(1/13) Why #Tesla can't handle curbs & why they can't surpass Level 2 automation: A brief explanation:
We've seen countless examples of Tesla's "Smart Summon" failing in the simplest environments. In the example below a Tesla drives over a curb but why can't Tesla see a curb?
(2/13) Tesla has 3 types of sensors: a low resolution camera array (8 cameras @ 1.2mp), ultra sonics and a forward facing radar (in the process of being removed). Radar is completely irrelevant here and Tesla's ultrasonics are also not configured for curb detection.
(3/13) For people interested in ultrasonic curb detection I'll link a paper below. So Tesla has to use their cameras to try to figure out where a curb is. It should be noted that their cameras are not really adjusted for this task, as they point too high.
(4/13) Tesla advertises that they have a 360° view around the car. This is demonstrably false as they have multiple blind spots. Blind spots are one of several reasons why Tesla is not able to handle curbs, they simply don't see them (at least not accurately enough).
(5/13) But I don't think that was the case for the failure documented above. In this case we can see at least two complications Tesla has to deal with: irregular painting and multiple light sources. Tesla has no way to measure distances, they gain depth information from ...
(6/13) ... images as they rely on machine learning algorithms which are trained to estimate (estimate!) depth from pictures from cues like perspective, scale, shadows etc. but if you watch this GIF you quickly see the problem: flickering & blurriness.
(7/13) It should be noted the depth information in the GIF above is of much better quality than what Tesla can extract from their feed but still you see chairs flickering in & out of existence, very low detail and a lot of blurry edges. This doesn't allow for autonomous driving.
(8/13) The summoned Tesla that hit the curb had to deal with a fundamental problem: curbs come in all different shapes and colours & as in this example they're lit from multiple angles. For the algorithm there are no (reliable) cues like shadows or lighting to extract depth info.
(9/13) As a result, Tesla doesn't see the curb and drives over it. It should be highlighted that Tesla cannot determine when their algorithms are reliable and when not. This is actually impossible, at least for now.
(10/13) Elon Musk and his cultists will argue, if humans can do it, NN can do it as well. This is a common pseudo-argument and a logical fallacy Musk uses to fool his gullible customers. He claims to have all the tools so the product is just a matter of work to be done.
(11/13) This is - of course - fundamentally false. We have absolutely no idea, not even a tiny bit how human vision works. I cannot stress this enough: nobody on this planet has an idea how human vision works. So Musk is obviously lying by claiming to have solved everything.
(12/13) Artificial neural networks are named because they look like what we see in a brain not because they work like it. Artificial neural networks have practically nothing to do with the actual human brain. If someone makes this argument, they don't understand any of those.
(13/13) Finally we understand why Tesla won't surpass L2 at least not with their current hard- & software approach. They lack the sensors & they lack the methods & knowledge how to achieve what they claim. It boils down to: "Surprise! Tesla didn't build a million brains in cars."

• • •

Missing some Tweet in this thread? You can try to force a refresh

Keep Current with Bullshit Exposed

Bullshit Exposed Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!


Twitter may remove this content at anytime! Save it as PDF for later use!

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

More from @BS__Exposed

21 Jun
(1/9) I cannot stress this enough: despite the fact that #Tesla fans will think this is good news, this is THE WORST NEWS for Tesla's FSD that has came out until now. Let me explain:
(2/9) First of all: No, Tesla has not the 5th most powerful supercomputer in the world, not even close. Though they have 1.8 exaFLOPS they only run at FP16. In laypeople's terms: very fast but not at all precise.
(3/9) Nevertheless, it's a large supercomputer tailored for their needs. So what's the bad news? Simple: despite having enormous resources Tesla hasn't even scratched the borders of L3 automation. #FSD in its current version can't even reliably detect objects & measure distances.
Read 9 tweets
20 Jun
(1/8) According to #Tesla conspiracy theorists, Tesla and #ElonMusk dominate the world of tomorrow. This is a compilation of a fraction of YouTube video thumbnails published by Tesla fans over the past year. The general theme is more than obvious: omnipotence.
(2/8) The fantasy of omnipotence is nothing new & is the most powerful tool to create a dedicated following. Tesla fans aren't fans of cars: they save the world with a genius leader who has unmatched powers & skills and who is (in their minds) already conquering other planets.
(3/8) Omnipotence is a compelling promise by Musk to his worshippers, follow him & you'll be part of his mission to change and lead the world.. the universe! Failing is not an option. This narrative reminds you of the playbook of all dictators around the world throughout history.
Read 8 tweets
17 Jun
(1/9) John Gibbs (@DrKnowItAll16) from the University of Georgia (@universityofga) makes a complete fool out of himself in a video where he invents numbers which he then refers to as data that he then says is evidence to prove $TSLA #FSD will be better than humans by end of 2021.
(2/9) Gibbs is so incredibly dense & uneducated that it is hard to even understand his verbal excrements. Essentially what Gibbs claims is that FSD will improve exponentially just because he decided to use an exponential function in a spreadsheet with made up numbers.
(3/9) He takes the number of driven miles and claims this number equals the improvement rate of FSD. To "make these numbers work" he then introduces arbitrary constants. One should note Tesla's mileage growth is not really exponential in the first place so everything is wrong. Image
Read 10 tweets
7 May
1/6 With #SpaceX in the news we look again at the profitability of their Falcon 9 rocket & what is going on with Starship. Without considering launch prices we can easily conclude that SpaceX currently doesn't make profits from these rockets, simply because of what Musk claimed.
2/6 We should keep in mind that Musk has nearly two decades worth of overstating and over-promising profitability and in general cannot be trusted with any statement. The competitor ULA has alleged that booster profitability would only be reached at ~10 reuses.
3/6 With this in mind we understand why SpaceX has to raise billions and billions for the supposedly cheapest spacecraft in history (Musk claims 1000x cheaper than anything else). The current Falcon 9 could in theory generate profit in the future but is many years away from that.
Read 6 tweets

Did Thread Reader help you today?

Support us! We are indie developers!

This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Become Premium

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!

Follow Us on Twitter!