-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla.com - "Transitioning to Tesla Vision"
- Thread starter run-the-joules
- Start date
Daniel in SD
(supervised)
The seat belt pretensioners are explosive devices and need to be replaced after being used. Different than whatever the mechanism is that locks the belt during hard braking.
stopcrazypp
Well-Known Member
As another pointed out the seat belt pretensioners are part of the crash module, the triggers are not the same as the ones that pull the belt in for sporty driving (I don't think that Tesla has that function anyways, my Model 3 certainly doesn't).
stopcrazypp
Well-Known Member
Not really, speeds below 15mph have low risk. They increase rapidly after that.
In Crashes, Low Driving Speed Can Cause Serious Injury and Death to Pedestrians, Report Finds
Well the argument is over if the report data includes rear end accidents (no matter the cause), and it clearly does. Sure, it doesn't break down the exact cause, but we don't really need to know that, given no matter the reason (whether from AP false positive or not), the prevalence of accidents overall is way lower than the NHTSA average (which may be underreported in the first place). Plus as a sanity check, the overall prevalence of rear end accidents according to NHTSA is around 32-33% of all accidents (depending on year), so it's not like Tesla's data is an outlier in terms of ratio to other accidents.
(For NHTSA rear end percentage, you can see under Chapter 2: Crashes > Crashes: Circumstances > Table 29)
Traffic Safety Facts Annual Report Tables
Actually this relates very well to the noisemaker argument also. When the driver is creeping up to a pedestrian, it is the driver's responsibility not to run them over. A noisemaker can't be relied on given the pedestrian may be distracted, wearing headphones, or simply do not care.
Actually it says all these are D (bold emphasis mine):
H3 Loss of Vehicle Lateral Motion Control
H4 Unintended Vehicle Deceleration
H5 Insufficient Vehicle Deceleration
H6 Loss of Vehicle Longitudinal Motion Control D
That's why there's the discussion above about if Tesla is using the rear facing cameras to detect if there is a risk of being rear ended. For example, if the car knows there is no car following, then it may feel safer to err on the side of braking (even if it may be a false positive). This would match with claims by some that the phantom braking tends to happen in empty roads. It may mean an unpleasant experience for the driver and passengers, but it may be "safer" if there was no car following in the first place.
Ahh, the scientific Forbes article that says it's "low" with no numbers at all. The title is kind of hilarious for your argument as well- "In Crashes, Low Driving Speed Can Cause Serious Injury and Death to Pedestrians, Report Finds"
Are you willing to let someone hit you with a car at 15MPH?
It's not.
Tesla compares cars on AP (which are on the highway) to the overall rate of all cars on all roads. We have zero data on Tesla on AP to Tesla's not on AP in exactly the same conditions, which is the only data that matters.
Did I say it was only unintended deceleration?
This is kind of a hilarious defense of a immature system. Tesla still doesn't trust forward facing cameras as much as radar, but they might be using a rearward facing, monocular camera to decide if they should brake for an estimated forward collision in case it may cause a rear end accident? Yet they can't avoid summoning of a ledge?
The clear reason it doesn't brake in traffic (if true at all) is that it locks on the car ahead, it has nothing to do with a car behind.
stopcrazypp
Well-Known Member
I'm just pointing out an example, there are other papers you can find with more detailed data if you look for it. And this is straying very far from the original topic, which is serious injuries/damage from phantom braking. I think it's be sufficiently demonstrated that the AP data Tesla has published so far is enough to cover that. I don't really think there are many if any scenarios where phantom braking will cause you to hit a pedestrian below 12mph.
There's discussion elsewhere, but one big difference with Tesla's AP is that it can be used (and is regularly used) on all roads, not just highways. It's not like ones like Supercruise that lock you to specific roads.
Tesla provides the data with AP off also and it's significantly lower than the NHTSA average (which given we know the rear end proportions are basically the same 35% vs 32-33%, we know the rear end prevalence is also just as low by the same proportion). That shows AP doesn't push the rear-end risk average over the general average from NHTSA stats. Sure, it'll be nicer if Tesla breaks down the data further by road type, but for this particular discussion it's not necessary.
You didn't, but you said it was the "highest rating" so it's misleading to not point out other factors, which Tesla may be balancing also.
Well we don't know that at the moment. People haven't pointed out if they were following a car or not or if a car was following them or not. As for trusting a particular camera, as we saw in the previous discussions it's all a matter of range and confidence levels, it's not an on/off determination (same with the amount of deceleration being applied, there are different amounts that can be applied, that can be adjusted based on detection: it's not an on/off thing). For the front, the car has to also determine if there is any object or barrier that it might hit (not just cars), for the back it only has to determine if there is a car approaching, so the task is not the same.
Last edited:
I have basic AP and I'll get phantom braking at the weirdest times. Some I can isolate to objects I've seen either on the side of the road, or in one instance a bag that flew in front of me on the interstate. But the rest of the times make no sense, and it's hard braking. I've got less than 2k miles on the car, and have had too many instances of phantom braking to trust it fully. I'm more alert when it's on than not, and for further clarification a good percentage is on TACC without AP.
I do feel that they'll fix these issues, but until then I'm very wary of using either TACC or AP.
I do feel that they'll fix these issues, but until then I'm very wary of using either TACC or AP.
1 year of ownership (car has radar, 4.18.2 firmware) and ~10k miles. I use AP virtually every time I drive on freeways and surface streets. No phantom braking.
The car does lift off the throttle and coasts in some spots, usually at exchanges when I'm driving under/over some other road.
The car does brake when driving by cyclists. It's gotten much better at judging distances and doesn't brake when there's >5' space.
The car does lift off the throttle and coasts in some spots, usually at exchanges when I'm driving under/over some other road.
The car does brake when driving by cyclists. It's gotten much better at judging distances and doesn't brake when there's >5' space.
I will say my no radar Y rarely shows cyclists or pedestrians. My radar X is MUCH better at that. Even when someone stands right in front of the car, the Y takes a long time to show them. Very unsettling.
So strange that everyone is having different experiences. My vision Y picks up pedestrians quickly, not sure about bikes. And in about 1500 miles, mostly on freeway, I haven't had any phantom breaking that couldn't be explained by a car merging in while I was in the slow lane. I wonder if there is some kind of HW variation, or the cameras aren't aligning correctly in some cars.
mikes_fsd
Banned
There's got to be something else going on here that is either being missed or not reported.
Just seems odd to have such diverging experiences.
I agree, that's why I'm wondering if there is enough variation in the HW or camera calibration to explain it. Some report lots of issues, and other none. It seems like there's no middle ground. Normally only see this when there are variations in how the HW is setup or the how the user uses it (I'm a software engineer in Aviation).
And this is why Neural Networks non-observable code are such a hot topic in aviation. They're very hard to understand where small variations cause big changes.
How would you go about certifying Tesla Vision to DO-178 DAL B?
I wouldn't think it would be possible, just the no dead code would probably prevent it. I'm amazed at what they have been able to do so far, and hope they can reach L4 (I don't think L5 will be possible in the near future).
It will be very interesting to see how Tesla and the rest of automotive goes and demonstrates these safety critical, ASIL-D (DAL-A) systems are acceptably safe, while having to use a completely new path, since as you say, historical methods won't work. It's going to be such an enormous leap from L2 to L3+, and there isn't a lot of data that Tesla's following any kind of structured development path that is focused on proving acceptable safety. The constant changes in vehicle configuration show a very early stage of development in this area.
FYI, L4/L5 isn't really functionally different. Nobody really cares if your Tesla can go offroading without you in it. It's still a human asleep in the back of a car with no driver, going 75 MPH down the highway or moving through dense city traffic. L5 is not a higher level of safety, it's just more areas where the vehicle can go autonomously.
Agreed, and that's the part I don't see them getting. I think there will always be some corner cases that aren't covered, but hopefully I'm wrong
linux-works
Active Member
some training we all got at work was iso26262 and the asil ratings. there's at least one person in our extended group who has many years of doing safety critical analysis of car designs. and we rely on that person a lot, since most code writing engineers know only a little about this safety coding stuff.
if you pay a lot of money and buy 'point and click' type process draw applications - that end up generating models - that end up generating code -- that can end up being safety certified and provably correct by the vendor (one of the reasons why you pay that vendor so much. SOOOO MUCH. but I digress).
tesla does not use autosar style 3rd party vendor code, to my understanding. in that case, they are themselves on the hook to prove its 'safe' and if anyone takes them to court, its only them that will show up, and not some other supplier with a decade or more of experience with this. and no, tesla does not have all that much experience in this. I'm not even sure they have solid asil-d where it counts. for one, there's no signs of redundancy in my model 3, that I can see. and the vision system is not even sensor-redundant, so that's a non-starter.
to get to full asil-d where lots of things can fail and the car can still be minimally safe - that's a big trick to pull off and I wonder if tesla's current architecture was designed with enough of that from the start. we already saw radar 'fall away'. if I was asil-d rating this car, I'd have lots of sleepless nights, I'll tell you.
which is an issue in Germany as the competition has no issues running their lane-keep assist up to 180kmh (or even higher)...
Similar threads
- Article
