Would agree with earlier comments; 50.2 is not very good, lane change is agressive and for the first time had to take control on multi lane highway as my Model S came really close to a pickup truck in the middle lane while I was in the left lane. I use AP a lot on the highway and never had that issue before.
-
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
I’ve seen the more aggressive lane changing(actually saw that in 50.1), but am not experiencing any of the other reported issues. If anything, I’m seeing the exact opposite: absolutely flawless autosteering on all freeways. That’s over on the order of 600 miles or so on 50.1/2. About 400 of those miles were 50.1 and the rest on 50.2.
Maybe @buttershrimp is right about calibration in these versions and my amount of driving just completed that really quickly?
Maybe @buttershrimp is right about calibration in these versions and my amount of driving just completed that really quickly?
My auto-steering on motorways has been solid on .50.2 too, but the steering wheel movements do feel a bit more aggressive and a little less reassuring. Haven't tried lane changes on .50.2 yet. Lane positioning feels slightly improved, seems there is less left-side bias showing...
As you can see, anecdotal comparisons based on memory are just that... anecdotal.
As you can see, anecdotal comparisons based on memory are just that... anecdotal.
I wonder how much of the satisfaction/dissatisfaction is geographic.
Seems happy 2017.50 drivers are on the west coast, more mixed eastward...from what I am seeing here.
FWIW, I still email tesla at [email protected] with Autopilot issues. I will give them software build information. Date, time, location information. And a description of the problems I had, and what I think is the problem. Suggest others do the same.
Seems happy 2017.50 drivers are on the west coast, more mixed eastward...from what I am seeing here.
FWIW, I still email tesla at [email protected] with Autopilot issues. I will give them software build information. Date, time, location information. And a description of the problems I had, and what I think is the problem. Suggest others do the same.
Drove from San Diego to Fresno for the 2nd time. First time was .42, second time .50.2. On the second trip I had to take over for AP2.5 numerous times. The first trip I didn't.
It also doesn't seem to be stopping for cars at a red light like it used to. I almost had a heart attack and had to slam the brakes the first time. Now I'm really alert. Overall it feels like a step back for me, but hopefully that is just a sign of progress as sometimes things get a little worse before they get a lot better!
It also doesn't seem to be stopping for cars at a red light like it used to. I almost had a heart attack and had to slam the brakes the first time. Now I'm really alert. Overall it feels like a step back for me, but hopefully that is just a sign of progress as sometimes things get a little worse before they get a lot better!
Strange. I was thinking it was geographic but my experience on 50.2 is all San Jose to San Diego, around San Diego itself, and up to Orange County and back. And my experience is the exact opposite of yours: with .42 I would have occasional times where I had to take over but with 50.2 I’ve had exactly 0(freeway driving). I’m on AP2, not 2.5, but it would seem odd that would be the difference.
Yeah that is odd. I drove from San Diego to Orange county a couple of times on .42 and only had one spot where it got a little sketchy. I might be going up soon and will compare with .50.2.
On my trip back from Fresno I had 2 occasions where the car just made a hard move across the lane line. Once there was a car there and I'm pretty sure if I hadn't taken over it would have it the car. First time it jumped to the left, second time to the right. Pretty sure both times were on curves.
Yeah, it's getting weird.... my firmware with 50.2 is getting better with every 100 miles. The bugs aren't going anywhere, but the car is much more impressive on the highway during high speed turns that used to auto slow too signficantly.
Here is a strange occurrence likely coincidental but who knows, today I attempted to pull off a similar thing to what @verygreen commented on when he heard about a guy on .50 making it though a roundabout. Though I’ve pulled off similar ideal path drives in the past, but today I tried something different. I used the Nav to set a destination on a road that was almost navigatable on previous firmware only to fail and require disengagements for the usual reasons. Note - It's not a roundabout though but a tough toll lane tunnel exit going INTO downtown Austin, kind of similar to the one seen in my hyper loop video (used to be funkytown on the website).... The interesting thing is the second time I tried this today ( much later in the day after a recharging near home wifi)... I did the same path on autopilot. This time, no navigation.... but following the same path.... the road dives under the road from the far left toll lane and back up on the right side of the highway where you promptly must exit.... Later in the day, 0 disengagements and 0 swerves as the road dips and turns...
Now, my conclusion is this: most likely this is a coincidence that I will test out again tomorrow. However, if the story is true about the roundabout from @verygreen's example, then my question for folks like @jimmy_d is whether or not the firmware could learn this quickly?Is it possible that the improvements on .50 as more many miles accumulate that is due to driving with navigation on? Can a NN be taught that quickly? As I type this I think.... placebo placebo placebo... but if it is true, I'm going to start setting my nav everywhere..... especially while driving on Austin's achilles heal for Autopilot, Austin (2222) between mopac and 360...
I wonder if other folks can try their difficult roads with the nav and conduct the same experiment?
I am not really sure what is it you did when and what happened when but NN is only used for visual stuff, and it's not learning "in the car".
Navigation on supposedly can help the car to know what lanes and turns to take, assuming suitable maps exist to guide the car on the selected route.
Navigation on supposedly can help the car to know what lanes and turns to take, assuming suitable maps exist to guide the car on the selected route.
That's not a small thing either then.... right? That would be a huge advance if the car actually behaved differently with nav on versus off. I keep assuming coincidence but the story of the person doing the roundabout is intriguing. If it assumed a fork in the road correctly because of navigation rather than 20% of the time it guesses going right with no lead car, it would be a very big deal.
Neural networks don’t typically change their weights(learn) at runtime. It’s an offline process that occurs on their servers somewhere. There are, of course, exceptions, but the kind of online learning you’re describing typically only happens in terms of customizing itself to a particular set of circumstances(e.g. a neural network for photography changing some parameters based on user input). I’m going with coincidence in your case.
Me too.... coincidence until proven tomorrow... I have the raw poorly handled footage.... I'll roll the clean improvement if it is consistently nailing it going forward.
Is it possible that my car could have learnt itself some NN during the long charging session at home via WiFi? It was creepy how the car did it correctly on the same firmware just 5 hours later . Again, I’m going with coincidence or time of day light and ideal shadow placement in this instance but still, it was really interesting after using Nav. Kind of jaw dropping, I need to try to film it today.
Not likely. In theory, it’s possible to run a forward pass(to predict outputs) and then also run a backwards pass(to update the network), but in practice, the learning rate would be infinitesimally small(learning rate diminishes over the course of training for modern neural networks) such that the relatively tiny number of samples it’s getting from your driving wouldn’t make much of a difference and, worse, and difference it did make would overfit to your particular circumstances.
Also, there’s the fact that @verygreen has, from what I understand, shown that only the vision/recognition algorithm is using the NN. That means it’s entirely a classification task, which requires ground-truth labels in order to train(of course, it can’t just learn from it’s own labels since that’s circular reasoning, and it’d always be “right”).
That said, there are forms of “online learning” that don’t involve actually changing the network. Usually that involves changing things like thresholds or other outside parameters to customize the network to a particular set of circumstances or a particular user. It’s possible some of that is going on here, though it’s hard to see how that would lead to the behavior you’re describing.
I really think there’s some sort of calibration that happens online. Definitely not changed to the NN but I do find that initially after some updates, turning precision and even lane recognition starts out a little flaky and as time goes on it visibly improves.
It could be so many factors — some sort of steering radius calibration, camera gains, or even map tiles downloading slowly in the background..... but the changes are mild in nature.
Similar threads
