Featured CR on Driver Assist Systems

I don't think the majority of incidences end in a fatality.

Right now, a fender bender involving a Tesla with Autopilot on is news worthy. Simple odds says such are also happening with the others. Why aren't they news. Part of it is because it simply isn't being reported the ADAS was on.

 

How often do local or state police even ask for those recordings? I strongly doubt their ask rate, let alone their rate of reporting up to the feds, is at all comparable to Tesla's automatic reporting.

Heck, even police department reporting of murders for the FBI's annual Uniform Crime Reports (UCR) collapsed in 2021, when the old simple forms were obsoleted and they were required to use the new and more detailed NIBRS system. 37% of law enforcement agencies responded by completely ceasing to submit any statistical crime reports at all. Some others submit only partial data, leaving barely half submitting full data. As a result, the most recent annual FBI crime summaries rely heavily on estimates to fill in for the missing data. I'm doubting their vehicle crash reporting will reach equivalent levels.

 

anyone else remember how many of us had to get bombarded by news / family with ongoing stories of the Prius supposedly lurching into buildings - cars - pedestrians etc when in fact it was drivers stomping on the wrong pedal?
Seems iirc CR was glad to weigh in on that not true debacle.
.

 

When any ADAS system can demonstrate and document a weighted all-conditions all-roads fatality rate under about 1 death per 100 Million Vehicle Miles Traveled while being used in driver-napping mode, or a good-conditions freeway rate very substantially better combined with declining to operate solo in poorer conditions, instead pulling over and parking until the driver wakes up and re-engages, then I will be more interested.

Anything less means that they are still not equal to 'average'(*) humans. Or at least what humans demonstrated before their Pandemic-era degradations.

(*) Note that this would be a weighted 'average' of the distribution of human driver performances. The way the crash safety statistical math works, one terrible driver can push crash rates up more than 10 or possibly even 100 perfect drivers can pull it down.

 

Wow. Once again the tesla autopilot fails to work the way that it claims to. In crash investigations, Tesla has assured the NHTSA that the car will detect drivers that are not capable of responding to emergency situations. That includes sleeping, reading the paper, etc.

Obviously this one is a fail. The driver was obviously sleeping. The "lane keeping assist" and dynamic cruise control look seem like a good thing when the person falls asleep... right up to the point when the autopilot turns off due to an unexpected situation, leaving the tesla as an unguided missile without brakes.

A year ago you posted that you had the "unplanned micro sleep events" while driving. ( Why Full Self Driving is needed | Page 3 | PriusChat ) You credited the autopilot with saving your life.

I thought it was odd at the time that you did so much driving while impaired.

 

1st irony is there nothing that says there was a crash from this instance.
irony of ironies ... if the video had been pulled from the phone of a corpse who crashed because (s)he was busy videoing a Tesla driver - instead of watching the road?
Nope ... the issue instead must be the person who may have passed out from diabetes or who knows what ... and it's the Tesla's fault. Yet the early version did not incorporate feel against the steering wheel. Again, presuming the worst. That's the fear mindset of those who can't choose to embrace improvements. That's what improvements are, the bare statistics that show these systems are better than man on his own.
Sorry but it's not a boogie man
.

 

They can only do this in small geocached areas with limited speed, and the leader Waymo has a lot more crashes and fewer cars than teslas autopilot in those small areas. This is likely because the driver of a tesla will take over in situations when the waymo or cruise vehicle will collide.

For non geocashed areas we are at level 2, which means a driver should take over. Since many crashes occur because of distracted driving, and the system can stop accidents with short distractions, a driver plus the system should be safer. My tesla's ADAS is much safer than it was in 2018 when I got it, but there are obvious situations to take over.

NHTSA is examining whether the system is so good that drivers rely on it and don't take over in these situations. They have found in some obvious situations when the drivers hand was on the wheel, but did not take over and we have some incidents where people have fallen asleep but in such a way that the system thinks they are paying attention. The bottom line is the systems are improving but they require a driver that is alert. There is a camera in my car that I agreed that tesla could turn on and monitor my attention when using the self driving beta, but they do not look on autopilot, and older cars do not have an inside the car camera. They may need to turn that on.

NHTSA says that the majority of accidents in teslas happen at night and in bad weather. Drivers need to be aware of these situations.

Absolutely which is why I presented the study with the same drivers driving anouther non tesla/porsche, versus their tesla or porsche. The tesla resulted in fewer accidents than the other car, and the porsche more than the other car. They also found tesla model 3 drivers 5x more likely to have excessive acceleration as defined by insurance companies, than than their other car. Still they were less likely to be involved in accidents.

 

I think you should stop making that claim about the study. The study itself was not available for examination. What you presented was a link to an article that claimed that there was a report about a study.

 

It is available but behind a paywall. I really do not want to violate the terms to give it to you, but you are welcome to buy it.

Your data that tesla's autopilot is less safe with an aware driver does not exist. If you have it please present it.

 

Who are you talking to? I've stated before that I don't have insider information about how many accidents have been avoided by the human driver's alert intervention when an accident was eminent. The last time I checked, Tesla only releases that information when forced to by the laws or regulations. Read more here: https://electrek.co/2022/01/12/tesla-forced-report-full-self-driving-beta-data-dmv-after-videos-showing-dangerous-use/


Austingreen is the one who insists that it's safer to drive a model 3 with a driver that is forced to pay attention than it is to drive the same car without that driver nagging.

But here's an easily found article that talks about the dearth of result data from Tesla, and it also points to independent data that has been voluntarily accumulated by 140 or so beta testers.

https://electrek.co/2022/12/14/tesla-full-self-driving-data-awful-challenge-elon-musk-prove-otherwise/

The independent data is at Home

That's really eye opening. In Virginia the average report is over 400 miles of driving between critical disengagements. That was in January of this year. California had as low as 2 miles AVERAGE between critical disengagements.

Definition:
Categories of Disengagements:

• Critical: Safety Issue (Avoid accident, taking red light/stop sign, wrong side of the road, unsafe action)
  

 

The Ca DMW is making the same mistake as you; thinking FSD is Level 5 when it is just Level 2. If you are going to ask for data from a Level 2 system, you'll need to ask for it from all. Did you apply the brakes when your Prime autobraked? If so that's a disengagement.

It won't take much to convince me Musk is hiding behind Level 2 to avoid oversight. The solution there is for regulators to close the loophole by better defining what the different Levels are in the law.

There are just 148 testers on that site. There were 160k users of FSD in 2021. It went to 400k last year. That data can very easily be skewed.

 

Found this online at https://tools4dev.org/resources/how-to-choose-a-sample-size/ It matches what I remember from high school teachings.

These testers are not providing data for precision analysis. It's showing trends that can be used to inform decisions about progress of the product.

While there may be just 148 testers who are making their data public, that should be enough for the use to which it's being put.

 

They test Tesla like they test any other vehicle: By indicating the make, model, trim, and year. If Tesla listed their trim as "2022.44.30", then that's what they would have to indicate. And if the feature they are testing is not dependant on the trim, they'll say so.

Tesla is not the only car that gets updates over the air. CR can't know the versions of the dozens of the softwares that run in these vehicles. It's interesting that you think they should.

 

From the link,
This advice is NOT for:

• Research studies conducted by universities, research firms, etc.
• Complex or very large surveys, such as national household surveys.
• Surveys to compare between an intervention and control group or before and after a program (for this situation Sample size: A rough guide).
• Surveys that use non-random sampling, or a special type of sampling such as cluster or stratified sampling (for these situations see Sample size: A rough guide and the UN guidelines on household surveys).
• Surveys where you plan to use fancy statistics to analyse the results, such as multivariate analysis (if you know how to do such fancy statistics then you should already know how to choose a sample size)

This survey is the opposite of non-random. It is an opt in survey. It is also complex with the variables being collected(state, time of day, weather, version, critical vs noncritical, whether same location or not). Even with an apt to automate some of this, the level of effort from the respondent is going limit this to people dedicated for whatever reason. This isn't a cold call or questionnaire being sent out to a random group to get a few answers.

This also isn't a survey of preference questions. The respondents need clear definitions of what a disengagement is.

From the survey page, "Braking: Braking required due to late deceleration, accident avoidance, or moving forward incorrectly / unsafely." Okay, let's say a kid runs out in front of the car. The car fires up the warnings, and applies the brakes. The driver also responds to the kid and warnings by also hitting the brake pedal.

Is that a disengagement? Maybe the car was going to stop on its own. Maybe not. Can't know now. Of course, I didn't want the driver to not brake for the kid. In case of a Level 4 or 5 car being tested, the date recording is likely finer to allow the researchers to go back and determine if a safety drivers intervention was actually needed in such a case.

Are these reported disengagements true disengagements, where the car failed to react in a situation, or are they the result of the drivers responding to the car's warnings?

PS: If a higher miles between disengagements is good, why does the report use angrier reds for higher numbers?

 

Of course it's not a random sample. In order to report, they need to have a Tesla. They have to pay $15,000 (non refundable, not transferable) for software that may or may not have bugs. Then they have to be willing to risk a being banned from the Tesla tests, superchargers, etc if they piss off Musk.

What are the chances of getting a qualified random sample by standing in front of Walmart with a clip board?

Yeah, I have kids running out in front of my car so often that I need to put a cow-catcher on the bumper to push them back onto the sidewalk.

Seriously, I have not had to do an emergency stop for a kid running in front of the car in the last decade. Not that the kids are non existent, but I drive the speed limit and keep an eye out for signs of movement. I've stopped long before the child is in danger. On the positive side, my

That's not a valid concern. By the definition above, ( "Braking required due to late deceleration, accident avoidance, or moving forward incorrectly / unsafely." in case you forgot) If the driver is braking for those reasons, then it's a valid disengagement.

My guess, if you don't have an ax to grind, you don't care what color you use. Are you suggesting that the data is somehow more valid if it's represented in baby blue or soft pink?

TL-DR: This is not a survey. The people who shared their data were, by necessity, self selected and not random. But that does not invalidate the data that they report.

 

You are purposely missing the point. For the minimum sample size of 100 to be statistically valid with a 10% margin of error, the samples have to be randomly selected. If not random, then more samples are needed to counter any possible selection bias introduced by the non-randomness. If CR used car message boards for survey, every car would have a crap reliability rating. An opt in data collection site could be no different than the maintenance section of a car message board.

You are being obtuse. There many potential cases where a Level 2 ADAS will give a warning, and the driver can take action before the car can correct itself. Despite the names and what Musk says, Tesla systems are just Level 2.

Everytime the speed or following distance on a DCC is adjusted is a disengagement then.

It does open it up to be questioned about selection bias.

You actually don't care about Autopilot or FSD. Otherwise, you'd be bitching like this in the threads about openpilot. There is about many cars using it as Tesla's with FSD. Has the same level of regulatory oversight, and cars can be running user modified software.

You are here because you believe only low power EVs can be efficient, and help with carbon emissions. Tesla's existence proves that wrong.