on the roof and bumpers, and a seriously big array of cameras and other sensors.

Today sees the un-stealthing of a new company called Otto which plans to build self-driving systems for long haul trucks. The company has been formed by a skilled team, including former members of Google's car team and people I know well. You can see their opening blog post

My entire focus on this blog, and the focus of most people in this space, has been on cars, particularly cars capable of unmanned operation and door-to-door service. Most of those not working on that have had their focus on highway cars and autopilots. The highway is a much simpler environment so much easier to engineer, but it operates at higher speeds so the cost of accidents is worse.

That goes doubly true for trucks that are fast, big and massive. At the same time, 99% of truck driving is actually very straightforward -- stay in a highway lane, usually the slow one, with no fancy moving about.

Some companies have done exploration of truck automation. Daimler/Freightliner has been testing trucks in Nevada. Volvo (trucks and cars together) has done truck and platooning experiments, notably the Sartre project some years ago. A recent group of European researchers did a truck demonstration in the Netherlands, leading up to the Declaration of Amsterdam which got government ministers to declare a plan to modify regulations to make self-driving systems legal in Europe. Local company Peloton has gone after the more tractable problem of two-truck platoons with a driver in each truck, aimed primarily at fuel savings and some safety increases.

Safety

While trucks are big and thus riskier to automate, they are also risky for humans to drive. Even though truck drivers are professionals who drive all day, there are still around 4,000 killed every year in the USA in truck accidents. More than half of those are truck drivers, but a large number of ordinary road users are also killed. Done well, self-driving trucks will reduce this toll. Just as with cars, companies will not release the systems until they believe they can match and beat the safety record of human drivers.

The Economics

Self-driving trucks don't change the way we move, but they will have a big economic effect on trucking. Driver pay accounts for about 25-35% of the cost of truck operation, but in fact early self-driving won't take away jobs because there is a serious shortage of truck drivers in the market -- companies can't hire enough of them at the wages they currently pay. It is claimed that there are 50,000 job openings unfilled at the present time. Truck driving is grueling work, sometimes mind-numbing, and it takes the long haul driver away from home and family for over a week on every long-haul run. It's not very exciting work, and it involves long days (11 hours is the legal limit) and a lot of eating and sleeping in truck stops or the cabin of the truck.

Average pay is about 36 cents/mile for a solo trucker on a common route. Alternately, loads that need to move fast are driven by a team of two. They split 50 cents/mile between them, but can drive 22 hours/day -- one driver sleeps in the back while the first one takes the wheel. You make less per mile per driver, but you are also paid for the miles you are sleeping or relaxing.

A likely first course is trucks that keep their solo driver who drives up to 11 hours -- probably less -- and have the software drive the rest. Nonstop team driving speed with just one person. Indeed, that person might be an owner-operator who is paying for the system as a businessperson, rather than a person losing a job to automation. The human would drive the more complex parts of the route (including heavy traffic) while the system can easily handle the long nights and sparse heartland interstate roads.

The economics get interesting when you can do things that are expensive for human drivers and teams. Aside from operating 22 or more hours/day at a lower cost, certain routes will become practical that were not economic with human drivers, opening up new routes and business models.

The Environment

Computer driven trucks will drive more regularly than humans, effectively driving in "hypermile" style as much as they can. That should save fuel. In addition, while I would not do it at first, the platooning experimented with by Peloton and Sartre does result in fuel savings. Also interesting is the ability to convert trucks to natural gas, which is domestic and burns cleaner (though it still emits CO2.) Automated trucks on fixed routes might be more willing to make this conversion.

Road wear

There is strong potential to reduce the damage to roads (and thus the cost of maintaining them, which is immense and seriously in arrears) thanks to the robotruck. That's because heavy trucks and big buses cause almost all the road wear today. A surprising rule of thumb is that road damage goes up with the 4th power of the weight per axle. As such an 80,000lb truck with 34,000lb on two sets of 2 axles and 6,000lb on the front axle does around 2,000 times the road damage of a typical car!

A recent news story from Utah describes a Tesla which entered self-park ("summon") mode and drive itself into the back of a flatbed truck raises some interesting issues.

If you had asked me recently what big car company was the furthest behind when it came to robocars, one likely answer would be Fiat-Chrysler. In fact, famously, Chrysler ran ads several years ago during the superbowl making fun of self-driving cars and Google in particular:

I have often written on the challenge facing existing automakers in the world of robocars. They need to learn to completely switch their way of thinking in a world of mobility on demand, and not all of them will do so. But they face serious challenges even if they are among the lucky ones who fully "get" the robocar revolution, change their DNA and make products to compete with Google and the rest of the non-car companies.

While I'm very excited about the coming robocar world, there are still many unsolved problems. One I've been thinking about, particularly with my recent continued thinking on transit, is how to provide robotaxi service to the poor, which is to say people without much money and without credit and reputations.

In particular, we want to avoid situations where taxi fleet operators create major barriers to riding by the poor in the form of higher fees, special burdens, or simply not accepting the poor as customers. If you look at services like Uber today, they don't let you ride unless you have a credit card, though in some cases prepaid debit cards will work.

Today a taxi (or a bus or Uber style vehicle) has a person in it, primarily to drive, but they perform another role -- they constrain the behaviour of the rider or riders. They reduce the probability that somebody might trash the vehicle or harass or be violent to another passenger.

Of course, such things happen quite rarely, but that won't stop operators from asking, "What do we do when it does happen? How can we stop it or get the person who does it to pay for any damage?" And further they will say, "I need a way to know that in the rare event something goes wrong, you can and will pay for it." They do this in many similar situations. The problem is not that the poor will be judged dangerous or risky. The problem is that they will be judged less accountable for things that might go wrong. Rich people will throw up in the back of cars or damage them as much as the poor, perhaps more; the difference is there is a way to make them pay for it. So while I use the word poor here, I really mean "those it is hard to hold accountable" because there is a strong connection.

As I have outlined in one of my examinations of privacy a taxi can contain a camera with a physical shutter that is open only between riders. It can do a "before and after" photograph, mostly to spot if you left items behind, but also to spot if you've damaged or soiled the vehicle. Then the owner can have the vehicle go for cleaning, and send you the bill.

But they can only send you the bill if they know who you are and have a way to bill you. For the middle class and above, that's no problem. This is the way things like Uber work -- everybody is registered and has a credit card on file. This is not so easy for the poor. Many don't have credit cards, and more to the point, they can't show the resources to fix the damage they might do to a car, nor may they have whatever type of reputation is needed so fleet operators will trust them. The actions of a few damn the many.

The middle class don't even need credit cards. Those of us wishing to retain our privacy could post a bond through a privacy protecting intermediary. The robotaxi company would know me only as "PrivacyProxy 12323423" and I would have an independent relationship with PrivacyProxy Inc. which would accept responsibility for any damage I do to the car, and bill me for it or take money from my bond if I'm truly anonymous.

Options for the poor

Without the proxy, robotaxi operators will want some sort of direct accountability from passengers for any problems they might cause. Even for the middle class, it mostly means being identified, so if damage is found, you can be tracked down and made to pay. The middle class have ability to pay, and credit. The poor don't, at least many of them don't.

People with some level of identity (an address, a job) have ways to be accountable. If the damage rises to the level where refusing to fix it is a crime at some level, fear of the justice system might work, but it's unlikely the police are going to knock on somebody's door for throwing up in a car.

In the future, I expect just about everybody of all income levels will have smartphones, and plans (though prepaid plans are more common at lower income levels.) One could volunteer to be accountable via the phone plan, losing your phone number if you aren't. Indeed, it's going to be hard to summon a car without a phone, though it will also be possible using internet terminals, kiosks and borrowing the phones of others.

More expensive rides

A likely solution, seen already in the car rental industry, is to charge extra for insurance for those who can't prove accountability another way. Car rental company insurance is grossly overpriced, and I never buy it because I have personal insurance and credit cards to cover such issues. Those who don't often have to pay this higher price.

It's still a sad reality to imagine the poor having to pay more for rides than for the rich.

An option to mitigate this might be cars aimed at carrying those who are higher risk. These cars might be a bit more able to withstand wear and tear. Their interiors might be more like bus interiors, easily cleaned and harder to damage, rather than luxury leather which will probably be only for the wealthier. To get one, you might have to wait longer. While a middle-class customer ordering a cheap car might be sent a luxury car because that's what's spare at the time, it is less likely an untrusted and poor customer would get that.

Before we go do far, I predict the cost of robotaxi rides will get well below $1/mile, heading down to 30 cents/mile. Even with a 30% surcharge, that's still cheaper than what we have today, in fact it's cheaper than a bus ticket in many towns, certainly cheaper than an unsubsidized bus ticket which tends to run $5-$6. Still my hope for robotaxi service is that it makes good transportation more available to everybody, and having it cost more for the poor is a defect.

In addition, as long as damage levels remain low, as a comment points out, perhaps the added cost on every ride would be small enough that you don't need worry about this for poor or rich. (Though having no cost to doing so does mean more spilled food, drink and sadly, vomit.)

Reputation

Over time, fortunately, poor riders could develop reputations for treating vehicles well. Build enough reputation and you might have access to the same fleet and prices that the middle class do, or at least much cheaper insurance. Cause a problem and you might lose the reputation. It would be possible to build such a reputation anonymously, though I suspect most people and companies would prefer to tie it to identity, erasing privacy. Anonymous reputations in particular can be sold or stolen which presents an issue. One option is to tie the reputation to a photo, but not a name. When you get in the car, it would confirm you match the photo, but would not immediately know your name. (In the future, though, police and database companies will be able to turn the photo into a name easily enough.)

Poor riders would still have to pay more to start, probably, or suffer the other indignities of the lower class ride. However, a poor rider who develops a sterling reservation might be able to get some of that early surcharge back later. (Not if it's insurance. You can't get insurance back if you don't use it, it doesn't work that way!)

It could also be possible for the poor to get friends to vouch for them and give them some starter reputation.

Unfortunately, poor who squander their reputation (or worse, just ride with friends who trash a car) could find themselves unable to travel except at high cost they can't afford. It could be like losing your car.

The government

The government will have an interest in making sure the poor are not left out of this mobility revolution. As such, there might be some subsidy program to help people get going, and a safety net for loss of reputation. This of course comes with a cost. Taxes would pay for the insurance to fix cars that are damaged by riders unable to be held accountable.

The alternative, after all, is needing to continue otherwise unprofitable transit services with human drivers just for the sake of these people who can't get private robocar rides. Transit may continue (though without human drivers) at peak times, but it almost surely vanishes off-peak if not for this.

Recently a reddit user posted this short video of an amazingly lucky driver in Japan who was able to turn his car around just in time to escape the torrent of the tsunami.

The question asked was, how would a robocar deal with this? It turns out there are many answers to this question. For this particular question, as you'll see by the end, the answer is probably "very well."

My recent article on a future vision for public transit drew some ire by those who viewed it as anti-transit. Instead, the article broke with transit orthodoxy by suggesting that smaller vehicles (including cars and single person pods) might produce more efficient transit than big vehicles. Transitophiles love big vehicles for reasons beyond their potential efficiency, so it's a hard sell.

Let's look at the factors which determine what vehicle size makes the best transit.

Perhaps the world's most exciting new technology today are deep neural networks, in particular the convolutional neural networks such as "Deep Learning." These networks are conquering some of the most well known problems in artificial intelligence and pattern matching, and since their development just a few years ago, milestones in AI have been falling as computer systems that match or surpass human capability have been demonstrated. Playing Go is just the most recent famous example.

This is particularly true in image recognition. Over the past several years, neural network systems have gotten better than humans at problems like recognizing street signs in camera images and even beating radiologists at identifying cancers in medical scans.

These networks are having their effect on robocar development. They are allowing significant progress in the use of vision systems for robotics and driving, making those progress much faster than expected. 2 years ago, I declared that the time when vision systems would be good enough to build a safe robocar without lidar was still fairly far away. That day has not yet arrived, but it is definitely closer, and it's much harder to say it won't be soon. At the same time, LIDAR and other sensors are improving and dropping in price. Quanergy (to whom I am an advisor) plans to ship $250 8-line LIDARS this year, and $100 high resolution LIDARS in the next couple of years.

The deep neural networks are a primary tool of MobilEye, the Jerusalem company which makes camera systems and machine-vision ASICs for the ADAS (Advanced Driver Assistance Systems) market. This is the chip used in Tesla's autopilot, and Tesla claims it has done a great deal of its own custom development, while MobilEye claims the important magic sauce is still mostly them. NVIDIA has made a big push into the robocar market by promoting their high end GPUs as the supercomputing tool cars will need to run these networks well. The two companies disagree, of course, on whether GPUs or ASCICs are the best tool for this -- more on that later.

In comes comma.ai

In February, I rode in an experimental car that took this idea to the extreme. The small startup comma.ai, lead by iPhone hacker George Hotz, got some press by building an autopilot similar in capability to many others from car companies in a short amount of time. In January, I wrote an introduction to their approach including how they used quick hacking of the car's network bus to simplify having the computer control the car. They did it with CNNs, and almost entirely with CNNs. Their car feeds the images from a camera into the network, and out from the network come commands to adjust the steering and speed to keep a car in its lane. As such, there is very little traditional code in the system, just the neural network and a bit of control logic.

Here's a video of the car taking us for a drive:

The network is built instead by training it. They drive the car around, and the car learns from the humans driving it what to do when it sees things in the field of view. To help in this training, they also give the car a LIDAR which provides an accurate 3D scan of the environment to more absolutely detect the presence of cars and other users of the road. By letting the network know during training that "there is really something there at these coordinates," the network can learn how to tell the same thing from just the camera images. When it is time to drive, the network does not get the LIDAR data, however it does produce outputs of where it thinks the other cars are, allowing developers to test how well it is seeing things.

This approach is both interesting and frightening. This allows the development of a credible autopilot, but at the same time, the developers have minimal information about how it works, and never can truly understand why it is making the decisions it does. If it makes an error, they will generally not know why it made the error, though they can give it more training data until it no longer makes the error. (They can also replay all other scenarios for which they have recorded data to make sure no new errors are made with the new training data.)

I frequently see people claim that one effect of robocars is that because we'll share the cars (when they work as taxis) and most cars stay idle 95% of the time, that a lot fewer cars will be made -- which is good news for everybody but the car industry. I did some analysis of why that's not necessarily true and recent analysis shows the problem to be even more complex than I first laid out.

To summarize, in a world of robotic taxis, just like today's taxis, they don't wear out by the year any more, they wear out by the mile (or km.) Taxis in New York last about 5 years and about 250,000 miles, for example. Once cars wear out by the mile, the number of cars you need to build per year is equal to:

Total Vehicle Miles per year
Avg Car Lifetime in Miles

As you can see, the simple equation does not involve how many people share the vehicle at all! As long as the car is used enough that the car isn't junked before it wears out from miles, nothing changes. It's never that simple, however, and some new factors come into play. The actual model is very complex with a lot of parameters -- we don't know enough to make a good prediction.

People travel more in cars.

It's likely that the number of miles people want to travel goes up for a variety of reasons. Robocars make car travel much more pleasant and convenient. Some people might decide to live further from work now that they can work, read, socialize or even sleep on the commute. They might make all sorts of trips more often. Outside of rush hour, they might also be more likely to switch from other modes, such as public transit, and even flying. Consider two places about a 5 hour drive apart -- today flying is going to take just under 3 hours due to all the hassles we've added to flying, even with the improvements robocars make to those hassles. Many might prefer an uninterrupted car ride where they can work, watch videos or sleep.

Vehicles run empty to reposition

Regular taxis have wasted miles between rides. Indeed, a New York taxi has no passenger 38% of the time. Fortunately, robocars will be a lot more efficient than that, since they don't need to cruise around looking for rides. Research suggests a more modest 10% "empty mile" cost, but this will vary from situation to situation. If you need the robotaxi fleet to constantly run empty in the reverse commute direction, it could get worse. Among those who believe robocars will be more personally owned than used as taxis, we often see a story painted of how a household has a car that takes one person to work, and returns home empty to take the 2nd person, and then returns again to take others on daytime errands. This is possible, but pretty inefficient. I think it's far more likely that in the long term, such families will just use other taxi services rather than have their car return home to serve another family member.

Cars last longer

The bottom part of the equation is likely to increase, which reduces the number of cars made. Today, cars are engineered for their expected life-cycle -- 19 years and 190,000 miles in California, for example. Once you know your car is going to have a high duty cycle, you change how you engineer it. In particular, you combine engineering of parts for your new desired life cycle with specific replacement schedules for things that will wear out sooner. You want to avoid junking a car with lots of life in the engine just because the seats are worn out, so you make it easy to replace the seats, and you have the car bring itself to a service center where that's fast and easy.

General Motors has purchased "Cruise," a small self-driving startup in San Francisco. Rumours suggest the price was over one billion dollars. In addition, other rumours have come to me suggesting that at least one other startup has been seeking a new round of funding at that valuation, but did not succeed due to the market downturn.

Reports released reveal that one of Google's Gen-2 vehicles (the Lexus) has a fender-bender (with a bus) with some responsibility assigned to the system. This is the first crash of this type -- all other impacts have been reported as fairly clearly the fault of the other driver.