A relatively simple solution for phantom breaking from overpasses, why doesn’t Tesla do this???

[Jeooosh]

Given that phantom breaking is in and of itself, dangerous, and

under the assumption that phantom breaking is caused by the car incorrectly assuming that the structure of an overpass is an object with which a collision could take place,

Cull the data from vehicles showing this pattern, check that there is an overpass there, and tell the car NOT TO DO THAT.

Teaching the car “hey, there is an overpass here so don’t consider something physical in those few feet something you would break for” seems obviously worth it on the whole. That would turn off forward collision actions for what, .005% of the time?

What am I missing here?

Thanks.

 

[cucubits]

What if there is a stationary object or a slow moving car right at the over pass? With this scenario there is a good chance it will ignore it and plow right in.

 

[Phlier]

It isn't just overpasses that cause phantom braking, unfortunately.

In fact, the vast majority of the places where I get phantom braking are completely unexplainable. No shadows, no weird road markings, nada.

If it was as easy as stated in the OP, you can bet they'd have fixed it a long time ago.

 

[Tam]

...What am I missing here?...

Missing LIDAR!

If it was solvable with cameras and radars, it would be done years ago.

I think the only current way to solve it is by adding LIDAR as Waymo has done it for years.

 

[Daniel in SD]

I swear that I've seen @verygreen show that they do map overpasses and ignore radar returns from them.
Obviously it's tradeoff between phantom braking and phantom non-braking (aka crashing).

Missing LIDAR!

Humans do it with two eyes! I believe that by the time we get to HW10 they'll be able to do it reliably with cameras.

 

[JayNJ]

Given that phantom breaking is in and of itself, dangerous, and

under the assumption that phantom breaking is caused by the car incorrectly assuming that the structure of an overpass is an object with which a collision could take place,

Cull the data from vehicles showing this pattern, check that there is an overpass there, and tell the car NOT TO DO THAT.

Teaching the car “hey, there is an overpass here so don’t consider something physical in those few feet something you would break for” seems obviously worth it on the whole. That would turn off forward collision actions for what, .005% of the time?

What am I missing here?

Thanks.

That is how they fixed it with radar. They ignored it when going under bridges.

 

[JulienW]

As someone who drives in the city I can say that even if this could work it would only eliminate WAY less than 5% of my phantom braking.

Overpass is a SUPER SMALL minority of over all phantom braking.

 

[Jeooosh]

It isn't just overpasses that cause phantom braking, unfortunately.

In fact, the vast majority of the places where I get phantom braking are completely unexplainable. No shadows, no weird road markings, nada.

If it was as easy as stated in the OP, you can bet they'd have fixed it a long time ago.

I agree that not all phantom breaking is caused by overpasses (never said that was the case..)

However I have in general witnessed a very strong correlation between a good amount of phantom breaking being just before an overpass.

Do a google search on phantom breaking and note that much of the conversations are are about it occurring around a *particular* overpass. This proves that it is significant.

I do a lot of cross country driving and notice the overpass thing. I guess this has not been your experience.

 

[Jeooosh]

Missing LIDAR!

If it was solvable with cameras and radars, it would be done years ago.

I think the only current way to solve it is by adding LIDAR as Waymo has done it for years.

"it would have been solved years ago" -- how can you be so sure? This is a software platform like any other.

 

[Jeooosh]

As someone who drives in the city I can say that even if this could work it would only eliminate WAY less than 5% of my phantom braking.

Overpass is a SUPER SMALL minority of over all phantom braking.

My experience might be different as I do a lot of cross country. However even in Baltimore I remember it *always* occurring around a particular overpass.

 

[Tam]

...Humans do it with two eyes! I believe that by the time we get to HW10 they'll be able to do it reliably with cameras...

Of course, humans can!

In theory, if humans can, then we just replicate how humans process the data and it's solved.

However, it's not solved because the machine's brain is not as sophisticated as a human's.

That's why Tesla's brain keeps getting better with HW1, then HW2, HW2.5, HW3.0, then HW4.0 coming then guess when the HWxxx will catch up with humans' brain?

 

[Tam]

"it would have been solved years ago" -- how can you be so sure? This is a software platform like any other.

The software is constrained by hardware. That's why Tesla has progressed from HW1, then HW2, HW2.5, HW3.0, then HW4.0 .

HW1 is constrained by a single camera while HW2 has progressed to 8.

 

[TresLA]

This sounds like not-an-engineer trying to solve an engineering problem. If it were that obvious, we’d have self-driving cars easy peasy! And “proof” may not be what you think. I rarely experience phantom braking (driving highways, city streets, interstates for 200 miles each way, etc. and maybe once every few weeks? Maybe it’s because I’ve adjusted to compensating for it, I can’t even remember the last time. I’ve had inexplicable gentle slow downs that needed a push of the accelerator pedal but rarely the dangerous sudden braking/deceleration. I’ve read many others’ similar experiences. I suppose that’s “proof” that YMMV?

 

[EVNow]

tell the car NOT TO DO THAT.

What am I missing here?

The car doesn't understand English.

 

[Daniel in SD]

This sounds like not-an-engineer trying to solve an engineering problem. If it were that obvious, we’d have self-driving cars easy peasy! And “proof” may not be what you think. I rarely experience phantom braking (driving highways, city streets, interstates for 200 miles each way, etc. and maybe once every few weeks? Maybe it’s because I’ve adjusted to compensating for it, I can’t even remember the last time. I’ve had inexplicable gentle slow downs that needed a push of the accelerator pedal but rarely the dangerous sudden braking/deceleration. I’ve read many others’ similar experiences. I suppose that’s “proof” that YMMV?

He figured out a solution that engineers at Tesla also pursued (though I'm not sure if they're using it now).
Here's a write up from The Verge:

A "geocoded whitelist" of objects like road signs and bridges will help prevent false positives (and thus automatic braking for no reason, which can be alarming and dangerous), and allow the system to notice the potential for crashes that might previously have been ignored.

Tesla preparing significant Autopilot upgrade, will use radar as 'primary sensor'

Will use camera, radar, and a 'geocoded whitelist' to avoid crashes

www.theverge.com

I think the blog post might have been removed in the great RADAR purge of 2021.

 

[jjrandorin]

For some reason, the Ops "this is so easy, not sure what I am missing?" for this type of issue reminded me of this spongebob squarepants "how to draw a circle"...

 

[Wigglesworth]

I suspect they're probably not super inclined to implement broad brush fixes like this for phantom braking, because while it might fix certain false detections of an object, it could easily create new scenarios where the car ignores an object that actually exists and plows right into it.

Phantom braking is a serious problem and definitely a big safety issue, but it's a lot better than the software getting overconfident and failing to brake for something that actually exists. I'd definitely rather they error on the side of having my car phantom brake, rather than failing to brake when needed.

I can understand why they might be a little slow to implement a fix/fixes for it, given the consequences turning down the sensitivity too much. This is the kind of bug fix that has basically 0 room for error. Fixing phantom braking 99.9% of the time isn't an acceptable result if the other 0.1% results in ignoring a real object ahead and going splat.

 

[TresLA]

He figured out a solution that engineers at Tesla also pursued (though I'm not sure if they're using it now).
Here's a write up from The Verge:

Tesla preparing significant Autopilot upgrade, will use radar as 'primary sensor'

Will use camera, radar, and a 'geocoded whitelist' to avoid crashes

www.theverge.com

I think the blog post might have been removed in the great RADAR purge of 2021.

The OP is giving a “solution” to be applied to the current state of the code. In any case, Monday night armchair quarterbacking doesn’t help anyone.

I should also add (I can’t seem to edit my comment anymore), that I was responding to autopilot phantom braking that the OP was referring to. I have experienced what might be considered as phantom braking with FSD beta, but nothing to do with overpasses. It’s just the overall sudden changes in velocity (direction and/or speed) as it tries to figure things out.