Autopilot is already improving.

[ohmman]

I really wish we had a better understanding of how the "learning" is supposed to work. I taught my son to drive, so I could probably teach this thing, too, if I knew how.

Since it's complicated, here's a great example video (23 years old!) about using neural networks to drive a vehicle. Things are different now, but the concepts are the same. We handle more features, we centralize processing, and our networks are much deeper (this one is only 3 layers deep). But you'll see how the confidence rises and falls depending on what the network sees. This is when AP would give you alerts.

Short video, but a classic and a decent introduction to the very basic concepts involved in this type of learning.

 

[Soolim]

I don't use the auto lane change (auto lane change disabled), because I activate turn signal, manual check that space is available, and manual lane change. In my style of driving, I would like the auto-steer to use the off ramp only when I activate turn signal (momentary). When lane widens and eventually splits, I would like auto-steer to keep left or right based on momentary turn signal to that effect.

Or, if auto lane change is enabled, lane change should only happen after 3 turn signal clicks. Any comment on this wish list?

 

[snort]

Currently AP is intended only for use on the highway. So in my view, it should currently always intend to stay on the highway. If the trainer (driver) is keeping the car on the highway most times, it should be learning to prioritize the trajectory and left lane marker as well as anything else that keeps it on the highway.

Yes! this seems like the right choice for the foreseeable future. until they have a robodriver that really is good enough to take over for an indisposed driver, they should make no bones about it being specifically for this narrow, straightforward, but very common use.

I have seen very, very little that indicates it is doing anything but follow lane edge markings or a leading car, all of which it sees through the forward looking camera and radar. I have seen nothing in my car's behavior that suggests its using GPS data in any productive way...in fact, the only places that make me think it might be there at all is where the car strays away from the middle or out of the lane entirely. given the data that I know they have (which is used to draw the lane lines on the IC) it would be trivial to stay in the middle of the lane. steering the vehicle is the easy part--it's the pattern recognition of the lanes in wildly varying contexts that's hard. but they're not keeping it in the middle, and the errors they are making are fairly typical of replicating what a sloppy driver might do...often the opposite of what a trained, skilled driver (e.g. a racing driver) would do.

they seem to have made some odd choices: when the lane markers diverge or one disappears, they seem to steer aggressively away from the one that remains. e.g. a right-hand exit on the freeway, they'll reliably take the exit. when the road widens, e.g. from a two lane, to two lanes with a parking lane on the side, it'll steer into the parking lane if it's empty. if the center stripe disappears, e.g. for a crossing, it'll veer /hard/ into the opposing lane.


Many years ago I wrote a clone of a popular video game. the "bad guys" would shoot completely randomly, sometimes hitting the player, most often missing. dozens of people asked me "how did you get it to aim at the user so well?" the answer is, I didn't. people tend to see what they expect to see. confirmation bias can only be only overcome by aggressive analysis and conscious effort to be aware of it and reject it if necessary.

--Snortybartfast

 

[mspohr]

I don't use the auto lane change (auto lane change disabled), because I activate turn signal, manual check that space is available, and manual lane change. In my style of driving, I would like the auto-steer to use the off ramp only when I activate turn signal (momentary). When lane widens and eventually splits, I would like auto-steer to keep left or right based on momentary turn signal to that effect.

Or, if auto lane change is enabled, lane change should only happen after 3 turn signal clicks. Any comment on this wish list?

On lane change, it's probably best to check to see if the lane is clear before activating the turn signal. This will prevent panic in anyone close to you in your intended lane.
It seems like a good idea to use the turn signal to indicate (to the car and to other drivers) that you intend to use the exit... however, the car might get confused and think you want to change lanes rather than exit. I don't know if the car can distinguish between the shoulder and another travel lane to the right.

 

[RyanT]

On lane change, it's probably best to check to see if the lane is clear before activating the turn signal. This will prevent panic in anyone close to you in your intended lane.
It seems like a good idea to use the turn signal to indicate (to the car and to other drivers) that you intend to use the exit... however, the car might get confused and think you want to change lanes rather than exit. I don't know if the car can distinguish between the shoulder and another travel lane to the right.

The other day it seemed to know the difference between a dotted line and solid line for lane changes for me. I was trying to do a lane change on a highway with a solid middle line and it was a no go. No other cars around.

 

[Soolim]

On lane change, it's probably best to check to see if the lane is clear before activating the turn signal. This will prevent panic in anyone close to you in your intended lane.

In my daily commute with high volume traffic and manual lane change, turn signal needs to be engaged or I could never change lane, since others won't know my intention. This has been my habit. Old habit is hard to change, that is why I have not enable auto lane change for lighter volume freeway driving yet.

 

[mspohr]

In my daily commute with high volume traffic and manual lane change, turn signal needs to be engaged or I could never change lane, since others won't know my intention. This has been my habit. Old habit is hard to change, that is why I have not enable auto lane change for lighter volume freeway driving yet.

Of course you need to signal the lane change. However, you should check to see if the lane is clear first.
I hope you aren't using the signal as a "request" to barge into an occupied lane.

 

[Chasedrgc1223]

Watch the instrument cluster where the vehicle path is drawn, and you'll see the difference in solid vs striped lines.

 

[donv]

How does autopilot do in areas where there is no data connection (i.e. 3G or LTE)?

 

[Soolim]

Of course you need to signal the lane change. However, you should check to see if the lane is clear first.
I hope you aren't using the signal as a "request" to barge into an occupied lane.

No worries about barging. I think I am a courteous driver and others are also quite civilized. By signalling ahead, they would let me in with more room to spare.

 

[gsxdsm]

This is one of the great things about the Tesla experience - so many of the owners are engineers and technically inclined that you can have really great conversations and analysis around the techniques used to enhance the driving experience. No other manufacturer has that.

 

[msnow]

One question I've had on the "leaving freeway exit ramps unintentionally" issue is wouldn't this have been easy to program out of the system using gps data? Surely Elon Musk and the rest of the team realized the cars would follow exit ramps. Couldn't they have programmed the cars at the launch to remain on the freeway unless the driver used the turn signal to indicate that a freeway exit was desired?

- - - Updated - - -



I'm surprised the autopilot wasn't programmed to look at map data ahead of arriving at a curve, calculate if the current velocity was too much for that curve, and then slow down. Even the primitive cruise control in my 2004 Infiniti QX56 SUV slows down when the road curves and speeds up again when it straightens out.

Well it sounds like Elon was listening...his post on AP update addresses this very thing.

BTW a cool username calisnow [emoji6]

 

[Cosmacelf]

Elon's post? Where?

 

[scottf200]

So Tesla uses Mobileye lane detection, but augments it with their own high precision GPS maps which tells it what inputs to use in a particular location.

It certainly appears that Mobileye does more than "lane detection". EM has mentioned HPP a few times ... which is Mobileye processing it appears.

"holistic path prediction" (HPP) is shown/discussed in Mobileye video presentations.

11-May-2015 Mobileye - http://finance.yahoo.com/news/edited-transcript-mbly-earnings-conference-044448949.html

We [Mobileye] have also a holistic path prediction [HPP], the ability to predict the path using a lot of context outside of the main lane detection mechanisms. It is relevant for hands-free driving, of course; you'd like to have a very high availability of lateral control even in situations where you don't see the lanes. But on the other hand, it's very relevant for the existing lane departure warning and lane keeping functionalities.

You can think of traffic light detection as an enabler for automated driving. When you're going into semi-urban areas and there are lots of junctions with traffic lights, you'd like to be able to handle those automatically. But on the other hand, you can think of traffic light detection as a very nice addition to the existing ACC functionality where when you engage ACC, the car would either warn or stop when you reach a red-light junction.

So, all of these new technologies are designed and targeting automated driving, but they also have a secondary use for the existing ADAS function and we see a lot of traction for them. For example, the two major wins that I mentioned at the beginning of my script, they all ask for these functions, regardless of whether it is going to be for automated driving or not. So the traffic light detection, the road profile, the free space, the HPP, all of them are already being sourced since a few months ago.

 

[snort]

Well it sounds like Elon was listening...his post on AP update addresses this very thing.

Elon Musk (@elonmusk) | Twitter

 

[sillydriver]

My own sense though is that we still do not know how much the improvements we are NOW seeing are due to algorithmic learning and how much is due to just accumulating better input data (map building). My hunch is that the latter is the major factor at this point, but undoubtedly the algorithmic enhancements are bound to come as well.

Judging by Elon's tweet about v1.01, I think you are absolutely right. He said:

Autopilot 1.01 coming soon: curve speed adaption, controller smoothness, better lane holding on poor roads, improved fleet learning!

It sounds like the algorithmic improvements are all packaged in the new release. His saying 'improved fleet learning' implies that there has been some learning in the original version. But if that learning isn't algorithmic, it must all be map building, as you said.

 

[heltok]

If true, this changes a lot. Call me impressed!

 

[ohmman]

Personally, I think this implies that there haven't been any model updates, and no improvement to date. I believe, as snort mentions upthread, that there's a lot of confirmation bias going on. I tried tweeting Elon back for more updates, to no avail.

I am attending the NIPS conference in Montreal in December. I'm very much hoping to run into someone from the Tesla AP team, if not a handful of people. Could get some more answers on the functionality if so.

 

[Magus]

I hear you. However what we are noticing is entirely in line with what we've been told. Intersections where I live, areas I've been to have web getting better with AP. There are curves that I visited when AP first came out that threw the alert to take over. Going over that same route yesterday was covered by AP without any alert or difficulties. It is a very specific sharp curve in the route that threw AP off 100% of the time roughly a week ago- requiring myself to take over every time. Now, no alert and AP does well in that sharp curve.

On my commute routes it may nudge towards exits it used to veer towards. AP is now able to cross intersections around where, without alerts as previously.

You are welcome to call this confirmation bias, and reach out Tesla. As a mental health provider myself I'm well aware of the concept. All of the reports again are completely consistent with what Tesla is reporting. Week to week improvements. Sill not "perfect" but most definitely noticeable and incorrect to call it confirmation bias. I understand the scientific bent, to be skeptical until there is actual data. However in this case, data as shown is this thread indicates that what we are seeing is accurate. In fact, the onus to provide evidence of confirmation bias is on you.


Let let me rephrase that. The evidence has been presented that AP is prepared for machine learning. This was clearly presented. So in order to hypothesize that we are experiencing confirmation bias, scientifically speaking you have to provide the data to show. Again, there is preliminary evidence of learning. The hypothesis that we are experiencing confirmation bias has much less weight in these circumstances.

 

[stevezzzz]

Personally, I think this implies that there haven't been any model updates, and no improvement to date. I believe, as snort mentions upthread, that there's a lot of confirmation bias going on.

I would expect some confirmation bias, but I have definitely seen improved Autopilot behavior in some, but not all, of the 'trouble' spots I drive on a regular basis.

One area where it has NOT improved is a particular lane split on I-25 southbound, in Denver, where the rightmost through lane becomes an exit lane (but it's also really a through lane, in the sense that it splits from the main through lanes, serves a couple of local exits and on-ramps, and then merges again with the main highway). When the rightmost lane becomes an exit lane, AP has no issues tracking in that lane; but when the split comes, AP consistently dithers about whether to continue in the established lane or veer left to stay on the main highway, until I have to take over to avoid hitting the start of the divider that marks the split.

Google Maps