Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Navigate on Autopilot is Useless (2018.42.3)

This site may earn commission on affiliate links.
Are autopilot override events flagged and sent to the Tesla autopilot team for analysis? Obviously they cant look at every event, but i would hope they’re using real world feedback. Maybe even use some sort of AI to look for trends.
The disengagements sometimes are.

What is always reported is the location where you disengaged (and how, i.e. brake pedal or steering or stalk). Anything beyond that is a big maybe.
If a disengagement was not triggered (i.e. you accelerated to override a phantom braking event) - that might trigger a snapshot if your car has a specific request for it in effect (a maybe).

The biggest feedback? You cause a huge accident on AP andmake national news - that usually results in some hasty actions to gather more data and prevent them in the future (e.g. after a flurry of "Tesla hit the firetruck" news they created a trigger to try and catch these conditions under the "accelerate into obstacle" guise - so I imagine there might be a fix for that eventually now, they also added the gore zones support and debugging after the infamous fatal CA accident, but in my experience the gore zones detection is still super glitchy and has a lot of false positives to the point of being unusable (by that I mean all such triggers I have seen were false positives)).
 
that is a *very* misleading choice of words if you ask me.

How many of the car companies you know are working on this? On the other hand who are the contenders? Intel (MobilEye), Waymo, Uber(?), ... none of them are car companies (those prefer to license ready made solutions anyway - it's safer this way too - somebody else is taking all the R&D risks! And the car companies can concentrate on their core strengths (or at least that's what PHBs/MBA would like you to believe as the good thing)). GM bought whoever does supercruise I think, but do they even aim at FSD ATM, any autonomous miles logged anywhere?
You can take a look at the rankings at Navigant Puts Tesla In Last Place Among Autonomous Vehicles. (Whether you agree with the rankings is a whole other ball of wax and should be discussed in another thread.)

GM bought Cruise Automation and they HAVE been doing autonomous driving testing in California, amongst other places. Per Waymo and GM still lead the pack in California’s new self-driving report cards, they did over 131K miles of such testing in CA public roads in 2017 vs. ZERO for Tesla.
GM’s self-driving division Cruise told the California DMV that it drove a total of 131,676 miles in 2017 — a 121,900-mile increase over 2016. Its vehicles (which have a range of hilarious names like Albatross, Sunbear, and Mongoose) reported disengaging 105 times during the 12-month period. A spokesperson touted the company’s “1400 percent rate of improvement” in the performance of its autonomous systems.
Cruise Automation posted Why testing self-driving cars in SF is challenging but necessary.

See reports at Autonomous Vehicle Disengagement Reports 2017.

For past history, see reports for Tesla and other testers at Autonomous Vehicle Disengagement Reports 2016 and Autonomous Vehicle Disengagement Reports 2015.

There's Honda Invests $2.75 Billion In Cruise Automation | CleanTechnica.

Nissan was willing to let a reporter ride along for
in 2017. Has Tesla demonstrated this to reporters riding along?

Nissan did reportedly launch Nissan and DeNA to launch self-driving taxi trial in Yokohama | The Japan Times
but it sounded like it was for a limited time and http://www.asahi.com/ajw/articles/AJ201807190038.htmland others implied it ended.

Cruise Automation has let reporters ride along but the the rides weren't perfect:
My Herky-Jerky Ride in General Motors' Ultra-Cautious Self Driving Car
My ride in GM's self-driving car offers a taste of what's possible

Here are some reports of Ford's efforts (for example) and some automaker partnerships:
Ford and Baidu team up to test autonomous cars in China
https://www.washingtonpost.com/loca...98119e-d2f6-11e8-b2d2-f397227b43f0_story.html
Ford triples the size of its autonomous test fleet | Autocar (from 2016)

Here's an example of another automaker doing such work: Hyundai Nexo hydrogen fuel-cell prototype drives itself 118 miles without human input.

Not sure if you keep up with Autonomous Car Progress, but if you did, you'd find the Nissan's been mentioned a few times (they're unlikely in the lead) and Cruise Automation has been mentioned a bunch.
 
Last edited:
I don't really have much positive to say about Drive-On-Nav aside from liking the visualization.

I don't really agree that it lowers the usefulness of AP because you're not forced to use drive-on-nav. You can still use AP2 just like you could before. With V8.X it would also occasionally fail on an automatic lane change.

I didn't get much experience with AP on AP1 because it became worthless to me because of some truck lust issues. What I really liked about AP1 was the the TACC was really good. With AP2 they try to do more with TACC in terms of detecting stopped cars, and it kinda ruins it because you get more false braking events. It's a balance Tesla needs to give the user more control over.

As to drive-on-nav in my limited experience it:

Slowed down excessively before an exit it was taking even though it was in the correct lane. Now maybe Tesla could argue that it was slowing down for merging traffic from the right, but I think they were mostly going straight. I would have slowed down a little bit to allow scissor merging, but it slowed down more dramatically. The car behind me went around me on the left and sped well ahead, and got back in. I don't know if it was in response to me or the traffic. There was a lot of merging traffic. He probably just assumed I was overly careful.

On the ramp I didn't like how it handled it. A normal human would use a consistent smooth motion, but it did a bunch of minor steering corruptions. Like a series of slight angles instead of a circle.

I didn't have any issues with the 360 degree visualization. I have an Model 3 so I don't think the rear camera glitches on it like it does on the Model S/X or at least I haven't seen it yet. I also didn't see anywhere where it didn't indicate red when I couldn't do the requested lane change. The visualizations were definitely improved over 39.7

The requested lane changes were mostly all really dumb

I didn't have any failed lane changing events with it, but my trip was short (30 miles each way)

My biggest beef with it is how it handles merging lanes, and other times where the lane gets wider. It's gets wobbly? I dunno the right way to describe it. It just starts to wonder towards the open space, and doesn't follow the lane line it should.

The best experience was when it automatically takes the exit. It feels really cool.

Will I use it? Probably not, but that's not saying much. I'm a huge fan of TACC only for various reason. Whats funny is AP on V9 (before this) really had me wondering if AP would pull me over.

This V9 drive-on-nav is definitely undercooked.

Now I do remain hopeful as this is something no car for sale today in the US does.

I'm really looking forwards to see where they make improvements.

To me this feels like it did when AP was initially released with AP2. Where the entire forum was expecting something so much better, but it wasn't nearly what we wanted.

I might try to develop a test for it. Where the test consist of multiple tries on a given route at various times.

Then come up with a scoring system for every function of it. I might have to involve other people to recreate specific situations where everyone is a willing participant.
 
L3 highway in good weather. 2021+. Yes. But dont confuse that with urban surface street environment.

Does Seattle perpetual rain count as good weather? I'm asking because that could delay the test by 9 months. Also, can you remove the + as that's cheating.

As to urban street environment I'm not entirely sure how that's relevant to L3. In any case I'm only talking about controlled highway/freeways with dividers. I'm also assuming that it will handle on-ramps to other freeways like drive-on-nav is supposed to handle. So it's exit to exit L3 driving in reasonably acceptable weather. (no snow, or heavy rain, sleet, etc).

We can setup a betting pool for those who wish to participate, and each significant region in the country will be part of the test (PNW, Southwest, Mid-West, North East, etc).

Maybe with the winnings going to autonomous driving related scholarship. So it's mostly for bragging rights, and something interesting to do.

The test will include an L2 drive-on-nav test, and FSD test (where we only test it in accordance to expectation for L3).

Maybe dates like these?

June 2019 - L2 Drive on Nav Test (assuming we have it without user confirmation then)
January 2020 - L2 Drive on Nav Test, and the first L3 test (if FSD is available)
June 2020 - L2 Drive on Nav Test, and L3 Test
January 2021- L3 Test and L4 Test
June 2021 - L3 Test and L4 Test

All bets will be placed 6 months before the test.
 
One thing I find interesting is the reaction to the initial post in this technically-oriented forum compared to similar posts in some of the more general audience forums. I've posted the exact same complaints in the general forums and gotten jumped all over by people implying that I'm the one with the problem. People bend over backwards to make excuses for the poor performance of the latest "features" (i.e. gimmicks). I don't understand it and it is becoming very old and counterproductive. I'm glad to see that technically-oriented people now are giving this iteration of Autopilot the real critical attention that it needs, not the attention driven by hype.
 
they did over 131K miles of such testing in CA public roads in 2017 vs. ZERO for Tesla.
Keep in mind the report does not include December so we'll learn soon enough Tesla did a bunch of miles during their aborted coast-to-coast video attempt.

I would not put too much weight into that since Tesla is super secretive on this matter for whatever reason (good or bad) and the information is incomplete. Similarly the ride-alongs are mostly not there (was not there this video of Elon in model3 driving with hands of steering wheel while giving an interview? Something he was then greatly criticized for).

So... this just highlights my point that car companies mostly depend on outside vendors to provide solutions to them.

Nissan website is super vague about what it is and when will come, but htey are big on how they are using NASA technology and such.

And so on, cannot research all of this now.
 
I would not put too much weight into that since Tesla is super secretive on this matter for whatever reason (good or bad) and the information is incomplete. Similarly the ride-alongs are mostly not there (was not there this video of Elon in model3 driving with hands of steering wheel while giving an interview? Something he was then greatly criticized for).
AFAIK, the video you refer to was simply auto-pilot w/no capabilities beyond what was deployed to customers, at the time. There wasn't any executing turns beyond lane changes, entering and exiting highways, going around roundabouts nor recognizing (and actioning upon) traffic lights and pedestrians.

Nissan tests fully autonomous ProPilot tech on Tokyo roads has another Nissan demonstration video. Was posted at Autonomous Car Progress. It shows turning at intersections, exiting and entering highways, going thru toll booths w/barriers, etc.
Nissan website is super vague about what it is and when will come, but htey are big on how they are using NASA technology and such.
As for NASA, at CES in Jan 2017 (I was there and at their booth), they did talk about what's at Seamless Autonomous Mobility: The Ultimate Nissan Intelligent Integration, an information center staffed by humans when an autonomous vehicle gets into a situation that it cannot and should not handle by itself. Whether or not this is just another idea and becomes productized (and can scale), who knows? You clearly can't have autonomous vehicles at highway speeds mixed w/other traffic coming to a halt when the car can't figure out what to do.

As for vague, you were probably looking in the wrong places. See Toward fully autonomous drive. Bloomberg - Are you a robot? matches up.
Nissan Motor Co. says it’s aiming to introduce fully autonomous cars to the market in 2022, as the Japanese carmaker strives to establish itself as a front-runner in driverless technologies.

The carmaker plans to add autonomous-driving functions step-by-step, first allowing more cars to handle single-lane driving by themselves, and subsequently navigate urban roads, including intersections, by 2020, said Yutaka Sanada, a Nissan senior vice president, at Bloomberg’s Year Ahead Asia conference in Jakarta.
Whether it arrives in those time frames, who knows?
 
Last edited:
I actually found that AP works significantly better in areas when nearby trees have leaves vs when they don't, likely due to the huge variations in shadowing.
AP could also be seeing the path of the road from where the trees are above the road in the distance, even around curves where you can't see the road itself. Also, maybe trees with leaves are easier to rule out as roadway than barren branches.
Great thread.

BTW, Teslas don't have to to run Tesla's software, do they?
This may be an opportunity for after market software... maybe there's a way to load from a USB and some secret left-right-up-down clicks of the thumb wheels ;)
Yes, but at what cost? It would be crazy expensive, and would need amazing lawyer shielding.
 
Last edited:
How many of the car companies you know are working on this? On the other hand who are the contenders? Intel (MobilEye), Waymo, Uber(?), ... none of them are car companies (those prefer to license ready made solutions anyway - it's safer this way too - somebody else is taking all the R&D risks! And the car companies can concentrate on their core strengths (or at least that's what PHBs/MBA would like you to believe as the good thing)). GM bought whoever does supercruise I think, but do they even aim at FSD ATM, any autonomous miles logged anywhere?
Besides GM w/their Cruise Automation, Nissan and Ford that I already mentioned and a chart by Navigant, Bloomberg: Who’s Winning the Self-Driving Car Race? has more players like Daimler, VW, BMW, Toyota, Volvo and Hyundai (already mentioned).
 
@wk057 I share many or your experiences with DoN and can also compare with my AP1 vehicle. I'm always excited to try out the feature, but agree it is not reliable (yet) and thus not useful. My litmus test is, "Would I suggest my wife use it to make driving easier." Mmm, no. (not yet) Still emphasize that I love getting the updates and seeing what the car can do (many successful automated freeway changes and exits).
I've been posting videos to my youtube channel (same name) showing the experience.

PS: Can I get your moded software on my 2015 S? :)

Specifically, I've experienced these limitations.
- random slow downs
- high sensitivity to phantom objects at overpasses (probably too much radar input?)
- occasional sketchy identification of vehicles in blind spot
- takes an exit thinking it is taking a freeway interchange
- very slow to assume freeway speed on the on ramp
- inconsistent suggestion for faster lane
 
  • Helpful
Reactions: TaoJones
@wk057 I share many or your experiences with DoN and can also compare with my AP1 vehicle. I'm always excited to try out the feature, but agree it is not reliable (yet) and thus not useful. My litmus test is, "Would I suggest my wife use it to make driving easier." Mmm, no. (not yet) Still emphasize that I love getting the updates and seeing what the car can do (many successful automated freeway changes and exits).
I've been posting videos to my youtube channel (same name) showing the experience.

PS: Can I get your moded software on my 2015 S? :)

Specifically, I've experienced these limitations.
- random slow downs
- high sensitivity to phantom objects at overpasses (probably too much radar input?)
- occasional sketchy identification of vehicles in blind spot
- takes an exit thinking it is taking a freeway interchange
- very slow to assume freeway speed on the on ramp
- inconsistent suggestion for faster lane

Ok if we're actually turning this into a can wk057 fix our problems wishlist:
1. Allow dumb cruise control when radar sensor is blocked with snow/salt/sand.
2. Allow disabling of sonics when there is a high degree of noise, i.e. snowy/salty/sandy conditions. Prevents false positive collision avoidance events. (yes, happens to me EVERY winter storm)
3. Move nags back to low confidence events instead of timers.
4. Preheat seats/steering only to max temp, no heater core.
5. Allow a 3rd "disable regen" function for slippery conditions, or immediately disable regen on yaw control event.
6. Allow dumb intermittent wipers, zero sensing. Optimize/adjustable speed/acceleration to prevent chatter.

Maybe Tesla will read this and take a hint.
 
And you can count on new readers coming to this forum and we can't reasonably expect them to do dui diilgence on each poster.


But we want to improve our discourse. Don't we? New posters should take the time to understand who they are speaking with/about, especially if they are going to go reflexively-negative ("go work somewhere else", etc.). Or should speak in much more guarded terms.

We all hate spammers, trollers, reflexively-negative posters, etc. They bring down the tone of the forum.

Your post seemed unnecessarily negative. You could have described the exact same experience of NavOnAP even pointing out its faults and areas for further improvement, without all that negativity. it would have been more effective at conveying the information both good and bad and it would have been more constructive. but not really a big deal. I do the same thing all the time but try to be better.

Humbly disagree. It was a well-written post backed with many observations. Emotion and opinion not only allowed but desired (at least by this reader).

But note that I still rated your post as informative because I thought you contributed facts and reasonable opinions (even where I disagree).

One man's fact-free, opinion-laden but hopefully not reflexively-negative reaction to your thoughts. :)

Thanks,
Alan
 
So far my experience is such that the NoA experience seems likely to depend greatly on your local traffic habits and levels. In light traffic it seems to be quite a nice start. In heavy traffic it is too timid to really interact well. Generally it feels like a very solid alpha version, much like AP2 autosteer did a year and a half ago. Hopefully improvements to this one come more quickly! The single update my car received has already shown improvement in the twerking behavior in the visualization.
 
  • Informative
Reactions: Cirrus MS100D
I like checklists. :p

Ok if we're actually turning this into a can wk057 fix our problems wishlist:
1. Allow dumb cruise control when radar sensor is blocked with snow/salt/sand.

I actually did this mod on my S a while back when I had a broken radar bracket that would eventually slide out of usefulness. The problem with it is that you get used to TACC and it limits your reaction time, even if you've been told by the car you're in dumb mode now. So, honestly, this is something that shouldn't be done. It actually is a problem resolved on the facelift cars, X, and 3... since the radar is protected.

2. Allow disabling of sonics when there is a high degree of noise, i.e. snowy/salty/sandy conditions. Prevents false positive collision avoidance events. (yes, happens to me EVERY winter storm)

This is a definite. Even happens to me in the rain. The car will start nudging to one side or the other thinking theres something to avoid. Ugh.

3. Move nags back to low confidence events instead of timers.

Better yet, just make it an option. Disabled, Confidence Based, Standard.... or something.

4. Preheat seats/steering only to max temp, no heater core.

Not quite sure why they don't just have the cold weather controls page directly in the app yet... this seems like a simple no-brainer.

5. Allow a 3rd "disable regen" function for slippery conditions, or immediately disable regen on yaw control event.

Definitely an issue in some conditions, although probably a harder problem to tackle than many. I think this is another problem that stems from the California-is-the-center-of-the-universe mentality.

6. Allow dumb intermittent wipers, zero sensing. Optimize/adjustable speed/acceleration to prevent chatter.
Maybe Tesla will read this and take a hint.

I think the RLS on AP1 works great, and they should put this back into production for AP2 cars. More options are better than less, though, in most cases.

---

As for my AP1 mods, I never really documented much of them anywhere because I never felt like I did everything that could be done... and also didn't have time to get back to the project. I did post some notes around here somewhere a while back, but I'll be damned if I have any idea where those posts are.

On a day-to-day basis, my AP1 mods are limited to automatic overtakes (and returns) on freeways, removing nags, truck-lust correction, and other minor improvements. These I use regularly.

Automatic overtakes work like this (super basic explanation):
  • Custom software monitors the environment as reported by the AP1 hardware
  • If the lead vehicle is traveling slower than my set speed, there is a travel lane to the left, there is not a slower vehicle visible in that lane, and ultrasonics are clear in that direction..... it starts a lane change process. This works as follows:
  • The car engages the turn signal and sounds a distinct tone over the speakers to note that, "Hey, I'm about to pass this guy... you should make sure all is clear."
  • The car does not initiate the lane change immediately. After the above, there is a full 3 second delay. The blinker remains on during this time.
  • If the human meat computer in the driver seat sees that the lane change won't be safe it can be aborted by tapping the blinker, disengaging autosteer or cruise, etc.
    • If aborted by the blinker, it will reset the state of things and eventually retry if conditions go unchanged... which is good for just delaying a maneuver for a courtesy or whatever.
  • If the lane change will be safe, no action is needed. The car will do the lane change. When complete it will disengage the blinker as well.
    • Ultrasonics are monitored during the change, just in case, as a minor safety thing... although the human meat computer should have checked anyway.
  • The software then monitors the car that's being overtaken.
    • Once the car goes out of view of the camera, it continues to extrapolate its position. It also uses the ultrasonics for some added confidence.
  • If other cars are ahead of the car being overtaken, those cars are essentially tagged as a car being overtaken unless they are traveling faster than or equal to my set speed.
  • Once it makes sense to return to the right, the car returns to the right.
    • A side benefit of the keep-right-pass-left logic is that it can result in overtakes on the right by default.

Note: Before people jump out with "This is so unsafe!"... remember, this is a level 2 driver assistance feature. The human meat computer must be aware and attentive. In no way would I say that the car should just blindly make a lane change with this setup. That said, it's extremely useful.

The result is that the car is usually able to seemlessly overtake a slower vehicle without ever decelerating.

The software will also recursively overtake another vehicle that appears ahead while already in an overtake maneuver, and eventually work to return to the original lane unless a number of factors make it clear that it shouldn't (like that target lane being exit-only or something). Lots of sanity checks. For example, a lane must exist for a full mile before being considered a target lane. It also checks nav data about the current road to double check if a lane is a ramp or not, when possible... although this data is pretty bad.

There is a lot of logic coded in... it makes the "thoughts" of the custom setup a bit more defined, mainly due to the predictive nature of the model used. It will continue to project vehicle positions for every vehicle ever encountered, until a max number of objects is encountered. It's kind of cool, although limited since it assumes a constant speed for the other vehicles it can no longer see. I also wrote an algo that predicts longitudinal movement of vehicles seen by the ultrasonics, which augments that data a bit and works okay. The result is that, in practice, rarely does it begin the process of an overtake/lane change while there is a vehicle in the way in the target lane. The only exception tends to be when moving back to the right near a merge... which I don't even think AP2 can handle as of yet, either.

I did try to program in courtesy lane changes for interchanges/merges, but just not enough data to work with to make it reliable.

Overall, it works pretty well considering the sensor limitations of the AP1 system. In most situations I use it in (long interstate drives) I can go whole stretches of interstate without ever touching a control in clear conditions.


That's it for things that are used regularly. Now, with my super pre-pre-alpha software/hardware, I've managed to make my AP1 S drive with near-zero interaction from my driveway to my old shop (6 miles, 6 stops/turns). I did this in a pretty hacky way, though, and probably not something I'd ever consider being usable as a real AP1 feature.

Basically I setup a database that monitored that route over dozens of human-driven passes and built up a pretty high resolution representation of the route using GPS data, reported lane positions at those coordinates, and other reported consistent factors with a particular location point. Through some clever software, I can use this data to augment the actions of AP1 on the route, including full stops and full turns. The result was surprisingly good, and I managed to do the route a few times with zero interaction. Most of the time, though, since AP1 can't see cross traffic, I'd have to disengage at intersections and reengage when clear.

I think this could be useful in some limited situations where the environment were closed/controlled, though, for moving the car autonomously... but probably not super useful in daily use for most. The benefits of the augmentation are amazing for general AP1 autosteer, though, and routes learned in the database are handled better overall than standard autosteer on either AP1 or AP2. So, most of the back roads in my town are well processed and autosteer works amazingly well on them while I use this mod, even without the full features of stopping/turning/etc.

Currently, this requires hardware I don't always keep hooked up, and haven't actually messed with for many months. But it's a great proof-of-concept.

The more interesting aspects of all of my mods.... they're things Tesla could implement without any hardware changes, since they could mod the actual software to handle it. In my case, I have to build on top of that with intercepts and external hardware... but no reason all of this can't be done in software on an AP1 car, from a technical perspective. Even the 6-year-old Tegra setup in MCU1 is plenty for running/storing my augment database setup.


The biggest problem with porting these mods to AP2 would be that the data produced by AP2 is super inconsistent. With AP1, I can drive a path at a constant speed, record the data, do it 10 more times, and compare the data of each pass only to find super minor variations in the information recorded about the path. With AP2, I logged some of this info and the same path run 10x might as well look like you were on another planet when each recording was made. So they definitely have some issues to work out to catch up to the quality of info produced by AP1.

I think the overtake stuff could be implemented on an AP2 car pretty easily, along with other improvements. The more advanced stuff would be limited by the noisy data produced by AP2, though.

Anyway, gotta run. Hope this clears up a little.
 
what do you mean?
I think they MAY mean that several years ago, the manufacturers put out statements that said they would take the vehicle liability for actions created or caused by the vehicles autonomous systems actions or failures. The German companies actually pioneered this nearly a decade ago - in their statements that they would "take responsibility" for the future autonomous systems. The potential blame or impact of a negative action by a vehicle simply cannot be allocated to the vehicle/drivers insurance company. The independent insurance providers can't possibly model or take on that risk, so the car providers (and today third party technology providers) are trying to quantify and indemnify some/all of it.
 
  • Helpful
Reactions: Boomer19
Not quite sure why they don't just have the cold weather controls page directly in the app yet... this seems like a simple no-brainer.
giphy.gif