The Tesla Self-Driving Advantage – 1.3 Billion Miles Of Real-World Data

2 months ago by Eric Loveday 51

Autopilot Data - Image Via Mitch Turck

Autopilot Data – Image Via Mitch Turck

Bloomberg’s Dana Hull recently posted an article titled “The Tesla Advantage: 1.3 Billion Miles of Data.

The article explains how Tesla’s real-world, on-road autopilot data places it leaps and bounds ahead of any other automaker in the autonomous driving segment.

Tesla Logo Camera

Tesla Logo Camera

Hull writes:

“There was, in hindsight, a clear element of risk to Tesla Motors Inc.’s decision to install Autopilot hardware in every car coming off the production line since October 2014. It paid a price, with federal regulators probing the deadly crash of a Model S while in driver-assist mode and critics slamming Tesla for rolling the technology out too soon.”

“But there was also a reward. The company has collected more than 1.3 billion miles of data from Autopilot-equipped vehicles operating under diverse road and weather conditions around the world. And in the frantic race to roll out the first fully functional autonomous vehicle, that kind of mass, real-world intelligence can be invaluable.”

It’s not just Hull who believes Tesla is leading the way in self driving. Nidhi Kalra, a senior information scientist at the Rand Corporation, stated there was no question Tesla has a big advantage, noting Tesla can learn at a much more accelerated level with an open fleet vs a limited amount of employees and trained drivers testing the systems.

Tesla software update 8.0 - Autopilot Enhancements

Tesla software update 8.0 – Autopilot Enhancements

Adam Jonas, analyst at Morgan Stanley for autos and shared mobility, chimed in too, stating that Tesla is “in a very unique position to push the state of the art of algorithmic driving and machine learning in personal transport.”

And even Richard Wallace, director of transportation systems analysis at the Center for Automotive Research in Ann Arbor, Michigan, had some positive notes to add:

“Whether they are ahead or not, Tesla certainly has tons of data. They will be able to analyze that six ways from Sunday and continue to tweak their algorithms.”

Most automakers, aside from Tesla, keep secret their autonomous driving programs, to some extent at least. Few automakers are willing to reveal how far along these programs are, but not Tesla who is openly testing autonomous driving with Tesla owners out in the real world.

autopilot-2-via-bjorn-nylundBack to Dana Hull, the author of the Bloomberg article. She states:

“Of course, not all miles are created equal: there are semi-autonomous as well as fully self-driving ones, real-world vs. simulated, highway vs. those racked up in tricky urban environments.”

“The 1.3 billion miles of data Tesla said it has collected represents those covered by its vehicles even when Autopilot isn’t switched on—it operates in “shadow mode,” with sensors tracking real-world data when it’s off.  In October, Musk said on Twitter that the number of cumulative Autopilot-on miles was 222 million.”

All of this would seem to put Tesla so far out in front that it’s unlikely some other automaker can catch up. Fortunately, it seems Tesla is open to sharing this data so that other automaker can benefit from this pioneering, on-road research project of sorts. Isn’t that nice of Tesla?

Source: Bloomberg

Tags: , ,

52 responses to "The Tesla Self-Driving Advantage – 1.3 Billion Miles Of Real-World Data"

  1. FFE 1 says:

    I am blown away by the technology every time I drive my S – what a world we live in. Things have changed a bit since my first car a 1960 Ford Falcon that I got used 40 years ago.

  2. Mark B Spiegel says:

    Here are some questions from someone who– as you know– is short TSLA and you have blocked on Twitter. Does that invalidate these questions?

    1) Why (according to Electrek) does the 2.0 manual say that the system needs hundreds of millions of miles of additional data before it can even provide the amount of driver assist of the previous system? The original system didn’t have close to that before Tesla made it public, so why aren’t the alleged 1.3 billion miles enough?

    2) What kind of data does Tesla have? (It’s the data itself that would be proprietary, so it should be willing to answer that.)

    3) How much data was really transmitted over expensive, bandwidth-limited cellular connections?

    4) Why for several of Tesla’s accidents did it claim it didn’t know if Autopilot was on because “the antenna was damaged” if data transmission would have occurred up until the point of impact?

    5) Why did Tesla need—post accident—to send inspectors out to study the “black boxes” if it was receiving so much data?

    6) How usable is whatever data Tesla did accumulate if it was all done for the previous Mobileye-based system?

    1. FFE 1 says:

      Hmm – are these questions addressed to me or the page in general? – I have no idea what you are talking about with the twitter stuff.
      You sound angry. I know it must be frustrating for you as one who shorts the stock and is a hater as the company. As Tesla distances itself from the competition and gets stronger everyday with every car sold it will be very hard for the late comers to catch up especially for autonomous driving. it sounds like they may be willing to share the technology as they did with the other patents they have released.

      1. Jay Cole says:

        InsideEVs Twitter feed has him blocked…but to be fair, if you have read his feed one would know why from our prospective. Dozens of tweets every day, most all of which self-serving a single position, delivered in a mostly juvenile fashion.

        Not going to lie, it’ll be a nice day when Mark clears in his TSLA position and moves on to another one…I don’t want to say what he would do for sure, but I suspect the frequency he mentions Tesla would drop astronomically (if he would mention them at all), in exchange for hourly diatribes on whatever the new pick is.

        I do have to give him credit for persistence, pretty sure Mark has been short now for ~4 years on TSLA, and he keeps trucking, I can appreciate that from a tenacious business point of view…but at some point it is just hard to listen to, so we don’t anymore…as is our right.

        As long as he discloses his position here and keeps it civil/on the level, he is still welcome to participate.

        1. ffbj says:

          Yeah, no amount of logical thought could explain why a person proven so wrong so often would continue to persist in a false hope/belief.

          I think it represents an element of the human psyche which harbors this trait which is useful in many cases. Shackleford trudging on against all odds, to save his men, but here it’s more akin to Scott who despite vast preparation failed to return from the South Pole and lost himself and his men, who refused to leave him as he lay dying, and died heroically too.

          1. Jim Seko says:

            I have a hunch he’s not shorting TSLA stock but is actually a paid troll for big oil

      2. Mark B Spiegel says:

        The questions can probably only be answered by Tesla itself, and they seem so obvious that it’s amazing to me that the media covering the company (Hull, InsideEvs, etc.) simply takes “1.3 billion miles” at face value without asking them. Heck, if I owned the stock I’d want them answered too (unless I were afraid of the answers), and if I didn’t own the stock I’d want them answered purely out of intellectual curiosity.

        Yes, I do hate deception (which I believe Elon Musk perpetuates and have documented multiple times), but I certainly not the concept of EVs in general– if I were an auto commuter with access to a plug (I’m neither) I wouldn’t hesitate to buy a Bolt (if I were willing to pay nearly 2x the price of, say, a Civic).

        Re. Twitter, InsideEVs has blocked me from following it there.

        1. Mark B Spiegel says:

          Should I take that to mean you can’t answer any of those simple questions either, Jay, or are you afraid of what the answers might be?

          I Tweet so much EXACTLY because of unquestioning articles like this one, first from Bloomberg and then from you guys. If other people demanded answers about this stuff from Tesla, I wouldn’t take it upon myself to do so!

          Regards,
          Mark

          1. ffbj says:

            Tilting at windmills.

          2. Nix says:

            Bullpucky.

            You tweet about Tesla because you are short TSLA and you want to sway public opinion against Tesla.

            You are a Concern Troll. (google it)

            We know you don’t actually personally care about anything you post, because you don’t actually own a Tesla and have no desire to ever own a Tesla. Yet you pretend to be Concerned! Oh so Concerned! about Tesla owners!!

            How stupid do you think we are? As stupid as you?

          3. Darth Wader says:

            Hey Mark,

            Just show up at CES or a motor show and confront Elon directly.
            How hard can that be?

            1. Anon says:

              Do you think it’s wise to encourage the pathological to get anywhere near Elon?

        2. Nix says:

          Mark, crying about being blocked on twitter? Really? What are you, a 10 year old girl?

          You got blocked. Suck it up. Who publicly whines and cries about being blocked on a social media site? Nobody owes you the right to post whatever, where ever you want in social media.

          Grow up.

        3. Pushmi-Pullyu says:

          Mark B Spiegel said:

          “Yes, I do hate deception…”

          That’s not very convincing, coming from a prolific professional FUDster. In fact, one might suggest it’s one of the most hypocritical things anyone has ever said.

          I agree with Jay here, Mark, and perhaps I’ll go even farther: Your persistence in continuing to very publicly advocate short-selling TSLA stock year after year, despite the almost mathematical certainty that you must have repeatedly lost substantial amounts of money on such an investment, suggests some sort of obsessive dementia on your part. But even if you do suffer from dementia, that doesn’t excuse your prolific campaign of repeated Big Lies.

          1. Nix says:

            I’m sure he has some lame list of his own Seeking Alpha posts to “prove” his claims…

            Because nothing proves you are right, like posting a link to your own Seeking Alpha story saying you are right.

            *laugh*

    2. Marsal G says:

      1 – different system, different hardware, different database. No conspiracy theories needed.

      2 – Perhaps you could read the release notes available online. But fear trolling is more fun, huh?

      3 – Are you crying because they aren’t sharing proprietary secrets?

      4 – Seriously?

      5 – Because they’re not as good as your buddy Needle**** meir and can file an official NTSC report based on pictures of a crash.

      6 – Oh crap, you’re right! Let me sell all my TSLA! You convinced me buddy!

      How’s your short position doing btw?

      1. Marsal G says:

        NHTSC

      2. FFE 1 says:

        Thanks Marsal

      3. Mark B Spiegel says:

        So 5 non-answers (as I suspected) and 1 (#1) that seems to imply that the alleged “1.3 billion miles” are indeed useless. Okay, thanks, and best of luck with your investments!

        1. FFE 1 says:

          so, to be clear, No car, no commuting, no plug = Troll

          I love the Civic comment. You do realize you can buy a BEV for less than the cost of a Civic after incentives right?

        2. ffbj says:

          Coming from you that’s more like a curse.

        3. Nix says:

          Mark cryBaby Spiegel said: “best of luck with your investments!”

          So funny you brought up investments.

          By the sheer volume of your whining and crybaby antics, I see the 50 point bounce in TSLA shares since December has your investments circling down the drain.

          How much money are you down the hole after TSLA went up by 50 that you are back here making a fool of yourself, with the stink of desperation dripping from every one of your posts?

          1. Nix says:

            Talking about investments, I can’t help but notice that your fund you troll for hasn’t even been able to keep up with the DOW in the last quarter.

            Even more embarrassing to you, is that somebody who invested $10,000 in YOUR fund at inception, would be down a full order of magnitude compared to an investor who put the same $10,000 into TSLA shares back in 2011 when you started your hedge fund.

            Man, I feel bad for your investors! They could have another zero in their holdings, if they hadn’t invested with a zero like you!

        4. Marshal G says:

          Speaking of non answers, I asked how your short position was doing, and I noticed you didn’t answer. I suppose trolling this website is a hint of how it’s going. I can ride the waves of bad press followed by good press, since I’m a long in a company that is going to grow at a ridiculous 50% yoy for the foreseeable future. You are basically treading water, hoping reality catches up to your hateful delusions and evil liberal conspiracy theories, but after the M3 comes out your fund is toast. I look forward to you being wiped out.

    3. Nix says:

      1) You are a trolling idiot who makes a fool of yourself on a regular basis, so you don’t know the answer because you intentionally don’t want to know.

      2) The rules and regulations for what data can and cannot be stored in automotive Event Data Recorders can be found here:

      https://www.law.cornell.edu/cfr/text/49/part-563

      The fact that you are whining just about Tesla, when 96% of new cars (including whatever you likely drive) all contain EDR’s. All of the details of the data collected is specified in that link, down to how often the data is collected, and how much is stored. Maybe if you spent 2 seconds breathing in between your financially motivated rants against Tesla, you would have time to actually research the target of your obsession.

      You are such an expert on Tesla, that you have to be schooled by the posters on every green car website you troll. Are you still reeling from your epic take-down you suffered over on GCC?

      3,4, and 5) The answer to all 3 questions is Asynchronous Batch Processing. Duh. Data is sent in batches after being collected and buffered, then compressed. All collected per the regulations I posted, and stored in the Data Formats specified by law, and buffered per the law, and available for data retrieval using a standardized tool, per the same law.

      Finally, if you want to know what data is stored, that “information collected belongs to the owner” of the car, regardless of who makes the car. So nobody is hiding anything like you imply, but since you don’t own a Tesla, it’s not your damn data to view. Too effing bad for you, Mark cryBaby Spiegel.

      http://www.ncsl.org/research/telecommunications-and-information-technology/privacy-of-data-from-event-data-recorders.aspx

      Now shuffle along little loser. Mr. supposed expert on Tesla’s, who gets his head handed to him every time he trolls green car forums, by people who actually do know what they are talking about.

    4. Trollnonymous says:

      What w whiny little B|tc4!!!!

    5. AlphaEdge says:

      Mark, almost all of us here are pro-EV’s and pro-Tesla, and for good reason.

      To constantly attack Tesla, cause you are trying to gain financially, is just lame.

      I just read your Twitter account. I’m all for dissenting opinions, and usually many dissenters have something positive to contribute, as they allow us to see the flaws we may not see, because we are enamored with the company or technology or we fail to think about the negatives.

      But you take the cake. Really do yourself a favor, and try to spend your time doing something else more productive to your own happiness.

    6. Nix says:

      Mark trolled:

      “1) Why (according to Electrek) does….”

      hmmmmm Why not go over to Electrek and ask about their story on their site if you have questions about what is going on according to Electrek?

      Oh, right. You can’t go over there and ask, because you are regularly ridiculed so badly for your misinformation over there, that you are a laughing stock. Now you come slumming over here to discuss stuff from electrek.

      Too bad you’ve become a laughing stock over here too. You are burning your bridges one site after another.

    7. Pinewold says:

      I will try to answer
      1) deep learning needs both positive and negative data for each input. If you add more cameras, you need to have 10,000 negative examples for each behavior you want (avoid, brake, accelerate…) for each camera. Each time a human corrects the computer, the computer learns. To make sure the computer learns the right lesson, (instead of a human bad driving habit of one individual) you want enough corrections to be confident the humans are right.

      2) we know Tesla has radar data, camera data, telemetry data(speed location, gas pedal / brake…) as well as stationary and moving objects. Based on all that data, there is “plan” of what to do and most interesting differences between what computer plan iand what the human did. Any time there is a big difference you want all of the inputs (radar, camera ….) so you can figure out who was right.
      3) , 4), 5)Tesla has wifi and 3G. In an accident, the car probable tries to upload something like the last 60 seconds of data using 3G. (Not a lot of data but enough to see the accident video and get telemetry data). Car knows accident based on sudden stop.

      If car can drive home normally, wifi is often available at home so daily corrections and all inputs for corrections can be uploaded over wifi.

      Tesla is probably not sending much data all the time. Map data can be downloaded over 3G as needed and music can be streamed, but video and telemetry are not constantly uploaded.

      5)If the antenna is broken it cannot be sent over 3G so all data is preserved, but must be retrieved by removing the data logging black box from the car. So you need to send a technician to get the data.

      6) Clearly less than 100% of the data is reusable since they had to turn everything off until AP2.0 can learn enough. on the other hand much of the code in the original system captured algorithms and data structures That can be reused. Knowing engineers it is highly likely there was at least a major rewrite of the code, but even that would be based on what they learned from their experience with MobileEye. So maybe the motion estimator needs to be tuned for new cameras, but knowing you need a motion estimator and knowing how well the current motion estimator works is not lost. This detailed real world data driving knowledge is where Tesla is miles ahead of everybody else.

      1. floydboy says:

        Really appreciate the answers Pinewold, but Mr Speigel doesn’t really appear to be interested in factual affirmations to the questions he posed.

        He instead appears to be doing this exercise to fulfill some sort of need to justify his anti-Tesla feelings and by extension, his financial position in that regard. It could be that an early belief that a company operating in a ‘high fail’ environment and in an unorthodox manner, would collapse and justify his skepticism, thus rewarding his financial position.

        Since it’s rather apparent that is NOT likely to happen now, he appears to be unwilling to modify, or even moderate his positions. The result is, now as that skepticism and financial position is very UNLIKELY to be rewarded, he resorts to more ‘colorful’ antics.

        It seems to be more self-affirmation than anything else, as I’m sure he knows his machinations aren’t the least likely to sway Tesla supporters.

      2. DangerHV says:

        Thank you Pinewold! Although Mark’s questions may be accusatory, you have made lemonade out of the lemon.
        Personally I have been curious to know more about the sensor data operation and you have provided very useful insight.
        I often stop reading the comment sections when all that is posted is bashing back and forth with no value to any of us. I’m glad I continued today. Thanks again!

        1. DangerHV says:

          Re. sensors, can anyone confirm whether or not Tesla’s have a physical pressure sensor on the accelerator pedal, as opposed to a position sensor? I need to know this to fully accept that the unintended acceleration claims, refuted by Tesla, are 100% driver error. On my 79′ Impala the pedal would get “pulled” down to the floor while using cruise control due to engine hesitation problem.
          Having only a position sensor does not satisfy my concern.

          1. Trey M says:

            I do not know the answer to your question, but there is a huge difference between your ’79 and anything else modern on the road today. Your example was the cruise system actuating the pedal for intended acceleration as the cruise system recognized the speed was below set and needed more power to get there. That vehicle had 2 cables to the carb, one from your pedal and one from the cruise actuator (either vaccum or electical operation). Today’s car is drive by wire with no physical connection between the pedal and the engine (in ICE vehicles). In an EV, you are simply telling the computer what your desired input is and the computer calculates the power to send to the motor. The pedal does not move or react to the cruise signal at all.

          2. Nix says:

            It uses 2 redundant Hall sensors to sense pedal position. All drive-by-wire pedals are mandated to have redundant sensors by DOT regulation. This is true for both ICE and EV’s.

            In addition, Tesla has a semi Drive-by-Wire braking system where the computer can apply the brakes as well as the driver directly activating the hydraulic brakes through a purely physical connection, so it also has 2 redundant brake pedal position sensors.

            You can follow my link above and read through the regulations for how EDR’s record and store the data for each of those 4 sensors, which is standardized by law across all car makers.

            All 4 logged data items will show not only that the accelerator pedal was depressed to the 100% position, but that the brake pedal was not depressed at all.

            For there to be a logging or sensor failure that caused unintended sudden acceleration, all 4 sensors would simultaneously have to all fail at the exact same instant. Then once the unintended acceleration was over, all instantly go back to working at the same time.

            This is nothing like your vacuum activated cruise control that is physically connected to your accelerator cable. I’m afraid if that is your only frame of reference, that the gulf between that old technology and the technology car makers have been using for the last 20+ years may be difficult to bridge using just that old technology as a frame of reference.

          3. DangerHV says:

            Thank you very much for the detailed info Trey M. and Nix!! The key is:
            “The pedal does not move or react to the cruise signal at all.” (Yes, I own a 20 year old car, cc doesn’t work at all)
            This takes away my concern that an electro-mechanical system malfunction could give an accurate reading from the position sensors, but there was no pressure exerted on the pedal.

      3. Doggydogworld says:

        Great reply, Pinewold. A couple of thoughts.

        1. “Each time a human corrects the computer, the computer learns”. This sounds like each in-car computer learns on its own. Isn’t it more accurate to say that when the human driver and computer disagree, the computer saves the prior X seconds of sensor inputs and sends them (later) to HQ? Engineers at HQ then filter all these events and feed them into the neural net s/w, which learns from them and updates it’s weighting, feedback coefficients, etc. The engineers then send the updated software out to all cars (possibly in stages, for verification purposes).

        6. Vision primitives such as motion estimation, edge detection, etc. are embedded in MobilEye’s chip. NVidia has their own algorithms (in s/w, mostly). NVidia (and probably MobilEye) also provide neural net s/w. I don’t see much opportunity for code re-use here. Tesla engineers would write code to reformat sensor data, integrate NVidia’s outputs with the car’s control systems and such.

        Data re-use depends on the type of data Mobileye provides to customers. Raw sensor data would be highly re-usable. If Mobileye only outputs intermediate results from their image processing engine, however, I can’t see how that would be very useful.

    8. Martin Winlow says:

      I don’t know the answers to your questions – sorry – but I have a few ideas… But that’s not why I’m posting this reply.

      I did a masters in engineering 30 years ago but then sort of switched career paths and policed London for 30 years. In the meantime I developed a strong interest in EVs and converted a petrol van to electric in 2009 and have driven EVs exclusively since. I now own several of them, the latest being a Tesla Model S.

      Frankly, I belive ‘shorting’ is fundamentally immoral. It is basically betting against the good fortune of a business and, therefore, the health, happiness and well-being of all its employees. It absolutely sums up everything that stinks about capitalism (it has its good bits, too) and the same sort of twisted, self-serving mentality is behind every recession we have ever known, especially the last one.

      The fact that I appear to be the only one that thinks like this is a sad inditement on our society – reinforced by the almost *complete* lack of sanctions against anyone involved.

      Anyway, just a thought!

      1. Doggydogworld says:

        Martin – short sellers play a key role in markets. They are the only people with a strong incentive to uncover fraud and other bad acts. Since dishonest companies do harm, not just to customers and investors, but also to honest competitors, exposing them is a public service.

        Of course short sellers are sometimes wrong. But it’s short-sighted to call them immoral.

    9. Steven says:

      Rather than leaving the “phone off the hook”, data is probably pooled for a period and then transmitted. Perhaps every five minutes, the car “phones home”, sends a compressed file, and then hangs up. If the antenna is damaged, the car may be holding the data upto the point when all motion stops. Make sense now?

    10. Someone out there says:

      7) Is it really relevant data if it’s just the same stretch of highway over and over agan? For example thousands of Californians using the autopilot on their way to work every day. The vast majority of this data I would expect to be nothing more than noise.

    11. Aaron says:

      Honestly, Mark, what specifically do you think you will gain with the answers to those questions? Not a single one would interest me as a shareholder.

    12. MDEV says:

      Do you have a job?

  3. Pushmi-Pullyu says:

    I’ll admit that when AutoSteer (Beta) was first enabled in Tesla cars operated by ordinary, untrained drivers, I thought that was “a bridge too far” for semi-autonomous driving.

    But with only one confirmed case of AutoSteer being in control when a fatal accident happened, and uncounted lives already saved, it’s now quite clear that Autopilot/AutoSteer is already safer than the average human driver, and of course it will continue to rapidly improve… whereas the average driving skill of human drivers won’t.

    Kudos to Tesla Motors for not only pushing forward the EV revolution, but also pushing forward safety in travel by automobile!

    Go Tesla, doubled and redoubled!
    🙂 🙂 🙂

  4. Nix says:

    “whereas the average driving skill of human drivers won’t.”

    Based on my daily commute, I can firmly confirm that since the invention of the smart phone, that driving skills have certainly gone down on the roads I drive.

    1. Pushmi-Pullyu says:

      As I understand it, there has been a measurable increase in the frequency of distracted driving, due to cell phone usage.

      The statistics are pretty frightening when you think about it. As I recall, statistics are that about 1 in 50 drivers are drunk. But distracted driving, due to cell phone usage, is according to some just as bad as driving drunk.

      And there certainly are a heck of a lot more than 1 in 50 drivers who are using their cell phone while driving!

    2. Trey M says:

      Agree 110%!! I see cars daily that swerve into other lanes, don’t brake when they should, don’t accelerate when they should. In most all of the cases I witness, when I pull alongside them, they have their eyes down looking at their phone.
      I am not totally innocent, but I refrain when there is traffic around.

  5. JR says:

    So interesting will Tesla without the Autopilot software installed also contribute to collection of data?
    I think Yes

    1. Bert says:

      Yes, all tesla’s are made with autopilot hardware and the software is also send to all tesla’s. The only difference is if the customer has paid to be allowed to use it or not.

  6. Get Real says:

    Can’t help but notice how desperate Spiegel is getting now as he consistently loses more and more money on his shorts.

    But really, Spiegel is little more then a leech who puts his personal gain ahead of everything that Tesla is and WILL do too advance the greater good of all humanity and the planet.

    What an unpatriotic douche.

  7. Loboc says:

    I seriously doubt that Tesla’s data is worth much.

    1. It is old data from an old hardware model. I’ve done conversions and data mining. It’s not easy to relate a Lawson system to a SAP system.

    2. The drivers of these vehicles are self-selected 1%ers. It won’t be statistically correct for other drivers.

    3. The data has duplicates. Maybe thousands of them. Driving the same routes over and over gives maybe meaningful data if there are deviations, however, if I take my commute as an example: there is little deviation and it’s the same route. What more can you learn the 1000th time the same thousand drivers run that route?

    4. What about all the corner cases. A motorcycle riding down the lane marker between cars. A paper bag in the road. How many times are these encountered in a mere 1.3B cases? Would the human reaction be a viable learning moment? Would the pothole still exist next month?

Leave a Reply