-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
.48 feels like AP2 finally passed AP1
- Thread starter jimmy_d
- Start date
-
- Tags
- Autonomous Vehicles
Ok, back from driving all over CA for the holidays. Started out after Xmas feeling really bad about 50.2 after the first day showed what felt like degraded line hugging on sweeping curves, but I had a lot of driving ahead of me so I just kept at it and tried to explore the boundaries over the next 1500 odd miles of AP driving. In the end I feel pretty good about it and I have a new personal best for my SF to LA run. 3 forced disengagements this time which is the best ever (out of maybe 20 runs on that route with AP of some sort) and significantly better than 5, which is the best I ever got on AP1.
I’m not sure what was up with my first days on 50.2 - was it bad luck? Was there some kind of calibration going on? Or did I just get used to the new quirks really quickly?
As for that personal best: from my first AP1 experiences in late 2015 through to AP2 on 50.2 my benchmark has been how well AP (TACC + AS) works to make driving LA to SF or SF to LA more enjoyable. Of course enjoyable is pretty subjective and it includes familiarization and comfort with a system in addition to the raw performance of the system itself. Part of AP getting 'better' for me is certainly me getting comfortable with where it works well and where it doesn't so that I can let my guard down some without actually taking risks. But even so I can come up with some numbers and one number I use is the number of disengagements that I'm forced to perform because AP is doing something I'm not comfortable with. Early AP1 would be maybe a dozen disengagements for SF->LA if I was being brave and it gradually worked it's way down to 5 on SF->LA (which is a bit different from LA->SF). When I switched over to AP2 on .38 I was back up to a dozen. As I got used to the characteristics of AP2 and how it differed from AP1 I got more and more comfortable on the long stretches where it's really good but the number of forced disengagements was still over 10. And the software got better too - .38 through .42 on AP2 would wander even on straight roads and that gives the constant sense that the steering is uncertain and something weird could happen at any moment. Recent AP1 doesn't have that at all, so this was a big shock moving over to AP2 at first.
In using AP2 over the last few months I encounter 3 important regular failure modes and a couple of annoying behaviors that don't qualify as failures. The biggest annoyance was straight line wander which was pretty severe in 38, 40, and 42 for me but which is basically gone in 48 and 50. Other annoyances were truck-lust and phantom breaking which AP1 had right up until I stopped using it heavily in Oct and which AP2 was still showing when I moved over to it at version .38. But I haven't had a truck lust or phantom breaking event in 1500 miles on 48/50, which feels like a pretty substantial improvement - especially on the SF->LA run where I would always have at least one phantom breaking event in the past (SF->LA is 400 miles) and truck-lust during a passing maneuver seemed to happen every 500 miles or so.
The failures I describe as 1) hill cresting fail, 2) hugging the lane edge fail, and 3) intersection fail. Let me describe each of them briefly before describing how they've changed over time.
Hill cresting fail is when the car has a sudden unexpected steering event while approaching a crest in the road where the lane markers get lost over the crest. I almost always intervene with this happens because the steering events can be pretty large. Also, my first ‘holy cow’ moment on AP1 in 2015 was one of these and I’m still traumatized. This kind of fail can happen on roads that undulate over a flat stretch as well as in hilly areas but it requires a significant ripple since, as far as I can tell, it only happens when the camera temporarily loses sight of the lane lines in an area about 100 feet ahead of the car due to the road surface curving downward and out of sight.
Sadly, hill cresting fail hasn’t improved much since the early days of AP1, but AP1 seemed to have a compensating mechanism where it would use a lead vehicle to supplement it’s guess about the lane location, sometimes. You could see this happening when, in the UI, the lane lines would suddenly disappear but the lead vehicle would be colored blue, which seemed to indicate that the car was ‘tracking’ that lead vehicle. That would help reduce the number of events. As far as I can tell it’s very rare for AP2 to rely on the lead vehicle for lane guidance - I haven’t seen the blue lead vehicle once in 6500 miles of driving AP2. That has the nice effect that when you’re in stop and go traffic your car doesn’t wander out of the lane whenever the car ahead of you does, and you don’t get “follow me” lane changes where your car makes a lane change ‘following’ a lane change of the car ahead of you. Basically - AP2 seems to have a much better general awareness of it’s own lane and doesn’t need to rely on the lead car pretty much at all, but then it doesn’t have the lead car to help out in hill cresting events when the lane lines suddenly ‘disappear’. One of my 3 SF->LA forced disengagements was a hill crest on the start of the grapevine in a location where AP has failed on me every single time I’ve driven that route. And not coincidentally it’s the location where I had my first ‘holy cow’ on AP1 back in 2015.
Only 1 hill cresting fail on the southbound grapevine is a record for me. It’s a pretty tough road segment for cresting events, so maybe there’s finally some substantial improvement showing up for this serious failure mode.
Hugging the lane edge fail is really more of a slow steering correct issue than anything else, but it results in the car moving far out of the center of a lane on a long curve and staying close to the edge of the lane, or even leaving the lane. You get the sense that the car is reluctant to steer too aggressively to match the curve and it just ends up riding the lane boundary on sweeping curves. In practice I notice that this mainly happens on decreasing radius curves and, at least up until 42 the big fail would be S curves where you’d be getting to the second turn just as the first one was getting corrected and end up leaving the lane on the second turn. I saw S failures pretty regularly on the freeways in LA and SF on 38 and 40, a bit less on 42, and I haven’t seen one yet on 48 or 50. Hugging the edge is really sensitive to speed. I see it a lot on freeways at 75mph but rarely at 65 and never at 55. For me it only results in a fail in situations where there’s a vehicle in the adjacent lane and that vehicle also happens to be hugging the same lane line. If the spacing to the adjacent vehicle approaches 24 inches I will intervene. Hugging the edge is also really sensitive to lane width. Controlled access highway lanes are almost always over 12 feet, with 13 feet being the most common freeway width here in CA. In those situations hugging failures are pretty rare even at 75mph - I have yet to see one on 48/50 but had a couple of them on 38/42. But it’s pretty common to see situations where lane widths below 12 feet get used ‘temporarily’. Mostly I encounter these in areas where extensive highway improvement work is being done and lanes have been tightened up to accommodate the work. Often this lane adjustment can involve abrupt changes to lane location from one stretch of road to the next. All the hugging fails I have seen on 48/50 have happened in these situations. On my personal best SF->LA run two of the 3 fails were lane hugging in an improvement area at the tail end of the grapevine where lanes had been narrowed to 10 feet and the speed limit lowered to 55, though of course the traffic flow rate was still 75.
So basically, it looks like lane hugging behavior is getting a lot better recently, with AP2 finally getting back to where AP1 is. It might actually be better than AP1 at this point. AP1 always failed a couple of times in the improvement areas at the tail end of the grapevine though the sample size is so small that it’s hard to be sure.
The third failure mode I call intersection fail, and it happens when you have two roads intersect, a merger, split, or on/off ramp. In those situations the lane lines get complicated so boundary identification can get confused and choose the wrong path or change direction mid-intersection resulting in a poor trajectory. The most classic version of this is when driving in the rightmost lane of a highway and the steering gets very wonky just as the ramp separates from the main lane. In early AP1 it wasn’t uncommon to have the vehicle try to exit and even in later versions you would occasionally see the car try for the exit, then change direction at the last minute and veer back onto the highway. Which is alarming to say the least. This behavior has almost disappeared recently but in more challenging situations can still crop up. I hadn’t seen it at all on highways as of late AP1, it came back for me on AP2 version 38, but I haven’t seen it on 48/50. Most remarkable to me is that AP2 48/50 seems to have become very good with surface street intersections. AP1 seemed to only be reliable with surface street intersections if it had a well behaved lead vehicle and at least one high quality lane line and even then it would wander when the lane lines disappeared mid-intersection. AP2 seems quite unconcerned with the lead vehicle steering behavior and extrapolates the lane across the intersection quite nicely, though on 38 I saw it lose the lane a couple of times on local 4 lane business district roads and had to intervene once. I’m not a fan of AP on surface streets and I only use it there to explore what the car can do BTW. As of 48/50 AP2 seems to have fixed the intersections where 38/42 was failing and I haven’t seen a failure on any of the benchmark intersections I’ve been using.
In summary, recent experience for me seems to put recent AP2 ahead of AP1 generally, though there’s at least one situation that I encounter once a month or more where AP1 still outperforms AP2 (hill crest with lead vehicle present). Lane hugging is still less in AP1 but as of 48/50 it doesn’t seem to be a bigger problem for AP2 in terms of the number of actual fails, though subjectively it can be weaker than AP1 on roads that have high quality lane markers. In situations where the lane markers are poor quality or absent AP2 wins decisively. Intersection fails on highways seem to no longer be a serious problem, and on AP2 they don’t seem to be a problem even on surface streets so long as the lane configuration isn’t excessively vague..
Of course, getting good use out of AP has always had a big component of the driver learning to anticipate bad situation and avoid them. In early AP1 I would avoid the right lane on freeways because AP was weak with merge/exit lane intersections. Even in late AP1 I would avoid the right lane when there was more than mild traffic because it wasn’t well behaved around aggressively or incompetently merging traffic. On AP2 now I find those problems to be not worthy of concern but I still watch for cresting roads, S curves, decreasing radius turns, and construction zones. This has all been a small price to pay for being able to relax and enjoy the scenery on the endless miles of rural highway between SF and LA. Now I seem to be in a place where, aside from the 5% of the route that runs through the mountains north of LA, the car can handle the rest of the route without any disengagements.
I look forward to the day, hopefully soon, where the car runs the entire route with high confidence and no interventions needed.
I’m not sure what was up with my first days on 50.2 - was it bad luck? Was there some kind of calibration going on? Or did I just get used to the new quirks really quickly?
As for that personal best: from my first AP1 experiences in late 2015 through to AP2 on 50.2 my benchmark has been how well AP (TACC + AS) works to make driving LA to SF or SF to LA more enjoyable. Of course enjoyable is pretty subjective and it includes familiarization and comfort with a system in addition to the raw performance of the system itself. Part of AP getting 'better' for me is certainly me getting comfortable with where it works well and where it doesn't so that I can let my guard down some without actually taking risks. But even so I can come up with some numbers and one number I use is the number of disengagements that I'm forced to perform because AP is doing something I'm not comfortable with. Early AP1 would be maybe a dozen disengagements for SF->LA if I was being brave and it gradually worked it's way down to 5 on SF->LA (which is a bit different from LA->SF). When I switched over to AP2 on .38 I was back up to a dozen. As I got used to the characteristics of AP2 and how it differed from AP1 I got more and more comfortable on the long stretches where it's really good but the number of forced disengagements was still over 10. And the software got better too - .38 through .42 on AP2 would wander even on straight roads and that gives the constant sense that the steering is uncertain and something weird could happen at any moment. Recent AP1 doesn't have that at all, so this was a big shock moving over to AP2 at first.
In using AP2 over the last few months I encounter 3 important regular failure modes and a couple of annoying behaviors that don't qualify as failures. The biggest annoyance was straight line wander which was pretty severe in 38, 40, and 42 for me but which is basically gone in 48 and 50. Other annoyances were truck-lust and phantom breaking which AP1 had right up until I stopped using it heavily in Oct and which AP2 was still showing when I moved over to it at version .38. But I haven't had a truck lust or phantom breaking event in 1500 miles on 48/50, which feels like a pretty substantial improvement - especially on the SF->LA run where I would always have at least one phantom breaking event in the past (SF->LA is 400 miles) and truck-lust during a passing maneuver seemed to happen every 500 miles or so.
The failures I describe as 1) hill cresting fail, 2) hugging the lane edge fail, and 3) intersection fail. Let me describe each of them briefly before describing how they've changed over time.
Hill cresting fail is when the car has a sudden unexpected steering event while approaching a crest in the road where the lane markers get lost over the crest. I almost always intervene with this happens because the steering events can be pretty large. Also, my first ‘holy cow’ moment on AP1 in 2015 was one of these and I’m still traumatized. This kind of fail can happen on roads that undulate over a flat stretch as well as in hilly areas but it requires a significant ripple since, as far as I can tell, it only happens when the camera temporarily loses sight of the lane lines in an area about 100 feet ahead of the car due to the road surface curving downward and out of sight.
Sadly, hill cresting fail hasn’t improved much since the early days of AP1, but AP1 seemed to have a compensating mechanism where it would use a lead vehicle to supplement it’s guess about the lane location, sometimes. You could see this happening when, in the UI, the lane lines would suddenly disappear but the lead vehicle would be colored blue, which seemed to indicate that the car was ‘tracking’ that lead vehicle. That would help reduce the number of events. As far as I can tell it’s very rare for AP2 to rely on the lead vehicle for lane guidance - I haven’t seen the blue lead vehicle once in 6500 miles of driving AP2. That has the nice effect that when you’re in stop and go traffic your car doesn’t wander out of the lane whenever the car ahead of you does, and you don’t get “follow me” lane changes where your car makes a lane change ‘following’ a lane change of the car ahead of you. Basically - AP2 seems to have a much better general awareness of it’s own lane and doesn’t need to rely on the lead car pretty much at all, but then it doesn’t have the lead car to help out in hill cresting events when the lane lines suddenly ‘disappear’. One of my 3 SF->LA forced disengagements was a hill crest on the start of the grapevine in a location where AP has failed on me every single time I’ve driven that route. And not coincidentally it’s the location where I had my first ‘holy cow’ on AP1 back in 2015.
Only 1 hill cresting fail on the southbound grapevine is a record for me. It’s a pretty tough road segment for cresting events, so maybe there’s finally some substantial improvement showing up for this serious failure mode.
Hugging the lane edge fail is really more of a slow steering correct issue than anything else, but it results in the car moving far out of the center of a lane on a long curve and staying close to the edge of the lane, or even leaving the lane. You get the sense that the car is reluctant to steer too aggressively to match the curve and it just ends up riding the lane boundary on sweeping curves. In practice I notice that this mainly happens on decreasing radius curves and, at least up until 42 the big fail would be S curves where you’d be getting to the second turn just as the first one was getting corrected and end up leaving the lane on the second turn. I saw S failures pretty regularly on the freeways in LA and SF on 38 and 40, a bit less on 42, and I haven’t seen one yet on 48 or 50. Hugging the edge is really sensitive to speed. I see it a lot on freeways at 75mph but rarely at 65 and never at 55. For me it only results in a fail in situations where there’s a vehicle in the adjacent lane and that vehicle also happens to be hugging the same lane line. If the spacing to the adjacent vehicle approaches 24 inches I will intervene. Hugging the edge is also really sensitive to lane width. Controlled access highway lanes are almost always over 12 feet, with 13 feet being the most common freeway width here in CA. In those situations hugging failures are pretty rare even at 75mph - I have yet to see one on 48/50 but had a couple of them on 38/42. But it’s pretty common to see situations where lane widths below 12 feet get used ‘temporarily’. Mostly I encounter these in areas where extensive highway improvement work is being done and lanes have been tightened up to accommodate the work. Often this lane adjustment can involve abrupt changes to lane location from one stretch of road to the next. All the hugging fails I have seen on 48/50 have happened in these situations. On my personal best SF->LA run two of the 3 fails were lane hugging in an improvement area at the tail end of the grapevine where lanes had been narrowed to 10 feet and the speed limit lowered to 55, though of course the traffic flow rate was still 75.
So basically, it looks like lane hugging behavior is getting a lot better recently, with AP2 finally getting back to where AP1 is. It might actually be better than AP1 at this point. AP1 always failed a couple of times in the improvement areas at the tail end of the grapevine though the sample size is so small that it’s hard to be sure.
The third failure mode I call intersection fail, and it happens when you have two roads intersect, a merger, split, or on/off ramp. In those situations the lane lines get complicated so boundary identification can get confused and choose the wrong path or change direction mid-intersection resulting in a poor trajectory. The most classic version of this is when driving in the rightmost lane of a highway and the steering gets very wonky just as the ramp separates from the main lane. In early AP1 it wasn’t uncommon to have the vehicle try to exit and even in later versions you would occasionally see the car try for the exit, then change direction at the last minute and veer back onto the highway. Which is alarming to say the least. This behavior has almost disappeared recently but in more challenging situations can still crop up. I hadn’t seen it at all on highways as of late AP1, it came back for me on AP2 version 38, but I haven’t seen it on 48/50. Most remarkable to me is that AP2 48/50 seems to have become very good with surface street intersections. AP1 seemed to only be reliable with surface street intersections if it had a well behaved lead vehicle and at least one high quality lane line and even then it would wander when the lane lines disappeared mid-intersection. AP2 seems quite unconcerned with the lead vehicle steering behavior and extrapolates the lane across the intersection quite nicely, though on 38 I saw it lose the lane a couple of times on local 4 lane business district roads and had to intervene once. I’m not a fan of AP on surface streets and I only use it there to explore what the car can do BTW. As of 48/50 AP2 seems to have fixed the intersections where 38/42 was failing and I haven’t seen a failure on any of the benchmark intersections I’ve been using.
In summary, recent experience for me seems to put recent AP2 ahead of AP1 generally, though there’s at least one situation that I encounter once a month or more where AP1 still outperforms AP2 (hill crest with lead vehicle present). Lane hugging is still less in AP1 but as of 48/50 it doesn’t seem to be a bigger problem for AP2 in terms of the number of actual fails, though subjectively it can be weaker than AP1 on roads that have high quality lane markers. In situations where the lane markers are poor quality or absent AP2 wins decisively. Intersection fails on highways seem to no longer be a serious problem, and on AP2 they don’t seem to be a problem even on surface streets so long as the lane configuration isn’t excessively vague..
Of course, getting good use out of AP has always had a big component of the driver learning to anticipate bad situation and avoid them. In early AP1 I would avoid the right lane on freeways because AP was weak with merge/exit lane intersections. Even in late AP1 I would avoid the right lane when there was more than mild traffic because it wasn’t well behaved around aggressively or incompetently merging traffic. On AP2 now I find those problems to be not worthy of concern but I still watch for cresting roads, S curves, decreasing radius turns, and construction zones. This has all been a small price to pay for being able to relax and enjoy the scenery on the endless miles of rural highway between SF and LA. Now I seem to be in a place where, aside from the 5% of the route that runs through the mountains north of LA, the car can handle the rest of the route without any disengagements.
I look forward to the day, hopefully soon, where the car runs the entire route with high confidence and no interventions needed.
Hmm maybe I just live on the edge more, but I made the entire drive from San Diego to San Jose with exactly 1 forced disengagement on 50.2(in northern OC, the car started following some old lane markers). The only other times I took over manually were getting off onto local streets for supercharging/arriving at destination and the local section of the Pancheco pass(as you come into Gilroy).
That could be. People seem to vary a lot in their AP use and comfort level. I think it can be really had to compare experiences in that sense.
That said, I'm really looking forward to using 50 on my next run north. Going north I generally get better performance than south and I think I might be able to do LA->SF with no disengagements on this next run. I generally take the 5 all the way to the 508 rather than going across to Gilroy, so I haven't got any experience on that leg, but it's nice to hear that it's working for you. I had wondered how that segment might go because of the curves. How fast do you drive it?
I used 50 to head north just prior to Xmas, but on that trip I went up the 101 so it wasn't a benchmark run for me. On that drive I did SD to Cambria without any disengagements, and then drove Cambria to SF going up the peninsula to SF and I think I only had one disengagement on the mountain roads going across from the 1 back to the 101 - lane hugging disengagement if I remember correctly. Using it through San Jose and on up the peninsula to SF was awesome. It was basically perfect on that leg of that trip.
50.2, as I've stated repeatedly, failed horribly at lane keeping on curves (of any kind). Basically it was lazy and would slowly fail to turn the wheel enough and drift past the lane on the outside of each curve.
50.2 did steadily improve (somehow??) but it was never confident and often oscillated oddly and otherwise did bizarre things (like losing the lane lines on a bridge with clear markings and trying to go into a bike lane).
50.3 is lights out. Wow. No complaints so far. Its like they made a mistake with 50.2 (used NN from pre .40) to rush it out for X-Mas and fixed it with 50.3 (duh, .42 like NN and behavior).
.50.3 will do a lot more curved intersections and will even do very very sharp curves (15mph listed) at like 3-5mph (slams on brakes at start of sharp curve and does it granny style with somewhat halting jerks of the wheel but subsequent runs are going faster and smoother). Even curved intersections it made 50% of the time are now 75% and smoother and less scary. Clearly I only do these tests when there is no one to the side or just behind or in front of me because when it fails it dives across the intersection and usually into a neighboring lane though one time it tried to turn into a perpendicular street (turned too sharply to the right). Its not pleasant but its cool watching the system evolve its capabilities and it seems to be learning within FW for me on these intersections (likely the mapping data).
If its not capable of making it, it doesn't learn anything (since you've gotta disengage) but if it makes it, even poorly, it has been continuing to improve for me.
Can do this curve now.
50.2 did steadily improve (somehow??) but it was never confident and often oscillated oddly and otherwise did bizarre things (like losing the lane lines on a bridge with clear markings and trying to go into a bike lane).
50.3 is lights out. Wow. No complaints so far. Its like they made a mistake with 50.2 (used NN from pre .40) to rush it out for X-Mas and fixed it with 50.3 (duh, .42 like NN and behavior).
.50.3 will do a lot more curved intersections and will even do very very sharp curves (15mph listed) at like 3-5mph (slams on brakes at start of sharp curve and does it granny style with somewhat halting jerks of the wheel but subsequent runs are going faster and smoother). Even curved intersections it made 50% of the time are now 75% and smoother and less scary. Clearly I only do these tests when there is no one to the side or just behind or in front of me because when it fails it dives across the intersection and usually into a neighboring lane though one time it tried to turn into a perpendicular street (turned too sharply to the right). Its not pleasant but its cool watching the system evolve its capabilities and it seems to be learning within FW for me on these intersections (likely the mapping data).
If its not capable of making it, it doesn't learn anything (since you've gotta disengage) but if it makes it, even poorly, it has been continuing to improve for me.
Can do this curve now.
Last edited:
That's really interesting about 50.2 improving for you. I get the same impression but I'm suspicious about whether I might be extrapolating from a couple of bad experiences. I mean, generally I'm extrapolating from a few bad experiences because in real number terms I don't see enough failures to be statistical, but when I'm trying to draw conclusions from two events it seems especially weak. When I get used to a failure mode and can see a failure might be coming I tend to remember it in less vivid terms than when it comes out of the blue and I'm caught by surprise. My initial bad experience with 50.2 was unexpected.
At what kind of driving speed do you usually seen lane hugging fails? I was seeing them mainly in moderately heavy traffic that was nonetheless traveling quite fast, which happens a lot in LA but pretty much never in SF. I did see some gross failures on wide open roads though.
Hearing about 50.3 from you makes me really interested in trying it out. I haven't received it yet.
50.3 is the real deal. You'll love it.
I had failures between 65mph and 75mph.
Several other people also reported sub par lane keeping with 50.2 that got better as they drove. Something was going on but sample size definitely makes it hard to really get to the bottom of why it was so bad and improved for so many.
JasJ
Member
50.3 is a big step back for me (from 48). Car (AP2) has always been center of lane for most of the last year, now with 50.3 the car loves the left side of the lane in a perfectly straight road. Doesn't matter if its a left lane, middle lane, right lane of a multi-lane highway with or without large vehicles on the left and/or right. Car now drives less than a foot from the left lane marker most of the time. My disengagements have gone from mostly Zero on Houston freeways, to several times a minute in busy traffic.
I did slow the car down to the speed limit coming down Pencheco pass, but at that speed, it handled the curves perfectly.
Joe F
Disruption is hard.
Sweeping curves for me on 50.2 and 50.3 are much worse than my jump from .44. I called it outside curve lust because at one point as I was dangerously close to a tractor trailer going through a curve, but it wasn't due to truck lust. The higher the speed, the more difficult the car has keeping centered.
Yeah. I was playing around with the speed quite a bit on this last run and it had a big effect on lane hugging. I'm pretty laid back on my road trips - never go over 75 and that seems to be plenty enough to create problems on curvy roads. I notice a lot of the traffic runs 80 or 85 by default and I was thinking that AP 50 was probably not going to work at those speeds on anything that has much in the way of curves.
Thanks @jimmy_d for your write up. Coincidentally I had one of those “blue” car moments on AP1 today.
This AM on my commute, my Tesla lost the lane markers due to the sun and a crappy highway. It started tracking a car in front. We were all going at speed (~70). We were in the middle lane.
A car passed me, and was “blocked” by another car in the left lane. That driver crossed in front of me, and for a moment that car turned blue [I still did not have lane markers... just the “grayed out” pseudo markers that approximate a lane]. I was waiting for my car to follow the crossing car into the right lane... which it did not.
After that car took off, my car resumed following the original “blue” car in front. So it looks like the software has corrected this problem now. Yes I still remember this problem from the early days! AP1 on 50.2
Interestingly the car didn’t even quiver or shift position as the other car moved across.
This AM on my commute, my Tesla lost the lane markers due to the sun and a crappy highway. It started tracking a car in front. We were all going at speed (~70). We were in the middle lane.
A car passed me, and was “blocked” by another car in the left lane. That driver crossed in front of me, and for a moment that car turned blue [I still did not have lane markers... just the “grayed out” pseudo markers that approximate a lane]. I was waiting for my car to follow the crossing car into the right lane... which it did not.
After that car took off, my car resumed following the original “blue” car in front. So it looks like the software has corrected this problem now. Yes I still remember this problem from the early days! AP1 on 50.2
Interestingly the car didn’t even quiver or shift position as the other car moved across.
Tjhappel
Member
Really weird how it’s differemt for all of us. 50.2 for me was just awful for 2-3 days then got ok but not before than .48 at all, them 50.3 hit... full game changer for me. There was no adjustment time needed for my car on this update, just killed it right out of the gate. It’s been driving freeway sections ive never made before repeatedly this week. Super happy with the progress and really can’t wait for “the big” jump we keep hearing about!
AP2 model x. Also hill cresting was greatly improved for me.
Thanks for the good info and posts guys.
AP2 model x. Also hill cresting was greatly improved for me.
Thanks for the good info and posts guys.
Thank you for a healthy reminder that AP1 evolves too and we need to keep that in mind in our comparisons.
Is there any correalation between the location (California area) and the people reporting strong performance of EAP on 50.3?
Could the improvment be related to the new autopilot map tiles discussed in this thread; Tesla Autopilot maps
I went trough 50.1, 50,2 and noticed 50.2 was very erratic to start with, but improved over time, calibration period needed I guess.
50.3 seems more consistent, but i still have some serious lane hugging in the sharper curves.
To me 50.3 is an improvement, but only an incremental one for me.
Could the improvment be related to the new autopilot map tiles discussed in this thread; Tesla Autopilot maps
I went trough 50.1, 50,2 and noticed 50.2 was very erratic to start with, but improved over time, calibration period needed I guess.
50.3 seems more consistent, but i still have some serious lane hugging in the sharper curves.
To me 50.3 is an improvement, but only an incremental one for me.
I am suspecting the same. At least with original roll out of 2017.50.2, it seemed that California/west coast posters were reporting signficant improvement. I noticed more dissatisfaction with folks from Chicago, Midwest, Texas, etc... as time goes on, they seem to be seeing improvement as well.
Yes, I'm in Chicago and 50.2 was a disaster. Driving to Midway Airport last night to get my wife, I noticed some lazy turning on sweeping curves (approached outside of each curve but, unlike 50.2, did not cross the lines but did get uncomfortably close and if someone was next to me it would've forced a disengagement). I was going 75mph and I was on 90/94 just like my prior failures with 50.2.
It's possible but I'm not sure it's maps related, maybe just tuning of control loops or vision that coincidentally favors certain kinds of roads?
I just looked at the production map tiles in San Jose and nothing's changed: na20170502: 2017-06-30 10:05
We're locked out of the development ones at this point, but there has been no change to the maps in recent months pushed out to customer cars. Of course it's possible that the algorithm could've changed on the car such that it's utilizing existing maps better?
daud
Member
Hadn't seen "the blue car" myself using AP2 either. Interestingly, this AP2 (50.3) got thrown off by what on the IC looks like another car crossing its lane and it briefly visualised that car in blue: youtube.com/watch?v=lJwp__0T2nI&t=22m40s
More subject matter from that video: duct taping a few sensors (and side camera’s - although, not relevant here yet, rendering the car’s right-hand side blind) doesn’t disable auto lane change. More like the opposite: allows for more dangerous AP behaviour.
Similar threads
