Hacker News Reader

Comment by musesum | Hacker News Reader

My use case was autopilot steering me off the left hand side of the road at 80 MPH during a sunset [1] My (WAG) guess is that the ML corpus was based on a model 3, while the model Y, I was driving, had placed the sensors slightly higher? I dunno; again, tis a WAG.

I am a huge fan of Tesla's permise. My Dad sold Vanguard Citicars [2], and I worked on GM's interactive ad for the EV-1. I put a down payment on Aptera's original EV and again on the resurrected version [3].

But, I backed out of buying a Tesla. Even though I disabled Autopilot and ran purely cruise control, the Model Y would brake for phantom obstacles. Such as: low rises in the road. Or: passing a 18-wheeler on the left. Driving from AZ to CA, the phantom braking was occuring every 5 miles or so. So, I had to drive 800 miles with a manual accelerator. Bummer!

[1] https://news.ycombinator.com/item?id=31504583

[2] https://en.wikipedia.org/wiki/Citicar

[3] https://aptera.us [edit]

View on HN

When my dad was teaching me how to drive, he told me that cruise control is only for when you have at least a hundred feet between you and the next car, and that you fully turn it off to pass other cars. I treat Autopilot the same way when I use it. I treat it as cruise control, if I use it at all.

Problem is that, with no cars around for hundreds of feet, with dozens of seconds of reaction time, when you hit a low rise on the AZ Interstate, the Tesla Y will abruptly brake. This is with Autopilot OR with simple cruise control. Pretty stressful. Bonus impact on range prediction. So, I couldn't use either.

I know this is a controversial take, but I'll say it anyway: anyone who doesn't feel safe driving with cruise control is a dangerous driver. Cruise control is a tool. So long as you understand how it works, and that it doesn't paralyse your right leg, you really can use it in any driving context. It is an alternative input mechanism to the accelerator pedal, nothing more.

Agreed. Modern cars with adaptive cruise control and lane keep assist are a breeze to drive if you maintain awareness of your surroundings and practice confident defensive driving.

No, it is useless on VAG cars in Norway. You can’t drive at the speed limit at 110 and suddenly have the car slow down to 80 because of “reasons”. That increases the risk of being rear-ended for no reason, even though the car behind you is supposed to be attentive

Granted, you seem to be able to disable it in the menus, but if you rent/use car sharing services, this may not be obvious to you

Human drivers sometimes perform phantom braking, but we don't realise it because we are able to justify it to ourselves.

> My (WAG) guess

Wives-and-girlfriends?

Wild-ass guess

Wild-ass guess guess?

Yes, like when people say PIN number or ATM machine. Or HIV virus or LCD display.

https://en.wikipedia.org/wiki/RAS_syndrome

Are there iterated versions of that?

JWT token

This is not a good example - JWT is the name of a format, not of the token stored in that format. JWT token just means "token in JWT format".

Saying "a JWT" is in fact more dubious, like saying "an SQL" instead of "an SQL string".

https://datatracker.ietf.org/doc/html/rfc7519 disagrees:

Abstract

JSON Web Token (JWT) is a compact, URL-safe means of representing claims to be transferred between two parties. The claims in a JWT are encoded as a JSON object that is used as the payload of a JSON Web Signature (JWS) structure or as the plaintext of a JSON Web Encryption (JWE) structure, enabling the claims to be digitally signed or integrity protected with a Message Authentication Code (MAC) and/or encrypted.

Even the quoted fragment is a bit debatable - while the standard does commonly use "JWT" to mean a particular token, the very first sentence is only grammatical if we take JWT to be the name of the standard/format. Otherwise, it's missing an article - it should have been "*A* JSON Web Token (JWT) is [...]" if it were consistently used to refer to individual tokens, not the collective format.

> like saying "an SQL"

An SQL is an implementation of the System Query Language isn’t it? Like MySQL, or MSSQL.

SSD drive

Honestly I don't think these are as often wrong as some people think. While PIN and ATM work well even when articulated (please input the PIN, the ATM is broken), others don't ("the HIV has killed millions"). LCD is more on the fence for me.

Other common examples that don't sound that wrong to me are "the IP/TCP/... protocol".

For a different reason, I also don't think "this API expects a JWT token" is redundant: JWT is the name of a standard, so the phrase should be parsed as "the API expects a token in JSON Web Token format".

Or, as recently, PUBG Battlegrounds, which reuses 2 letters from the ititialism.

> the Model Y would brake for phantom obstacles. Such as: low rises in the road. Or: passing a 18-wheeler on the left.

Fwiw, I had similar phantom breaking experiences with my model y.

One of the recent updates fixed it such that phantom braking rarely happens (maybe once per 1000 miles driven or less now versus almost every time I passed an 18-wheeler on a two-lane highway before).

Yay; progress! My experience was from last year. Will stop whining, now.

Autopilot cannot be enabled until the position of the cameras is calibrated. That's why it requires you drive on well-marked roads for several miles when you first receive your car, before it works.

Through that is not how neural nets work.

It's anything but trivial to make a neural net properly abstract over some "camera position parameter".

More over it's nearly impossible to be sure it did properly abstract in all cases. I.e. it might have abstracting in every case but some arbitrary edge case which for a human looks no special at all but for some arbitrary reason is for the NN.

Anyway this is highly speculative and might very well be unrelated to the given Teslas behaviour.

I’m pretty sure some things get normal algorithms, that not everything is a giant neural net.

Hard to know without being on the team, hence my WAG modifier to the guess. My reasoning is that the Model Y is built on the same chassis as the Model 3. So, a quick and dirty solution is to use the same training set results. But, it most likely is something else.

This is a roundabout way of saying that maybe - just maybe - the problem is only for the model Y.

Am rooting for Tesla to fix it. Maybe a Tesla engineer looks at hacker news? Already filed a complaint with the Tesla Sales Manager. Or perhaps the log of my screaming in terror did the trick. Is prosody for QA sentiment a thing?

I don’t think they segregate their training data based on car models. I mentioned in an above comment, but Comma.ai doesn’t do this and their devices support tons of cars. It would be very odd if a company far smaller than Tesla was able to figure out how to account for different camera positions and Tesla wasn’t. I bought a 2022 model year car and plugged the Comma device in and it just worked, and they would have had pretty much no training data from my car at that point. Just my speculation though.

Are there more recent papers? I see one from 2016 [1] But, then again my assumptions are a bit dated as well. Was thinking: could a different horizon on a CNN mid-layer trigger a false positive? Perhaps, classify a slight rise as a bumper or some other obstacle?

Maybe a simpler system, like Comma.ai's cameras has looser tolerances. Somewhat akin to the one or two eyes of a Human driver.

Maybe it is policy. I could imagine the brand hit to Tesla for every crash - even due to driver error. Maybe phantom breaking is an artifact of erring on the side of caution. Maybe lawyers got involved. (The horror!)

Anyway, idle speculation, this.

[1] https://arxiv.org/pdf/1608.01230.pdf

I have a Comma 3 in my car, and it has a calibration phase that takes like 3 minutes of driving. The device works for many different cars of different sizes, and their driving algorithm uses neural networks. As far as I know you still have to get the device fairly close to centered on the windshield for it to work, but clearly you can still do some sort of calibration based on driving data alone. Maybe Tesla can account for even more deviation in camera position because they know ahead of time where the cameras are mounted?

I'm not saying it's impossible or will take long, I'm saying it's not easy to implement and depend on your pipeline subtle bugs can sneak in you might not be able to find with any testing.

I have zero knowledge of Tesla cars, but do wonder why this can't be pre-calibrated at the factory?

I don't know, but the calibration also has to be re-done if you have a camera replaced.

How long ago was that, Autopilot has been improving rapidly and since they stopped using radar I haven’t experienced and phantom breaking.

Reply on news.ycombinator.com