Winner-take-all effects in autonomous cars
ben-evans.com
ben-evans.com
Today we have left and right blinkers, brake lights, the horn, and some 'body' language when you use the whole vehicle to show your intent. But with fully autonomous cars, we could be giving out signals that are rich with information.
"My exit is soon, I'd like to change to the right hand lane everybody!"; "People up ahead of me seem to be stopping suddenly, watch out as I may need to do the same!"; "The right lane is blocked at 43.718045, -79.507663, pass the word back to everyone so they are prepared".
If we can build an open protocol for how cars can share this rich information, then as more and more cars are autonomous, the roads will improve continuously. Even non-autonomous cars could start listening to this data and sharing it with the drivers.
On the other hand, if we don't build an open protocol then we could start to see, for example, all the Waymo cars helping each other out- but not helping the Teslas or the GMs. There's a winner-take-all effect in action.
While reading this article, I had a paranoid convergence of ideas.
I was reading another article the other day about stock trading strategies, and how some particular strategies only work if a small number of people are using them; once they go mainstream, they lose their advantage.
The example used was a whole lot of funds jumping into a particular strategy, which then effectively killed the strategy's effectiveness, and everybody lost.
My paranoid fear is something similar in autonomous vehicles; a vicious cycle of systems reacting to each other, because they're not able to communicate directly.
Like when you're walking towards someone, and you both try to step to the side to let the other pass, but you step to the same side; then you correct by stepping the other way, but so does the other person.
Also if anyone's interested, the article I was referring to is this: http://www.mauldineconomics.com/the-10th-man/black-monday
It's way too much data to send over LTE, so you need to have a bunch of disk space to store it, and then you need your users to be OK with you abusing their WiFi in the middle of the night to send gigabytes of data back, which they won't, because we have data caps.
Whereas if you operate a fleet, you can just connect a data transfer cable to the car every night and dump all the data.
Obviously you can compress the data; maybe throw away the vision and point cloud data, and just keep the object locations, but is control with already recognised objects the real problem here, or is it the actual recognition part?
Related... it will be interesting to see how well autonomous cars handle mechanical failures as they (and their tires) age. Good drivers really distinguish themselves in the moments where the inputs and behaviors that are expected do not occur (ex: tire blow out).
I do think it will figured out in some future where there's true autonomy and they're proven far safer statistically. But you're right that the general public isn't going to shrug when $AUTOMAKER's car runs over a skateboarder and the company says "We are committed to safety and we'll try harder next time."
There's also the uncanny valley when no one's paying attention although they're legally supposed to be.
Unfortunately many examples from recent years (acceleration pedals, air bags, tires) show that after few weeks everybody will forget about the issue and move on.
Autonomous cars will be much better at handling these types of situations - in your example, the machine will know what's going on few milliseconds after it happens and will be able to adjust speed and direction almost instantaneously. All you need is lots of sensors.
Which rapidly reduces the MTBF at which point you're stuck with an overpriced manual drive vehicle.
What will make autonomous cars better is that they can take all of the input from thousands of blowouts, add "training" from actual experts, and load the resulting techniques into every car on the road, ready for execution the instant something happens.
There are two things where human beings far outpace AI in competence. The first is the ability to recognize when inputs just simply don't make sense and thus the intended algorithm is going to return garbage results. A related, but important issue (particularly if you talk about networking cars together) is being able to recognize malicious inputs.
The second issue is that humans are good at higher-order planning. An example as it relates to driving is what we do when we can't tell the state of the traffic light (direct sunlight, say). Humans naturally look for ways to indirectly infer the state: the status of the traffic light in the cross direction, the current disposition of cars around the intersection.
Autonomous cars will only be better at handling these situations if the engineers are aware that these situations need to be handled and take measures to make the cars aware of them. And history suggests that is sufficiently unlikely to happen to make me fearful of the future.
And yes, we can deduct light color from other clues, but we can also go through a red light because we were preoccupied with our cellphones. AV will on the other hand have access to multiple cameras, looking at the light from multiple angles, possibly with some polarizing filters that will make this non-issue.
I wouldn't be surprised if Tesla (and other autonomous driving providers) directly offer customers their own insurance and liability coverage. The auto insurance companies of the world see the writing on the wall, yet have no data to accurately model or price such a product. It might be a good thing too, since it providers a second layer of incentive to improve the underlying autonomy technology.
I keep hearing, and used to assume, this. But observe public reaction to e.g. the Tesla autopilot accident: measured and muted.
Counterbalancing negative reactions to machines are generational shifts in our perception of technology. I haven't seen a study on this, but from memory, what was the last new (or, for franchises, first-in-franchise) film that portrayed a man-vs-machine war? The closer we get to a population of people mostly born after the mid-1980s, the better the reaction appears to likely be.
There's quite a few recent rogue AI movies, but they typically have a few characters stopping the machine rather than a war. Avengers 2, I Robot, Eagle Eye.
Of course. If somebody is a much-better-than-average driver, then suddenly their probability of getting killed is going to be equalized with the rest of the population and may become higher as a result.
The most likely outcome in the near term is that self-driving is an auto accessory you can order for a car. That's what the car-makers have in mind, and most of them are talking about shipping it for the 2020-2021 model year. In that business model, it's not a big money-maker for whoever sells the self-driving technology. They get to be a Tier I auto parts supplier, squeezed hard on price.
There's a lot of noise about "transportation as a service". Remember that Uber loses money, even after squeezing its drivers really hard and making them buy and maintain the cars. Actual "transportation as a service" is going to look more like car rental and Zipcar, and will probably come from companies that already operate fleets of cars. They already have "rent by the hour, day, week, or month". They're just adding "rent by the trip". Also, bear in mind that "transportation as a service" requires level 5 autonomy - no driver required - so the vehicles can move around empty to pick up a passenger. That's going to come a few years after Level 4.
Autonomous cars don't require car to car communications. They don't need to be "connected" at all. Chris Urmson pointed that out when he was heading Google's self-driving effort. From a safety perspective, it might be better if they're not connected. The enthusiasm for car to car communications is coming from the infotainment and advertising people.
For examples: https://www.theguardian.com/technology/2016/aug/18/self-driv....
They have also been used in DC, Vegas, Perth, the Netherlands and other places.
It might not be a total replacement for cars but they are self-driving vehicles that are moving passengers now.
If I had access to these running at 30km/h only on major streets in my city I'd probably get rid of my car.
None of these systems seem capable of an ordinary airport parking shuttle route. That's embarrassing. Those things should be performing better.
Uber loses money because it pays the drivers - if you eliminate the driver you eliminate most of the cost. And if you can use simpler electric vehicles which are cheaper to drive and maintain, then you have yet another significant advantage.
I don't think anybody knows what exactly will happen, but there is clearly an enormous opportunity. The trucking industry alone in the United States is nearly a trillion dollar market. Of course when you add in all the complexities of actually scaling the business it doesn't become pure profit, but I'd be amazed if you couldn't print money if the technology worked.
For example, if a driver can rest while such car travels the highway(or a traffic jam), he might agree to a 10% pay cut(Via pays decently today, unlike Uber), and that would be enough to win.
That's a level-4 service with strong network effects and thus nice margins, with time.
As for how that translates to advantage in level-5, that's a complicated question, but a few ideas:
1. Brand and technology advantage due to much more miles.
2. Sharing. Maybe some share of the global transportation demand would continue to be shared.
3. Politics. On-demand shared transportation solves moat of the slowness of public tranaport. Maybe we could also solves the experience part(and easier for 3 person shring than a bus). At that context, dedicating special road lanes for sharing may become politically realistic, and than cars lose on speed.
Same goes for rush fees.
Or maybe, just owning the customer order layer, would be enough to win on optimization of single person cars moving around a city, giving a 10% advantage in trip time , and that would be enough , with time, for a mono/duo-poly.
But even if not and it's a low margin business, there would still be a ton of money to be made.
It's reasonable to assume that these standards will get more stringent over time as accidents, insurance, and compatibility become forced issues (likely though lawsuits). The companies that have the data and systems to meet these standards will be able to keep up but it'd be hard (read: impossible) for a new entrant to catch up after the initial players are further down the evolutionary curve.
There's a term for this twin to regulatory capture. In essence, the amount of regulation in an industry scales with the age and development of the industry. Difficulty of entrance scales scales with regulation. Combine this with any economies of scale and network effects present, and you have a state-subsidised incumbency bias.
An outcome could well be that both Tesla and Google both meet the high government standards, but in doing so they've had to create such high performing autonomous systems that the consumer can't tell the difference. They both work equally perfectly because they have to be to be allowed on the road.
If this is the case then the autonomous system becomes a non-factor in the purchase of a car. Either you're in the market for an autonomous car or you're not.
Imagine trying to calibrate your third-party self-driving software for every different car geometry, sensor placement, transmission system, and steering sensitivity across a few different car manufacturers. If you screw it up for one car model, people die and it's your fault.
Now let's say waymo chooses a sole partner to avoid this, GM, for example. Now waymo and GM have to work together to build a car, with a huge expertise barrier between the two teams. Will GM be willing to redesign their decades-old electronics platform around Waymo's self-driving system? Will Waymo put in the effort to make their sensors easily maintainable by GM mechanics? If the relationship doesn't last forever, much of this effort is wasted.
Tesla has expertise on both sides and has a mandate across the organization to make the integrated system work. Not to mention a huge head start handling UX and safety regulations in the field. They may not have a monopoly, but I expect them to be the first to level 5 and the leader in autonomous driving market share for years to come.
I think it depends a lot on the specific partner. Noticeably WayMo has partnered with Chrysler, rather than Ford, and Chrysler seems in a position where they would retool to do what WayMo told them to do, because they have no self-driving effort of their own, so WayMo is their only hope if this becomes real. WayMo might just buy Chrysler at that point.
Self-driving is such important (just after eliminating combustion engines) that we could upgrade existing cars with cheap ROMB boxes. Vehicle GPS tracking system costs about $30. ROMB box would cost about $60.
Let's say a car ROMB received info about the white truck, while your car cameras and vision recognition systems see just a cloud and any truck in 50 m range, that conflicting info should cause the car to slow down.
The infrastructure would need to be designed such that large scale outages are unlikely I guess.
The problem is who pays for it, and who authorizes its installation. It's really something that has to be done at the governmental level. And I don't think there's any will to pay for it.
There are other issues of course. One is that in an Urban environment transportation issues would be better served by good public transport than self-driving vehicles.
The other is that I think once the government get involved, the numerous sociological issues associated with self-driving cars (which I personally think are more significant than the technical issues) would become apparent.
A lot probably comes down to the size of the moat. If some fairly modest dataset is "good enough" that implies needed data will be widely available can be used by anyone for their automation algorithms. (Regulation could also force some level of standardization at this layer.)
Alternatively, truly vast data sets could turn out to be the difference between decent assistive driving systems, autonomy on highways, and more broadly useful and enabling self-driving technology. Such data sets could end up being outside the capabilities of a few companies who got there first.
Given that humans do without detailed data of this sort, I think it should be okay to leave manufacturers and service providers to figure out how proprietary or standardized the datasets should be.
I personally might never trust/desire an autonomous car, especially one connected to the internet, a phone network, or a car-to-car network. In a world where everyone has your best interests at heart, it seems like a nice feature; in the one world we've got, it doesn't.
Any specific suspicious death may come and go, but the fact that something like this is believable with conventional networked cars is enough reason to be very skeptical of security and government access (and criminal access) claims of manufacturers, no matter how well-meaning.
Ultimately, I am alive because nobody cares enough to murder me.
I don't see what's wrong with having fewer things in your life which can silently kill you.
If you get in a new car, you are now completely at its mercy. There are few or no precautions you can take; you might as well lay on your back, and strap your arms down so your belly stays up.
What you're really doing is making it cheaper and easier to kill, misdirect, or trick you. This means that people who wanted to do these things before, but couldn't afford it, now can.
Humans learn by using data from their interactions with the environment, since childhood. That's a ton of data and there might not be a way around that.
Example: Driving with cruise control on at 60 mph. The car in front is going ~59.5 mph, so you slowly gain on them. Why do the cars not communicate their speed, so your car can adjust speed to maintain a safe distance? Even cars of the same brand can't do this today. What will be the catalyst for this changing?
I think more interesting would be smarter traffic control systems which communicate with cars in cities to advise traffic flows - a simple example being a (wirelessly) counting down light rather than red/amber/green.
How do you deal with spoofing info for/from another vehicle?
What happens when someone decides to see what happens when they `inter_car_radio.broadcast(myspeed=0)` or `inter_car_radio.broadcast(myspeed=120)` when the speed limit is 80?
- a signed (regulator-mandated) signal for cars from other manufacturers;
- an encrypted signal intended for cars from the same manufacturer.
The first signal would be signed by the root CA of the car maker to indicate that its communications can be trusted.
The encrypted signal would of course carry an information-rich payload for coordinating a convoy to travel at the same speed, brake, signal a turn, etc simultaneously, for instance.
From full-autonomy onwards, it is very hard to predict anything. There will still be manufacturers of course, but the manufacturer-consumer structure of the market could completely change. I think there's a decent chance personal ownership will decline.
Removing drivers from ride-services will drive down costs, a lot. Probably enough to change ownership incentives and market dynamics.
If that happens, 2nd order effects will also be big. People act very differently when they pay per-km explicitely, even at reduced rates. Tolls (economists love tolls, btw) become more palatable when rolled into a per km price. There might be fleets of little, light cars servicing city centres.
In this chaotic context, I think it's hard to predict anything about market structure.
The 3+1 companies of sufficient scale to dominate the market are: Google (Android), HERE (was Navtek, was Nokia), Tom Tom (Apple iOS), and Inrix (startup) - which does not create its own maps, but has captured enough driving data to compete with the Big 3.
The only other major map provider is Microsoft Bing, but it does not (AFAIK) have significant products based on driving data, although it does have 'T-Drive' and various other research efforts. Losing the battle for mobile OSes has crippled its efforts.
Patents are effectively a grant of monopoly.
Also strong vertical integration can still ruin the market.