I think that we are talking days, not weeks, or probably that Teslas can have multiple teams testing various hardware and algorithms at the same time: Tesla plans on producing 500k cars per year from 2018. If all of them have this equipment, and they drive a conservative 10k miles per year in average, Tesla will gather data from 5 billion miles driven after a year. Next year they will have 15 billion miles. After 6 short years that will amount to more than 100 billion (B) miles accumulated.

For comparison, Google's self driving cars have driven less than 2 million miles until now and unless they team up with a car manufacturer or launch their own car, their project is soon irrelevant. I imagine this would only be accentuated as tech talent would prefer to work where their algorithms can be tested in a couple of days rather than where such test would take years.

EDIT: And just imagine how cool it is that since their cars drive all over the globe, in all traffic situations, under most traffic rules, in all conditions, and both winter and summer, day and night at the same time, a team can actually test changes to an algorithm on cars selected for certain conditions (500 cars in Minnesota's winter weather at the same time as 500 in Australia's summer, 500 during daylight in Asia at the same time as testing 500 driving at night in Europe, 1,000 cars in rainy weather, 1,000 cars in sun, 1000 cars in fog, 1,000 cars in snow, all within hours).

I've seen quite a few Google self driving cars on Castro in Mountain View. http://farm5.static.flickr.com/4082/4939372588_47979bceec_z....

That picture doesn't do it justice during the most busy hours of the evening, where there's tons of people crossing and bicycles flying about in every direction.

Given that, I've never seen a crowded street in Napoli or a crossroad in Ho-Chi Minh, but I can imagine they are a completely different level of complex all together.

As noted elsewhere, the announcement was that the necessary hardware will now ship standard. Musk has said elsewhere that the software is many times harder than the hardware.

My three tests for self-driving cars would be:

1. Drive through Milan on a Saturday afternoon, from the outside of the city, and park on the opposite side of the centre of the city.

2. Drive through London during a weekday, similar thing... cross the centre and find a parking space and park.

3. Drive through Paris during a weekend, same deal.

Those three cities are the most stressful to drive through for various reasons.

With Milan it's speed and decisiveness, small spaces and tight navigation.

With London it's the mix of traffic with a high cycling % along with a good mix of motorbikes and heavy goods vehicles as well as a phenomenal number of pedestrians who will walk out at any moment.

With Paris it's the speed of the ring-road (and the challenge of handling variable speed traffic on and near it) as well as the tight spaces on the streets and the numbers of scooter riders.

I also like the diversity of traffic signals and signs across those cities.

When I see a self-driving car manage those cities I am going to be impressed. It will exercise so much of their systems to do any one of those.

PS: I like that they are all fashion capitals... it would be a good marketing campaign to throw in Tokyo and NYC and present an accessible and fashionable angle to a set of super complex technical achievements.

Why don't some of the teams impress me and try driving somewhere with black ice, strong crosswinds and blowing snow. If their systems can't drive in negative environmental conditions, they will be largely worthless in many locales.

In my opinion, the fully autonomous vehicles shouldn't need to handle those situations. Leave that up to the human to navigate.

Fully autonomous should be able to handle the NORMAL day-to-day driving responsibilities. If you're going to require those situations that you mentioned, we'll never have fully autonomous vehicles.

NOTE: when I say fully-autonomous, the car will still have a steering wheel and someone sitting in the driver seat. The google car that had no steering wheel would never happen in the "real-world" only tightly controlled environments.

If we never see a car that can actually drive itself without human intervention, I will be severely disappointed. All that work and hype just for some driving assistance systems?

They may be difficult conditions, but people drive in them all the time. Something like 30-40% of the US population drives in snow and wind for months at a time. They're not uncommon. What's the great utility of a car that can only perform its function in ideal circumstances, i.e. on a closed track.

I'd compare it to having a convertible car---it has one lovely feature, that only really proves it worth during elusive weather conditions.

This will make driving in those conditions even more dangerous. If people start relying on self driving cars, their driving skills will deteriorate. If computer can't drive the car safely, rusty driver won't do better job.

Say Dakar for example (2.5M people): no road signs, pedestrians on the highway, constant honking from every directions, more varied "vehicles" like horse carriage and "car rapides" (small buses).

Having recently returned from Paris, I was completely flabbergasted by the driving patterns. It was surreal. I am excited about Tesla and this self-driving car, I really am. If it can in fact navigate around the city of Paris on a Saturday I'm not sure how I would react. Probably wouldn't believe it.

It knows where it is, it knows where it needs to go and they already have an "aggression" where they'll pull out a bit to signal to other drivers they're going.

The really complex part for us is that we can't easily keep track of all the things around us and have very slow reactions. The computer knows the distance of all things around it accurately and can respond in milliseconds and doesn't get flustered or angry. Or am I missing something, is there a lot of complex reasoning?

I ranked them in order of crazy. Milan is the most crazy of crazy. As are the motorways around Milan. I am pretty sure that the shared philosophy to not use the brake is that if they just go faster they can be ahead of any trouble about to occur.

I'm in Rome, which isn't quite as bad, but yes compared to driving in the US it's completely different. Every day I have to deal with someone cutting me up, or doing something otherwise stupid. You need to drive quite aggressively here, and in a self-driving car I'd imagine the journey would take twice as long, as it would keep stopping to give way to avoid an accident - and drivers would definitely try to take advantage of that.

Urban really means not clearly rural. But there are a lot of "urban areas" 10s of miles outside metropolitan areas that people would consider exurban with forests and homes on significant acreage.

Any folks who grew up on a farm who can confirm?

I wasn't allowed to fell the trees but I had to use a chainsaw and cut the branches off the felled trees and then section the denuded trees into small enough rounds they were suitable for busting into firewood with an axe. I was given the option to learn to drive my grandfather's stick shift Toyota pickup truck in order to move all the wood with that, or move it by hand. It was an easy decision.

But, yes, until you're talking about more serious off-road/bad road (which is challenging in a completely different way), there's nothing especially hard about driving on a private dirt road.

Also driving the shitty back roads in Finland was a lot of fun in my old Honda Civic which didn't have power steering or ABS (let alone any kind of traction control and stabilization you have in newer cars). Driving a modern car with all the modern aids is very easy on dirt. On snow some of them can hinder you though (some automatic gearboxes just don't know what to do when none of the wheels have any traction) but for the most part it is very easy too if you have winter tires.

Driving is very easy. Too easy I think which causes people to drive way too fast for the situation quite often.

Main issue for driving on dirt roads/snow for automated cars I see is just seeing the road itself correctly not with controlling the car.

https://www.youtube.com/watch?v=RjrEQaG5jPM

Edit:

This video from a motorcyle gives a nice first-person perspective:

https://www.youtube.com/watch?v=BqOLehUdqkU

Humans have been known to "follow" the car in front of them when its not appropriate. Being parked on the side of the freeway (especially on a curve) is one of the worst places to be. I recall hearing that people hit those parked cars quite often.

Fun issues to think about.

At the same time, it would beam this data back to HQ and to other cars, telling them to avoid the area of uncertainty.

Except the Google car is in the wild, collecting miles every day (as are the Uber vehicles), while Tesla's miles using this technology are all theoretical.

New hardware needs hours in the field, to verify failure rates and performance under real world conditions. Early deployment means Tesla's hardware will be ready for full service as soon as software is available. Autonomous driving then becomes "just a software problem", with hardware eliminated as a variable.

Does anybody make an antenna-level middle box for GSM phones? Something you hook into the antenna cable as a firewall/monitor?

My favorite part is that the VPN keys were called "carkeys.tar"

They still have access to the car and it's logs, however. Unless you remove the SIM card and disable Wifi, then you are truly on your own.

Besides products sold to governments agencies, the only 2 options I know of to make a GSM middlebox are OpenBTS or the osmocom stack.

You'd probably have better luck getting at that data through onboard diagnostic ports or the like.

1. When the driver has full control of the vehicle

2. When the 'autopilot' is engaged and the driver is ready to intervene if necessary

So if the AI passes the safety test on Type 1 data, Tesla can promote it to being tested on Type 2. And if it passes that safety test it can be promoted to full autonomous control.

The 'autopilot' mode effectively does for Tesla what Google's test drivers do, but for free and on a much larger scale. Seems to me Tesla have a very strong hand here.

So if there's an accident, Tesla can check to see if the autopilot would/could have avoided it. If they can turn round to lawmakers and say that "X% of accidents could be avoided if hands-off autopilot was legal" it should help speed up the regulatory side of things.

"It would have avoided this accident" (by braking, steering). It can say nothing about the future, "... but it would still have been in a collision 0.42 seconds later".

For all the collisions it would have avoided, there's another subset where it would only have "delayed" the collision. But that won't be mentioned. Because it doesn't fit the narrative.

This is why "partial automation" for the initial data collection still produces valid results. You can replay data against updated models and do "what if" testing without actually sending the car back out on the road again.

Eventually, (and with a lot of handwaving) the difference converges towards 0 and the network gives the right answer.

In this case, what you need is mostly "what would most humans do?"

There would be things to refine about that (e.g. prevent speeding; analyse how humans reacted right before crashes etc. and improve on responses), but as a starting point it is immensely useful.

Which is to say, if Google wants self-driving Google Maps vans, then collecting data using Google Maps vans makes sense. But if they want general-purpose self-driving cars, then collecting data using Google Maps vans will only give them a very narrow set of data.

In the OP where I mention outfitting an example 2019 car, that car might be a current-model-year car with roughly similar vehicle dynamics to the proposed 2019 model. The only thing that they are paying particular attention to is the specific sensors in use and their placement on the vehicle. This apparently is critical to development of the driving model, the camera on the test vehicle must be in the precise position and direction and must react in the same way as the production model or the data is bordering on useless.

Now, with a full 3D pointcloud and enough sensor data you might be able to translate one recording into a lower-resolution resampling to model the production version, but I've not heard of anyone doing that. Test data is collected on the production sensor suite, no changes allowed.

Definitely two data sets that should be put together.

I'm assuming that once a Google Maps van has covered an area, it intentionally avoids that area until a significant time later.

And should there be places, terrains, urban conditions etc. that Tesla doesn't cover, they will see it and can hire a number of chauffeurs to drive around there. Even if they hired 1,000 drivers to gather data in certain geographical areas at certain hours or weather conditions that they lacked input from it would be a drop in the ocean compared with the cost other car manufacturers face to keep up with Tesla now.

You can't just install autonomous driving systems in any combustion engine car without changing lots of systems. Tesla is electric fist. Everything is already wired, ready to be measured and in feedback loops.

All I'm saying is that the claim of N million miles should be taken with a grain of salt as long as you do not know the entropy that these miles contain.

Said another way: Is there a downside to parallelizing the collection process across a broader horizontal population of collection vehicles over a shorter period versus the traditional 1-2 year collection period?

From the positive side, they will collect and be able to evaluate a much /wider/ range of driving data from a far larger set. They are much more likely to run across 'better morons' both in operators of the Tesla and in operators of other cars.

I would say that heavy rain is incredibly different from a blizzard. Snow has a much greater horizontal-to-vertical velocity, reflects more light, and flutters around. Rain mostly moves in straight lines, even in high wind. Not to mention traction is completely different on water vs. slush vs. snow vs. ice, as well as with good snow/rain tires vs. worn all-season tires.

I'm sure Tesla knows all this, but just in case you haven't driven in a blizzard before, you will need to drive very, very differently from driving in heavy rain.

Not sure what an autonomous car could have done in such a situation but anticipate that maybe we should preemptively pull over for a bit before it starts raining.

"I'm sorry, Dave, I can't do that. It's raining."

"Are you kidding me?! He's dying! Go! Now!"

"Did you try applying pressure to the wound?"

Let's hope that we never have cars that decide it's not safe for us to drive, "regardless of who's driving."

That's a common misconception, but I believe it was debunked by snopes.

The answer is: If the measurement period for the average over multiple systems is long enough that each random variable has visited a representative sample of states which make up the majority of contributions to the mean value, then the system is ergodic.

[1] https://en.wikipedia.org/wiki/Ergodicity

In the future, would Telsa use these data for other purposes? ie after analysing the data, it finds that my driving behaviour risk and has a higher chance of accident. It could trigger high insurance premiums or restriction on which car I can purchase etc.

What exactly is the problem with punishing people for risky driving? Or rather, let them pay a corresponding premium to cover for the (statistically expected) higher damages they will cause? Keep in mind that driving behaviour is something the user has full control of, in contrast to e.g. disabilities, for which of course no insurance premium should be paid.

I actually think this is a good idea, and would like to hear opposing opinions.

We could determine their liability insurance premium more accurately!

Why don't we film everyone in bed to determine if they use condoms?

We could determine their health insurance premiums more accurately!

Driving like a moron in a 2T vehicle is.

Risky driving isn't a private problem. It's a public problem.

And the government already monitors your on road behavior and punishes you for engaging in risk. Speed cameras and traffic cops exist, in case you didn't know

However, presence of alcohol in sweat or by another means if detected and taken by systems to feed data to insurance companies it will make judicious decision in charging the person a higher premium.

Because, drunken driving risk lives of others on the road and yourself. If a person is not in a position to live on a agreeable consensus of social security model without putting others life at risk, jungle suits him.

https://en.wikipedia.org/wiki/Auto-brewery_syndrome

Also:

http://www.consumerreports.org/cro/car-insurance/credit-scor...

The interesting question is if it might actually be a smart strategic move to announce such a use of data. It's a bit counterintuitive as that would limit your customer pool. However I think there are benefits to limiting your customer pool to more "positive" cases. Let's say Tesla would announce that they'd collect data on reckless driving, warn the drivers and upon ignoring the warnings a couple of times inform insurers about it. That would probably result in less reckless drivers buying their cars which might actually not be horrible since that would also reduce the number of accidents etc. The big issue is of course that it would violate all reasonable privacy laws I can think of (I guess they could make it opt in).

[all assuming non-automated driving]

I don't think that is a good analogy ... many "sicker" people have played no role in their sickness which may not be related to behavior or decisions.

A better analogy would be: Why don't mountain climbers and scuba divers[1] pay higher insurance rates ?

I sort of think that they should, however I don't have a good answer as to what happens when an injured stunt skier is broke (and broken) ... do they get no medical care ?

[1] Or whatever - you get what I mean.

So stuff like a helicopter evac, access to a hyperbaric chamber and the ability to call a number and speak to a doctor who specializes in diving related illness means I already do pay more for insurance than someone who isn't a diver.

Statistics are obviously useful, but they are a big gun that sometimes points towards feet.

With enough data we wouldn't need to just rely on measuring speed. Reaction times, steering behaviour and lots of other driving terms I am not familiar with can all come into play. Cross reference this with actual accidents and you could come up with a fairly accurate model of what makes a safe driver.

You are correct. My problem with that approach is twofold:

1) There is no clear definition of "good enough";

2) The balancing mechanism between improving model accuracy and simplifying the model (for predictability/market competitiveness) involves actors with very unbalanced bargaining power (insurance companies vs very small subsets of customers)

If I can state the problem in another way: When you use a very refined statistical model for insurance premiums you are basically creating dynamic pricing for insurance services. Dynamic pricing hurts competition in the market through the development of an exaggerated bargaining power by one side of the market. Consumers would be worse off. What they gain in justice of their pricing (I get lower premiums because I'm a safe driver), they lose in bargaining power (via fragmentation of the consumer market).

We we don't have are any legal or moral frameworks for passive, always-on methods of observation. Until we do, we shouldn't be letting private companies arbitrarily determine whether a driver is dangerous or not.

https://www.motorists.org/issues/red-light-cameras/increase-...

I believe we have established red light cameras are the cause of danger in a large number of circumstances, with yellow light time being reduced to increase government revenue.

Not to mention the lack of due process.

I don't think the government should rely on private companies algorithms to determine danger or guilt.

Otherwise, we should punish you with a speeding ticket, because you will sometime in the future with some probability exceed a speed limit. Would that be ok for you?

The realistic one for me on this is when they show aggregate data to insurers to get lower premiums for their cars (or indeed provide the insurance themselves at low rates and use the data to get excellent reinsurance rates). That seems a very tesla thing to do.

I think a combination of LIDAR and cameras is probably going to be required for a truly autonomous car, at least for the first decade or two.

https://www.tesla.com/nl_NL/videos/full-self-driving-hardwar...

The video clearly shows a fully autonomous vehicle in a range of driving conditions. How about accepting that the sensor suite is adequate?

The video shows a fully autonomous vehicle in a range of (relatively) easy driving conditions.

1. There is not a single "sticky" situation in that video. Handling 90% of situations is (relatively) easy; it's the last 1% that is extremely hard.

2. It's not clear how many tries it took to record that video, but likely more than one.

3. At the end of the video, the car drives in the opposite lane of where it's supposed to. That smells of a highly-staged demo, not a general purpose solution.

In short, this video in no way proves that the sensor suite is anywhere close to being sufficient. Can that car repeat the same trip at night? In rain? Can it drive down Castro St. in Mountain View or University St. in Palo Alto at lunch time? Etc. etc.

The reasons are multiple. 1) building and upgrading a car model take them at least 2 years... in the meantime, they have time to clock the kilometers anyway.

2) They have no idea how to build that type of data "massive" environment. They just don't know. Like all our "connected car" data came onto one single thread acceptor on one single machine. No load balancing, no scaling nothing, until August when the vacations congestion overflooded it. So now we have 4 threads in that single machine. I could ramble on and on on this type of mistakes they do. They just don't know. And this year, they asked the team to reduce their budget. Because it is how the car industry works. After two years to kickoff the project, you have to cut your cost by 5% to 10% every year.

However, this would mean adding sensors and 12 years ago, these would have cost much more along with the data requirements.

After that a car is pretty much seen as completed, it will go into the factory and to the customer and probably get minor updates and fixes during the first year. However development already starts for the next model. Data and information from previous cars won't help you that much, since the systems in the new car will often be very different (as I said, most things are developed externally and from different suppliers, reuse is limited).

I wonder what safeguards the manufacturers have in place to determine that those tests are not gamed and are honest as presented. I am thinking of VW emissions scandal. Given the stakes it would be interesting to know what level of trust is involved vs. verify.

This move sounds like a terrible idea with respect to what you say!

It's like aerospace companies deciding to forego testing and calling that a "brilliant move!"