Leigh Drogen Profile picture
16 Oct, 26 tweets, 5 min read
1) Before moving to Montana in July, we bought a Tesla Model Y, the first car I've owned since college

The drive across the country and the last few months have been...illuminating as it relates to the reality of driverless tech

Here's a thread on what's real/fake
2) I'll preface this by saying overall the car is totally worth it, and regardless of everything below, I still wouldn't have bought any other car, it's just....Elon is way out over his skis when it comes to the truth around the driverless stuff and how far off it is
3) The basic issue is that the car uses cameras and that causes a ton of edge cases, like so many that I can't imagine this basic paradigm ever being able to cross into true driverless

It's basically an 80% good system, but that's seemingly a zillion miles from 100% good
4) Many of the different algorithms that govern the behavior of the car just seem like they haven't been thoroughly worked on, like they were a first pass

This is likely far from the truth, but that in and of itself says a LOT about how hard the last mile is so to speak
5) Tesla likely will never get to where it really needs to go without adding technologies other than video, like LIDAR

But I can see them getting a bit further by tweaking those algos and feeding them a lot more data, sometimes it feels like just not enough data
6) If you put the car on a highway with no road work it's very capable, but the second you introduce anything out of the ordinary it freaks out

For example, it won't merge into the other lane when the one you're in is closing with cones, it will just run into the cones, WTF?
7) I almost killed us 15 minutes after leaving the dealership in NJ when I didn't realize it wouldn't recognize the pylons were closing our lane

It also won't navigate two lanes going to one without cones when it has to negotiate with a car in the other lane to go front or back
8) The car has a propensity to stop super short when another car crosses the road in front of it, or even when the car in front of you is turning off the road, this is super dangerous because the car in back of you will rear end you, and you're the idiot at fault
9) I get what's going on here, the collision avoidance system is very conservative, but you would think that there's enough data to augment behavior for a car in front of you turning off the road, or crossing in front with plenty of time so the breaks don't slam
10) sometimes it'll just freak out for no reason and stop short because it probably thought it saw something, but it's just a shadow or other artifact, again super dangerous
11) The system that is supposed to set the max speed limit doesn't work, because the cameras don't recognize all the speed limit signs, and it feels random when they do and don't

If there's anything around the sign, above or below it, no dice
12) I assume there is also an internal database for road speed limits, but that too isn't accurate, and sometimes the car will keep going say 55mph when the speed limit has actually dropped to 25!
13) It can't make turns at a light or a stop sign, totally incapable of pulling that off successfully because the system is overly conservative regarding things that are moving around in its vicinity, it drives like a scared child driving for the first time, which is dangerous
14) The traffic control system where it asks you to push the gas pedal to confirm going through a green light is buggy as crap and doesn't always ask you, then it just stops short out of the blue, again super dangerous

It doesn't know how to handle flashing yellow lights at all
15) The car is too aggressive in terms of how fast it takes turns, it does slow down enough into the turn but the g forces it pulls are too much for you to be comfortable because it's accelerating half way through the turn instead of out the back of the turn

Seems fixable
16) If there's any issue with the painted lines on the road, it freaks out and doesn't work, again this isn't scalable because construction takes place a lot of places and screws that up

We have good roads here, but imagine NYC, or basically any back road, no way
17) The auto lane change system is dumb as a rock, basically useless because it always wants to be in the right lane, which causes it to have to continuously change lanes to move around slower cars no matter how aggressive the setting you place it on
18) Another pretty basic algo is how fast the car accelerates from a stop at a light, it's way too aggressive off the mark and slams your head back into the seat a lot, my wife is starting to get annoyed, again just feels like they didn't work on that algo enough for some reason?
19) This is a sample of the issues, and it's not that any one of them is the end of the world, but when you put them together and look at why they are there you realize the basic framework is never going to get them to full driverless because it's more than incremental
20) All of that said, I use the driverless every day, it's awesome, with a handful of interventions I can go from my house to the airport or hockey rink while eating a sandwich or looking at my phone, you get used to managing the car more than having to drive it
21) And while it's still an 80% good solution to a really really hard problem, some of which involves a machine working properly having to deal with stupid human drivers (hard!), I think it's important to keep some perspective at just how incredible what's actually here now is
22) I can say without a doubt that if you nominally manage the car on roads you're familiar with, which reveal many of the edge cases as you drive them again and again, it's vastly safer than driving yourself regardless of whether you're eating a sandwich or looking at a phone
23) I missed a few other deficiencies in the autopilot I forgot to note

The car’s ability to negotiate highway traffic in order to change multiple lanes to take an exit is extremely poor

The reason again is that it’s too conservative when changing lanes
24) an important part of driving with other humans on the road is being the right amount of aggressive, if you’re too passive it’s actually dangerous, it’s a similar issue as the cars from two lanes merging into one
25) it’s also another example of where LIDAR may be required to solve the problem because it’s a matter of both depth perception and speed

The car in a sense is not making enough inference at probable future behavior from the cars around it

Maybe it’s just not enough data again
26) it all gives you a really strong appreciation for just how much processing power your brain is doing completely in the background when you drive and the various algos it’s running

Also should make you even more scared of driving drunk

• • •

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

Keep Current with Leigh Drogen

Leigh Drogen 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!

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!