Please complain to Tesla via your preferred route. The more people that complain, the faster they'll fix it. If we don't complain, they'll just figure that we'll get used to it and leave it as it is.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
"Fleet Speed" (mis-)feature in 2019.8.6+
- Thread starter rnortman
- Start date
Please complain to Tesla via your preferred route. The more people that complain, the faster they'll fix it. If we don't complain, they'll just figure that we'll get used to it and leave it as it is.
is this bs also on 2019.12.1?
On 2019.8.5 I already found disruptive the fact that the reduction from like 80 to 45 while taking an exit is sometimes borderline dangerous, since most ramps the drop in speed is not required to be that drastic and other cars get really pissed at you since it's almost like brake checking people...
They need to fix this...
On 2019.8.5 I already found disruptive the fact that the reduction from like 80 to 45 while taking an exit is sometimes borderline dangerous, since most ramps the drop in speed is not required to be that drastic and other cars get really pissed at you since it's almost like brake checking people...
They need to fix this...
I assume so, based on the fact that my communications with my service center and customer support are that I have to wait for a software fix. If they have a software fix already and just aren't giving it to me I'll be annoyed.
Please write to both your local service center and Tesla CS. Emphasize that cruise control, unlike the rest of Autopilot, is not in beta and that this is a defect in cruise control which they are obligated to fix since they broke it. Not that this is getting me very far but the more people who complain forcefully the more attention they'll pay to it -- and hopefully they will think twice before breaking non-beta features in the future without any way to disable or roll back the change.
Last week, Musk pointed out the flaws in using HD maps to control autonomous driving, since the recorded maps may not match actual road conditions, and that the only strategy that would really work is for the system should respond to current conditions, similar to how a human does today (using a combination of Tesla's radar, cameras and proximity sensors).
"Fleet Speed" has the same problems as HD maps. That data is based on past conditions - not the current conditions.
Yesterday during a 30 minute drive on one interstate that had recently completed construction (adding new lanes) and a toll road currently under construction (also adding lanes), I had 4 to 5 instances of rapid braking in the middle of heavy highway speed traffic. This was entirely due to incorrect speed limit adjustments, which sometimes would abruptly lower the TACC speed and sometimes lower speed without adjusting TACC.
If Tesla is going to use "Fleet Speed" for automatic speed adjustments, they should implement several improvements ASAP:
While it seems like a great idea to use "fleet learning" to better control vehicle speed, the current implementation has some flaws which should be addressed - soon...
"Fleet Speed" has the same problems as HD maps. That data is based on past conditions - not the current conditions.
Yesterday during a 30 minute drive on one interstate that had recently completed construction (adding new lanes) and a toll road currently under construction (also adding lanes), I had 4 to 5 instances of rapid braking in the middle of heavy highway speed traffic. This was entirely due to incorrect speed limit adjustments, which sometimes would abruptly lower the TACC speed and sometimes lower speed without adjusting TACC.
If Tesla is going to use "Fleet Speed" for automatic speed adjustments, they should implement several improvements ASAP:
- Provide an option for audible and/or visual notification that a Flee Speed adjustment is coming up (like a lane change)
- Fleet Speed should be provided in real-time from Tesla's cloud server, based on data which is updated to reflect current road conditions (based on the previous Tesla vehicles who just drove that section of road)
- Fleet Speed should be recorded only when the vehicle is operating at or above the posted speed limit, and there aren't any vehicles ahead causing the speed to be reduced
- And, there should be an option for disabling automatic speed adjustment
While it seems like a great idea to use "fleet learning" to better control vehicle speed, the current implementation has some flaws which should be addressed - soon...
Yeah, this is the sort of variability I see. There is one spot in my commute where it consistently -- almost every time -- drops the speed. But the way it drops the speed is not consisitent -- sometimes the TACC set speed drops, sometimes the speed limit drops but set speed is the same, sometimes the speed limit display disappears completely. It's a little different every time.
Fantastic idea, though I'm not sure they are capable of doing this in the current code. It would need to know in advance that you are going to take that exit, even if you aren't using NOA (because this "feature" seems to operate whether you are using NOA or not, and whether you are following a Navigation route or not).
Terrible idea -- the sample size is way too small. If that guy who drove through there 5 minutes ago had to slam on the brakes for some transient condition like a small traffic knot that has since cleared up, should they perpetuate that event to you? And then the next guy? Causing more transient slowdowns because any time one car hits the brakes in dense traffic it has the potential to create a traffic jam where there was not one before.
The fact is, there is not enough data to do this based on historical conditions, especially if you limit to recent history (like the past 30 minutes). The only way they can make this feature work is by doing what you alluded to at the beginning of your post -- actually make the system as smart as a person -- make it use its sensors to understand current conditions faced by the vehicle right now and respond appropriately. But this is really hard.
I agree about using the data only when there are no vehicles in front that are limiting your speed, but I think to do what they want to do, they have to be able to limit you to below the speed limit. I think the idea here is to avoid the problems they've had with people trusting TACC to take sharp curves and sharp exits properly, with some accidents as a result. So they actually have to drop below the speed limit on those curves/ramps if this "feature" is going to fix that problem.
I think the take-home message is that while that is a real problem, this is not the way to solve it -- or at least, they should only do this when using "self-driving" features like Navigate on Autopilot or any future "FSD" features. When I am using the system as an L2 system with my hands firmly on the wheel, I actually want it to trust me to be in charge and not try to overrule my decisions unless there is an imminent impact (which is the job of AEB). I want to be given the option to consciously and explicitly choose "this is driver assistance but you're in charge" system operation vs "I'm going to assume you've fallen asleep or crawled in the back seat and do what it takes to keep you from suing Tesla when you get in a wreck" operation. Because the former will bring back the smooth, predictable TACC and lane-keeping performance we used to have, and the latter will feel like a nervous new driver for a long time to come... maybe forever on HW2.x. Let them hand that liability back to me by giving me an explicit choice of operating modes.
Yes, please.
quickstrike12
Member
Very well said @rnortman
Elon(or maybe it was Karpathy) said it himself. People don’t drive around knowing and using google maps posted speeds, or using lidar shooting out of their eyes etc.....they use vision.
This is how a human knows to go 75mph on the roads to work at 6 am in no light traffic and no construction cones and how we tend to slow up a bit when we see traffic cones and or more traffic later in the day or whatever.
Or maybe when it’s a light rain or it’s dark ...whatever. We use our vision and make a decision AT THAT MOMENT on the speed that’s appropriate for the conditions.
Somehow to get past where we are now to FSD someday they will need to solve this with machine learning. NOT fleet speed, and not HD maps ........
Possibly some way of looking at the cars all around and “keeping up with the FLOW of traffic? Why brake from 75 to 55 somewhere (posted or not) if the cameras and radar can see cars all around not doing this. I know this isn’t a perfect solution either. Just one example.
But a great point brought up is currently there is not 2 different software approaches in design. Maybe there should be.
We with EAP and a AP2.X chip and perfectly content with L2 autonomy need a set speed and let me control some of these things. Keep my software simpler for that solution. Don’t overthink it.
And FSD design that will need to be way more adaptive since it will not have human oversight. And may be able to run dual HW3 chips and make many more adaptive decisions
Just a thought.
Elon(or maybe it was Karpathy) said it himself. People don’t drive around knowing and using google maps posted speeds, or using lidar shooting out of their eyes etc.....they use vision.
This is how a human knows to go 75mph on the roads to work at 6 am in no light traffic and no construction cones and how we tend to slow up a bit when we see traffic cones and or more traffic later in the day or whatever.
Or maybe when it’s a light rain or it’s dark ...whatever. We use our vision and make a decision AT THAT MOMENT on the speed that’s appropriate for the conditions.
Somehow to get past where we are now to FSD someday they will need to solve this with machine learning. NOT fleet speed, and not HD maps ........
Possibly some way of looking at the cars all around and “keeping up with the FLOW of traffic? Why brake from 75 to 55 somewhere (posted or not) if the cameras and radar can see cars all around not doing this. I know this isn’t a perfect solution either. Just one example.
But a great point brought up is currently there is not 2 different software approaches in design. Maybe there should be.
We with EAP and a AP2.X chip and perfectly content with L2 autonomy need a set speed and let me control some of these things. Keep my software simpler for that solution. Don’t overthink it.
And FSD design that will need to be way more adaptive since it will not have human oversight. And may be able to run dual HW3 chips and make many more adaptive decisions
Just a thought.
This is absolutely the only way EAP on HW2.x is going to actually be smooth and pleasant. HW2 just doesn't have what it takes to respond appropriately in all circumstances -- though since they did promise "on-ramp to off-ramp" to EAP purchasers, what they really ought to do is upgrade everybody to HW3, but nevermind that for now. Assuming they're not going to upgrade us to HW3, give us an option for basic TACC+Autosteer that doesn't pretend to be anything else (no NOA, no auto speed changes) that the system can't actually do smoothly and reliably. If I enable NOA, or ask for "guardian angel mode", then fine, switch to super cautious mode.
While the AP/TACC software is operating along a navigation route, the software knows there is a Fleet Speed adjustment coming up, just as it knows there is an upcoming lane change.
While the speed adjustments often occur on exit ramps or highway interchanges, the major problem is the automatic speed adjustments that occur on a highway, when the vehicle is operating at highways speeds, and the vehicle slows down without any warning, posing an increased safety risk when in the middle of other high speed traffic.
For S/X, the software should display a warning on the dashboard display of an upcoming speed adjustment, and provide a mechanism for overriding that adjustment, just like there is a way to do that for upcoming lane changes.
While the speed adjustments often occur on exit ramps or highway interchanges, the major problem is the automatic speed adjustments that occur on a highway, when the vehicle is operating at highways speeds, and the vehicle slows down without any warning, posing an increased safety risk when in the middle of other high speed traffic.
For S/X, the software should display a warning on the dashboard display of an upcoming speed adjustment, and provide a mechanism for overriding that adjustment, just like there is a way to do that for upcoming lane changes.
It just occurred to me, given the recent news that Tesla is being sued about the accident last year where Autosteer slammed somebody into a concrete gore, that maybe Tesla's actual goal here is intentionally slow down when a gore is visible. I wonder if the recent NNs have a gore detector and the car slows down when it sees one... whether directly in the path or not. That would be a really obnoxious way to "solve" this problem. I certainly hope that's not what's happening here.
Come to think of it, the spot where I have the most problems, there is a gore on a curve. So as you're approaching it, that gore actually is directly in front of the vehicle, though the lane obviously curves so as long as you follow the lane you're not in any danger of hitting it. But a not-very-smart NN doesn't know that.
Come to think of it, the spot where I have the most problems, there is a gore on a curve. So as you're approaching it, that gore actually is directly in front of the vehicle, though the lane obviously curves so as long as you follow the lane you're not in any danger of hitting it. But a not-very-smart NN doesn't know that.
Don85D
Member
My premise for FSD has always been that it is not viable until cars communicate with each other. Vehicles on a ramp become the eyes for the vehicles coming up behind. Construction zones are the same. When the fleet becomes all vehicles on the road ahead of you there is safety. In fact the only risk is that drivers not using FSD may be overwhelmed because average speeds are higher than they can manage without intensive attention to driving skills.
For now simple TACC is all that I need while we wait for all vehicles on the road to communicate with each other using a WW standard protocol. Until then we wait.
For now simple TACC is all that I need while we wait for all vehicles on the road to communicate with each other using a WW standard protocol. Until then we wait.
While V2V (Vehicle to Vehicle) and V2I (Vehicle to Infrastructure) would help, it will be many years before these are fully deployed and could be relied on for autonomous driving, and that's assuming those systems are bulletproof and not susceptible to malicious actions.
V2V/V2I is like HD mapping or Tesla's "Fleet Speed", this information is useful but should only be used as secondary information to what the vehicle is able to detect in real-time, since you can't guarantee that any of these secondary sources reflects the actual road conditions.
If Tesla's sensors and software can't detect the current speed limit (like a human driver) and adjust to the road conditions (curves, wet pavement, heavy traffic, missing lane lines, construction zones, debris in the roadbed, potholes, fog/rain/snow, …) then Tesla will never be able to achieve their AP goals.
We're supposed to get traffic sign detection soon for AP2+ vehicles, that should be a big improvement over the poor quality offline speed limit database or using "Fleet Speed".
V2V/V2I is like HD mapping or Tesla's "Fleet Speed", this information is useful but should only be used as secondary information to what the vehicle is able to detect in real-time, since you can't guarantee that any of these secondary sources reflects the actual road conditions.
If Tesla's sensors and software can't detect the current speed limit (like a human driver) and adjust to the road conditions (curves, wet pavement, heavy traffic, missing lane lines, construction zones, debris in the roadbed, potholes, fog/rain/snow, …) then Tesla will never be able to achieve their AP goals.
We're supposed to get traffic sign detection soon for AP2+ vehicles, that should be a big improvement over the poor quality offline speed limit database or using "Fleet Speed".
dark cloud
Active Member
jeffro01
Active Member
Okay so that explains all of the sudden speed oddities I keep running into... Tesla needs to start documenting these changes and putting them in the release notes. This is absolutely ridiculous...
Jeff
Jeff
We do not need fleetspeed, what we need in regards to (maximum) speed is:
- speed sign recognition on AP2.x+
- AP not taking blind corners full speed because it cannot calculate the curve path. SO IT SHOULD ASSUME ITS A 180 DEGREE TURN or atleast use the map data guestimate the curve
- daytime/night time speed limit differences for road segments encoded in the map data
- speed sign recognition on AP2.x+
- AP not taking blind corners full speed because it cannot calculate the curve path. SO IT SHOULD ASSUME ITS A 180 DEGREE TURN or atleast use the map data guestimate the curve
- daytime/night time speed limit differences for road segments encoded in the map data
quickstrike12
Member
I think speed signs might be a good place to start but there must be more. The posted speeds are not what the FLOW of traffic in a major city is by any stretch and if it keeps coming across an area where it sees a sign posted as 55 while the speed of traffic is 65 or even 70 is gonna be a problem. And many of them (the signs) are used as very conservative numbers that people don’t actually use. Example large sweeping curves to a new hwy where it recommends 45. People are maintaining speeds like 70 usually. Many of these scenarios here in Dallas.
I don’t say it’s wrong and it might be better than just using map data like we have now, but there must be more.
It all goes back to “how does a human do it” ?
For FSD to be successful they must solve this.
Thanks for this. This is humorous-level total incompetence by the Tesla "autonomy" team. It shows that they still don't have a clue what they're doing with driving policy. Autonomous cars? Never at this rate.
Luckily you can just turn off TACC and Autosteer. Fleet Speed doesn't activate when they're off, right?
Well, that is true, but I didn't pay this much for a car expecting to get one without cruise control. I would be OK with this feature if it only activated when you were using the "advanced" driver assist features like NOA. But when I'm driving the car and I just want adaptive cruise control, the car needs to just give me adaptive cruise control -- maintain my set speed unless there's a car in front of me going slower, in which case it should match that car's speed. Period, end of story, give me that as an option at least.
I was in a fight again with the 'feature' this morning. On a stretch of highway near my house has a limit of 130km/h, but my Model S just keeps going down to 110km/h.
I keep pushing the leaver back up to 130km/h, but the car then slows down again to 110km/h. Super annoying feature!
The feeling I can't shake is that the people who write this software never drive the cars in real life.
Sometimes I just want to take a Tesla engineer with me for 30 minutes and show him/her all the small flaws which are in Autopilot. \
I keep pushing the leaver back up to 130km/h, but the car then slows down again to 110km/h. Super annoying feature!
The feeling I can't shake is that the people who write this software never drive the cars in real life.
Sometimes I just want to take a Tesla engineer with me for 30 minutes and show him/her all the small flaws which are in Autopilot. \
Similar threads
