Autopilot during rush hour traffic .... a parting of the seas occurs .....

[BozieBeMe2]

Yes it does.

Tesla can make a neutral network to turn the damn wipers on but uses basic if/then logic to accelerate the car in traffic.

In some cases, I suppose that your right But, in regards to the wipers, that neural network that you are referring to, is nothing more that a vision network. The last time I checked, Tesla is still using the forward camera to detect raindrops obscuring the field of view, pixelation.
The same, 'If then/' logic,
What you may be asking for is a 'Common Sense' logic, which I believe needs a neural network of 'Cross Talk', to the other cars to be sure of what those cars will do.

 

[chronopc]

Umm, that's kinda the entire point of having the car do the driving.

You can correct 100,000+ machines one time. With humans you have to teach them all separately; including the unteachable.

From this thread it appears that some people are perfectly fine with this behavior as it is exactly how adaptive cruise control work. They're designed to be simple because doing too much can cause issues.

The benefit of having the car do the driving is that it will drive the way it was programmed to vs. people who behave randomly.

I am perfectly okay with a behavior as long as it's consistent.

Speeding up to fill up a gap during stop and go traffic is really normal. In cities where there are a lot of traffic, people purposely do this so no one can get in front of them.

 

[JeffK]

Speeding up to fill up a gap during stop and go traffic is really normal. In cities where there are a lot of traffic, people purposely do this so no one can get in front of them.

It is normal, but unfortunately, it's the not checking your mirrors when changing lanes that causes trouble with this.

 

[conman]

You already have a feature that adjusts AP to match this. They called it following distance.

If you have a tight following distance set, the car has to react aggressively, especially on slowing down - there's no room to do anything else, and the current 1 setting seems to be inside of the car's effective reaction time (which is much longer than the theoretical reaction time from the systems installed,) so the car has to brake harder than the car in front.

On higher settings, the car uses the extra distance to reduce the sharpness of the reaction, driving defensively.

Even on setting 1, the take off at traffic lights is glacially slow, so no.

 

[chronopc]

It is normal, but unfortunately, it's the not checking your mirrors when changing lanes that causes trouble with this.

I think the car does a good job at changing lanes. I noticed it's really conservative so it waits for a big gap before it does it.

 

[Saghost]

Even on setting 1, the take off at traffic lights is glacially slow, so no.

Not sure why you think the two are connected.

Yes, the car is very hesitant coming off of a complete stop in all distance settings.

That doesn't change the fact that changing the setting makes a big difference in both braking and acceleration beyond the initial bit.

 

[SigNC]

I thin EAP acceleration in general is too aggressive, even in chill mode. I hate how it hammers down when i switch lanes into a clear lane. That being said i'm sure the distance it can check looking back in an adjacent lane is limited so perhaps it accelerates hard on purpose.

 

[Cheburashka]

There was a thread in the Autonomy forum that talked about how many cameras Tesla was using for AP functionality. Have they gotten past using the normal and wide angle cameras yet?

Not yet..

In some cases, I suppose that your right But, in regards to the wipers, that neural network that you are referring to, is nothing more that a vision network. The last time I checked, Tesla is still using the forward camera to detect raindrops obscuring the field of view, pixelation.
The same, 'If then/' logic,
What you may be asking for is a 'Common Sense' logic, which I believe needs a neural network of 'Cross Talk', to the other cars to be sure of what those cars will do.

Good points.

I don't think it needs to talk to the other cars necessarily. If my brain does not interpret an open space ahead of me in traffic as an opportunity to floor the pedal, the car shouldn't have to either. That's the whole premise of EAP with all it's computing power, cameras, sensors, etc. It should be more than just a dumb system.

 

[conman]

Not sure why you think the two are connected.

Yes, the car is very hesitant coming off of a complete stop in all distance settings.

That doesn't change the fact that changing the setting makes a big difference in both braking and acceleration beyond the initial bit.

That's precisely my point. I didn't think they were connected. I was referencing the fact that we'd like another setting to adjust the autopilot acceleration separately from the lead car distance.

 

[Economite]

My car does this when I first got it so I think each car still needs to learn it on it's own. It may share codes and references from the cloud but your car will learn it and in the future it won't accelerate that fast. I'm not sure how I fixed my issue. Maybe driving it on manual trained it or it just learned on it's own when it's on AP. My AP gets better each day on it's own even when there isn't any firmware updates. There is only 1 issue left that will need firmware update but after that I'm confident the car can do FSD once that's fixed.

Posts like this are what leave me really worried about the way Tesla communicates to its users how AP actually works.

All of Musk's random tweets about "machine learning" and "deep neural networks" have left drivers with all sorts of ideas about the extent to which their individual Tesla needs to be "trained" by the driver and about the ways in which the cars passively provide data to help improve future builds.

It's possible that Tesla has designed AP to "learn" and "be trained" the way you describe. But I don't think Tesla has ever explicitly said that's the way things work. Personally, I doubt it is. And if this is the way things work, then Tesla should more clearly explain that.

As it stands now, with things left vague, it seems like Tesla is almost encouraging AP users to keep trying AP in situations it hasn't been able to handle in the past (either to provide "training" or to check to see whether AP can now handle the situation). Such unguided experimentation seems like a bad idea, especially if it is based on an incorrect belief about how "machine learning" works.

Lots of folks criticize the Mountain View crash victim for using AP near a part of the road where he knew from past experience it didn't work well. To me, the talk of "machine learning" almost encourages this sort of experimentation. After all, if he didn't keep trying AP in that situation, how would he know that it still wouldn't work there. And, once it was successful a number of times in a row, how would he know that he had just been lucky and that the "problem" wasn't actually "fixed." The problem here is that if AP is 99% successful at navigating a stretch of road (and this is actually probably a low success rate), on average a driver can keep using it in that stretch and 99 times it will be successful, but on the 100th time, boom. Hard to not get complacent with those odds.

 

[Saghost]

That's precisely my point. I didn't think they were connected. I was referencing the fact that we'd like another setting to adjust the autopilot acceleration separately from the lead car distance.

Okay, but while the slow initial acceleration from a stop seems consistent across distance settings, the dynamic response at higher speed certainly isn't - and that's what I thought the discussion here was mostly about. You can't tell the car "follow really close" and expect it not to accelerate hard in an effort to follow really close when a gap opens.

On a higher setting, it won't do that (but you'll get idiots trying to go around you even though you're going just as fast as the car in front because there are a lot of really stupid, aggressive drivers out there.)

 

[jtpassat]

Heres a video of what I was describing as “slight of hand”

This car is on Autopilot. What happens next?