-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Autopilot is already improving.
- Thread starter mobe
- Start date
I have had similar experiences. I have over 1500 miles with autopilot so far. My commute is about 130 miles total. 90% of it is on the highway with autopilot on. There are days where it seems that autopilot is improving in certain areas. However, it also seems to struggle on those same areas on different days. I've done the commute in day and night, in relatively stable weather conditions so far. If autopilot is improving for other Tesla owners, it certainly does not seem to be improving consistently for me.
sillydriver
Member
Here is an anti-fail report: AP worked really well. On 10/28 I drove on the interstate in a literally blinding rainstorm, so that some cars in the right lane were using their emergency blinkers. I was in the middle of three lanes moving too fast for the conditions in order to keep pace with light-moderate traffic. It was late afternoon with flat gray light. I didn't even consider trying AP for the first several miles because of the conditions, but noticed the dash display showed it was still seeing the lane markings. I could barely see the lane markings with the amount of rain on the road. So I tried it for another dozen miles with both hands full on the wheel but not resisting the steering movements. It worked! It was definitely seeing the lanes, since I was too far behind the car ahead for AP to follow it. In fairness, because I was in the middle lane it didn't need to deal with the lane markings around exits, which I don't think could have been seen at all.
RationalOptimist
Member
Quick Q. Amidst the impressive reports of AP learning, have we ruled out the possibility that it is simply learning locally, not from the fleet? e.g. might there be an algorithm that says "if driver corrects at GPS coordinates X,Y, then switch to tracking other lane at that point." We know Tesla thinks a lot about people's regular commute, and isn't it possible that it's simply optimizing to avoid repeated driver take-overs? To test this, someone who's seen improvement on a particular road should invite a friend to try that same road for the first time. (Note: I v much want to buy the daily-improvement from data from the entire fleet story, would just like to hear from someone who's ruled out this other possibility.)
I would expect it to be learning both ways. The way to test this would be to drive the same route (where it veered to the exit ramp for example) several times in a row. Because the fleet information downloads daily (from what others have posted) there should be no improvement on subsequent runs. If it updates locally then each run should get better.
Sorry for the confusion. The slowing down is seen using 7.0 with Auto Steer engaged; the speed variations are very odd and quite repeatable. I haven't confirmed the behavior with TACC alone under 7.0. Prior to receiving 7.0, TACC had no trouble holding speed through that section.
Andyw2100
Well-Known Member
Update: I tested TACC-only speed control on this same stretch of roadway today. I observed the same speed variations even though I was doing all the steering. It's just weird. Particularly since it holds speed beautifully when driving the same stretch in the other direction.
Last edited:
Cyclone
Cyclonic Member ((.oO))
sillydriver
Member
Good question. We know that the software is running even when not engaged: the display on the driver's console is continuously updating the AP's understanding of the situation. I’ll have to try it again with Autosteer turned off in Settings.
One other note: the problem occurs only in one direction; there is no speed variation observed driving the same stretch in the other direction.
_____________
Update: Even with Autosteer turned off in the Settings, TACC will not hold a set speed through this particular stretch of road driving eastbound; westbound, it's flawless. Furthermore, the odd speed variations form a pattern that repeats almost identically from run to run. There has been no learning that I can detect.
Last edited:
BriansTesla
Old school meets new tech
Non AP car here but wanting to contribute to the discussion. Since my MS learns from me where to raise the suspension using GPS coordinates, couldn't the Autopilot use the same type of data to learn how to adjust steering?
Example: At XY coordinates going direction Y, driver corrected Z so adjust accordingly. Seems simple in principle, probably isn't so much in execution.
Example: At XY coordinates going direction Y, driver corrected Z so adjust accordingly. Seems simple in principle, probably isn't so much in execution.
Difference is that's just a GPS trigger with a single output (suspension height). That's not machine learning - it's a setting. The learning part means taking knowledge from what the driver does given a set of conditions, and interpreting that to what the car should do given a similar set of conditions. As I've mentioned a few times upthread, generalizing is the most important aspect of this type of problem.
I wish my MS would learn to stop changing my suspension when I drive past a street on which I have it set to raise suspension.. :wink:
BriansTesla
Old school meets new tech
Of course we know the Autopilot is doing this learning. My thinking is that some of the work could be done at the vehicle level. You know, avoid the pothole in my driveway, things like that.
Khatsalano
Member
Drove from SF to LA today, almost entirely on AP. I can really feel that the concentration of Teslas in LA is not like the Bay Area because the self-learning corrections are much sparser here. The car does really weird stuff like hard-duck right when the lines are clear and there is no reason to swerve like that. Back in the Bay, this was the behavior on Day 1, but now, those are all gone.
Back in the Bay Area, I feel like I work with the AP as a team; human and machine working towards the same goal. Today, it was incredibly stressful to drive on I-5 S from Burbank to Anaheim through LA because I felt like I was constantly fighting the AP on every lane bias, every curve, and at times, these hard ducks that required a manual rescue. I guess we need more training down here, or as the Electrek article puts it, a tug on the AP leash.
- K
Back in the Bay Area, I feel like I work with the AP as a team; human and machine working towards the same goal. Today, it was incredibly stressful to drive on I-5 S from Burbank to Anaheim through LA because I felt like I was constantly fighting the AP on every lane bias, every curve, and at times, these hard ducks that required a manual rescue. I guess we need more training down here, or as the Electrek article puts it, a tug on the AP leash.
- K
beeeerock
Active Member
The resolution of the GPS, especially at speed, is unlikely to be tight enough to actually safely follow a GPS track (if that's what you're suggesting - not sure). But I suspect that the system does learn that "the left lane edge paint anomaly at (approximate Northing and Easting) northbound should be disregarded - go straight and don't deke left". And perhaps, "it turns out the right lane appears at N: E: is a climbing lane and not a right turn lane - safe to follow fog line".
Where I am, I suspect much of my travelled lane km's have been tracked by me alone, or mostly me. Since the summer ended and especially since version 7 dropped, I haven't seen another Tesla on the highway I travel most often. Obviously, I'm not out there 24-7, but Tesla density isn't overly high in this neck of the woods! So I suspect much of the 'training' I've been seeing has largely been from my car, maybe a pass or two by another small number of Teslas. On one day early in using 7.0 when I had some slack in my schedule, I actually turned around and tried problem spots again to see if I could reproduce the error. I generally couldn't and I was left with the impression that AP learns very quickly! However, there have been a few occasions since that time where previously improved behaviour has regressed back to what I saw the first few times through. But it seemed to learn the correct behaviour again. Not sure how or why that is happening!
Given that the system is apparently active in that it shows the lanes on the display, I have to assume that it's watching how I drive and comparing my behaviour with what it would have done if AP was actively in use. Much like the theory that REM sleep (or the type of sleep where you're paralyzed - I'm hardly an expert on sleep!) is a good time to have dreams that test scenarios - falling off the cliff etc. - because you can't damage yourself, I suspect that AP is doing the same. Not able to act dangerously, but able to test scenarios in a dream state! It would be interesting to know how or if AP's 'watching' differs from a drivers 'panic corrections' with respect to how the learning occurs. This is likely something Ohmman could comment on based on his experience?
I made a video of the problem here: TACC with 7.0 - unexpected slowdowns
It seems that the built in sat nav plays a role, this has been confirmed by Tesla. It sounds like you're experiencing the same behaviour.
Soolim
Member
Are you implying that AP learning is close to real time and that quick? What is your theory on how the group leaning works?
Similar threads
