Tesla Pseudo LIDAR Using Just Cameras And Imminent Acceleration of Self Driving Improvement

Tesla can compare two or more video frames to confirm the depth of objects and generate pseudo LIDAR. They can convert images to a birds-eye view representation of what is around a car. They are also able to train with an unsupervised method.

Many technical people and some business people with better views of the behind-the-scenes development at Tesla are getting far more confident about Tesla’s full self-driving in 2-3 years.

Tesla is massively increasing the rate of improvement of full self-driving in mid to late 2021 and onwards with the full release of the Dojo training supercomputer and increases in the processing of data.

Tesla will create their own chips for the Dojo self-driving training supercomputer. Tesla will be using NPU chips (Neural Processing Unit). Currently, it takes 3 days to do a training session for self-driving. The deployment of Dojo will allow for larger training sets and reduce the training time to 7 hours. This would allow three training sessions per day instead of two per week. The new hardware and software will enable unlabeled and more unsupervised training. Parts of Dojo were already online in August-October 2020. More parts of Dojo have been released or are being released with the new FSD betas.

SOURCES- Tesla, Dr Know it All, David Lee Investing, Solving the Money Problem
Written By Brian Wang, Nextbigfuture.com (Brian owns shares of Tesla)

36 thoughts on “Tesla Pseudo LIDAR Using Just Cameras And Imminent Acceleration of Self Driving Improvement”

  1. In the meantime, someone needs to do an Uber for drunks who need to be driven home in their cars—call it “Tip-see” 😉

  2. For another example, on Youtube (which is part of the Alphabet empire) the search engine can be given the exact title of a video, with the uploaders user name, and still relegate it to page 3 or 4 behind a murder of results that have a couple of the same keywords but come from big money media companies with close financial ties to YT.

    If it's a "politically extreme" video (or if the uploader has other videos judged as such) then it may not appear in the results at all.

  3. Are you suggesting that Duckduck is producing curated results? Or is it that their "less intelligent" algorithm is not second guessing Brett's search terms?

  4. You can see how cut-throat the competition is between insurance companies by looking at their "cost of float" which is in their company reports.
    That is, how much the money they pay out, minus the premiums that come in, compares to the total amount of money they have sitting around.

    For competitive markets, (US car insurance for eg.) the cost is positive 1 to 3%. That is, they pay out MORE than the premiums. The way they can make a profit is that they have all that money sitting around for years until they have to pay it out, so they can earny a few % per year by investing it.

    Note: This is not the level of competition we see in more controlled markets like health insurance.

  5. The "quality algorithm" at Google IS curation! They staff their "trust and safety" team with left-wingers, and then the algorithms are trained to replicate their bias, in addition to various black and white lists they maintain.

    Again: If a search engine actively avoids returning the results you are looking for, it's broken. YOU may want Google curating what you know, but I don't.

  6. Then you cant use a quality algorithm, you have to use curated results maintained by like minded people or a wild card search.
    Else that would be like using threshing when you really want to throw away the wheat and keep the chaff.

    Quality and authoritative sources are dependent on evidence and not opinion, the highest quality of reality is demonstrable.

  7. I evaluate search engines by how well they find what I'm searching for. If they direct me away from what I'm searching for, in favor of what somebody else thinks I should have been searching for, that's a failure.

    How am I supposed to form my own opinions, if I'm systematically directed away from one side of the argument? If every time I try to see what the people on one side are saying, I get sent to the other side attacking what they're saying, instead?

    That the censors happen to be more in agreement with you than me doesn't make them any less censors. It doesn't really make them any more objective, either. That's just confirmation bias on your part.

    The bottom line is, I don't want a search engine spoon feeding me opinions in the guise of 'protecting me from misinformation'. I would be a fool to trust them to accurately distinguish "wrong" and "we disagree with it".

    The moment you trust ANYONE with that power, they abuse it. Really, they only try to obtain that power in order to abuse it.

  8. When I'm searching for something ideologically neutral, I use google, because they're superior on topics where they don't have any motive to put a thumb…

    That's because you have different standards for different things. You seek results that are high quality and from authoritative sources for "ideologically neutral" topics, and you seek low quality results for the rest.

    The problem isn't google's algorithm, it's your selective love of low quality non-"ideologically neutral" sites and "conservative news sites" love of posting low quality, unsupportable and delusional content to attract people trained on that stuff.

    You cant expect disinformation sites like project veritas, which chronically fabricates stories with their artful editing, to be filed on the same bookshelf as a normal news site.

    Chronic unsupportable assertions and intellectual fraud are not the hallmarks of high quality and does not make one an authoritative source of information. You cant expect a high quality search engine to make an exception for your low quality sites because they're supposedly a "conservative news" site, that would be manipulating the results.

    Dont ask the rational world to lower their standards, ask what passes for the "conservative media" to increase their quality and embrace evidence. They should stop profiting from damaged people by reinforcing and preying upon their delusions.


  9. Brett's answer seemed on target given the concerns you stated.
    Maybe give some examples of things you think might get past unsupervised training?

  10. Neutrality is probably too much to ask for, I'm satisfied if I have a variety of biases available, instead of being limited to one. Each eye has its own limited perspective, but binocular vision compensates.

    In fact, I think it was the cult of journalistic objectivity that opened the path for them to largely purge journalism of competing biases, and produce the current monoculture.

  11. Lidar cannot read street signs, see lane markers, or tell what color the traffic light is. They are good at creating precise point-cloud distance data however it remains to be seen how precise we need distance info to be in the far field. Up close it needs to be precise but that can be effectively done up close with radar, visual cameras and structured light, and ultrasound. In the far field, ok distance measurement is probably ok with relative speed regarding your vehicle being more important.

  12. My cynical crystal ball says that rather than dropping rates over all, they will offer slightly lower rates for operators of autonomous cars to accelerate the rate of adoption, but not enough lower to threaten their projected profits.

  13. I follow this stuff a lot on imagesensorsworld blog. From what I have seen, I am not really sold on active imaging systems like lidar due to their low resolution and susceptibility to interference when several of the systems start to be used side by side. Passive systems have the edge in my opinion. Recently a company called Brightways was covered with what they called "gated vision" technology that performed quite well. Check them out: https://www.brightwayvision.com/

    Also recommend following image sensors world if you are interested in self driving cars since image sensors and lidar are critical components.

  14. I have a plugin that lets me screen specific websites from google search. So whenever I run a search and click on an article which is too biased, I add that site to the blocklist. It's pretty long now.

    Neutrality is an underappreciated virtue among journalists these days.

  15. I'm wondering… can the pseudo-LIDAR cameras be dazzled by low sun or oncoming headlights?
    Seems like they'll need some kind of iris or dynamic filter to adjust almost instantly to any changes in ambient light, no matter how extreme the change. Otherwise these cameras won't be any good for judging distances until the light conditions return to something the system can handle.
    So if the system is dazzled, just as people are in those circumstances, what happens?
    I get that cameras are cheaper, and if this works it can be retrofitted to all existing Teslas (with enough cameras) via OTA software patch, but proper LIDAR seems far less susceptible to blinding in such conditions.

  16. I agree that the costs caused by accidents would drop. Insurance rates ought to drop as well, and will if there is proper competition between insurance companies. I fear there will be a tendency for the companies to keep rates higher than the loss statistics should justify. I expect a lot of public controversy over that point.

  17. Your first reply was off my point. This second reply does not ease my concern. I have my doubts that reviewing results would catch the kinds of things I am concerned the unsupervised training could get wrong. Reviewing, at least the way it often is done, is too passive.

  18. You may be right about me missing the point. But I assume the results of the training are examined, not just assumed to be correct.

  19. Photographs might be better than LASER if the computer can be fast enough to extract the data from it because a LASER beam will only reflect off one point at a time and has to very quickly scan the area as the action is occurring. A photograph takes it all in together.

  20. Might be belaboring the point, but all autonomous cars need to achieve is a statistically lower accident rate for the insurance industry to start lobbying for allowing them on the roads. Currently, the US average is around 1.1 deaths per 100 million miles driven, reduce that by 10% and they save money.

    I also wonder if autonomous cars will reduce the severity of collisions, as they will react faster, even in circumstances where they can't prevent the collision. So we may see the fatal and injury causing collisions reduce at a faster rate than the over all frequency of collisions.


  21. Well, you've improved safety 99%, once you've got the baseline performance equal to an undistracted human. It isn't there yet. In particular, while humans are distractable, we're still pretty good at reacting to out of the ordinary circumstances. An obviously empty plastic blows across the road? Ignore it. A similar bag obviously full of stuff sitting on the road? Avoid it, who knows what's in it.

    Your self-driving car right now will tend to treat them both the same.

    But I'm confident it will get there, because the machines are improving, and the people aren't.

  22. yes, it is a hard fact, humans can be distracted. The pretty girl, a pretty view, a couple fighting by the side of the road, the phone, the radio dial… The assumption is computers can't. Right there you have improved safety 99%.

  23. bb. wish you the best of luck. Betting against Musk is an incredible feat of strength. He has assembled the best engineers and has resolved production to a level of basic brilliance. He is a hard worker and a brilliant mind, when you have that combination and a system to attain your goals, well, there is nothing in your way.

  24. It doesn't have to be perfect, it just has to be better than a teen who just got their driving license. If we'll let a human that bad drive, why not a machine?

    And if humans can drive with pseudo-lidar, aka "depth perception", why not machines?

  25. We're talking here about human lives, pseudo is not enough. If it won't be as good as LIDAR, even if it will achieve 99% accuracy of it, still not good/safe enough, because once again we are talking here about more potential accidents and human lives

  26. Unsupervised training doesn't mean they have the cars driving around on city streets unsupervised. It's a technical term for training a learning engine where you're not actively telling it what to learn, but instead letting the learning fall out of the data.

  27. We actually have testimony from people within Google to support the idea that Google is deliberately censoring search results on an ideological basis. It's easy to do: You just hire left-wingers to do all your 'objective' evaluations, and bias automatically happens.

    Not censoring mention of the candidates, but instead issue groups and particular news stories, such as reporting on the Hunter laptop.

    And, yes, when I want search results that haven't been manipulated to hide from me things Google thinks I shouldn't find, I do use DuckDuckGo.

    When I'm searching for something ideologically neutral, I use google, because they're superior on topics where they don't have any motive to put a thumb on the scale.

    See, for instance: https://www.realclearpolitics.com/articles/2020/09/20/google_pushes_conservative_news_sites_far_down_search_lists_144246.html

  28. "engineering.stanford.edu/magazine/article/are-search-results-biased-along-partisan-lines"

    In a world where well supported and quality information is valued, it's unusual to expect sites who cater to low quality delusional information will rise to the top. Not everything is a conspiracy, the only problem here is a large % of people embrace low quality delusional information as a sacramental principle of their world view. They expect, in vain, that particular brand of sour cream should naturally rise to the top and it never will.

    If you want a naive key word search that will present every site on the web that embeds the meta keywords such as "hillary, pizza, kids", use a search engine like duck.

  29. "After many years of improvement, Google algorithm still doesn't give perfect search results"

    In fact, Google's algorithm intentionally gives less than perfect results. Setting aside the ideological filters designed to make it hard to find things Google doesn't think you should be looking for, (Which are a larger and larger factor as time goes by.) and sites gaming the system, there's paid promotion of results to the top which wouldn't naturally be there, and they intentionally screwed up the search string parsing so that people who didn't know how to properly construct a search string would still have a chance of finding what they were looking for.

  30. Fool me once, shame on you, fool me twice…

    We've heard this song from Musk before, probably to generate loads of free PR. How many times will Tesla/Musk be able to do this before it stops working?

  31. I'm out of my area of expertise, so I might be misinterpreting, but the notion of unsupervised training of something safety-critical, such as self driving cars, seems dangerous. It seems to me that would allow too high of a chance that the training would result in at least some cases where the car would be trained to make choices that yield results anyone who evaluated them would deem to be clearly the wrong results.

    I am a supporter of the development of self driving cars, not opposed to it. I hope I'm wrong about this uneasy feeling about unsupervised training, but that is the feeling I get from only reading the text in this article (I did not view any of the video).

  32. Here is a good summary of the different approaches regarding autonomy. Tesla pure ADAS and perfect algorithm approach may prove to be a more of the same dead end. After many years of improvement, Google algorithm still doesn't give perfect search results, but the end result is still far less likely to kill people than autonomous driving algorithm. There is a misunderstanding that the Musk and his blind followers have made, intelligence is derived from consciousness, an element in itself, Artificial Intelligence will help, but not replace ours completely, some of it may be required in driving, adding maps with information for driving autonomy and additional sensors may allow for sufficiently bypassing this obstacle. I see in the future added to that a network of human and self driving cars that provide information for driving traffic as a whole. One type of sensors and pure algorithms – is probably not enough.


Comments are closed.