Model X Crash on US-101 (Mountain View, CA)

[thegruf]

the only reason the distance setting of 1 (or 4 or 7) can be relevant is if a car in front suddenly changes to another lane leaving a corresponding time for the Tesla to re-acquire objects in front. I recall that Tesla specifically warn about this circumstance.

The radar is inherently a doppler device and as such does not "see" stationary objects. Tesla however have been developing exactly this capability through software advances. Unfortunately for reasons unknown it was unable to prevent this accident.

 

[S'toon]

The Wright Brothers went after their goal very patiently and methodically, with a rich trove of both practical and scientific knowledge to build upon. I much enjoyed reading the book by David McCullough about their signal achievement.

You mean the second manned powered flight in history?

 

[Krazaak]

I don't understand AP. From all the posts I've read, it's frontal detection seems to be dismal. Calling it AP is a scam. It should be called BP (Beta Pilot) or GP (Guinea Pigs). Eon Musk is effectively crowdsourcing the testing of AP. This product should never have been sold. I will never buy the AP product for my M3. And yes, I am an engineer, but the kind that addresses ALL scenarios.

The mistake is that thinking autopilot means something more than it does. I've used autopilot for 1000s of miles and it has never once done anything dangerous, thrown me towards a barrier or crossed outside my lane. The reason it has never done anything wrong, is because I've either been looking out my windshield and have disengaged AP when it was obvious it wasn't going to handle the markings well or I've had my hands on the wheel and the vehicle's attempt to steer where it shouldn't resulted in an automatic disengagement.

The danger isn't in having autopilot enabled, it makes the ride more relaxing, especially in stop and go traffic. The danger is in thinking you can do something else while it holds course for you.

I also firmly believe that driving with autopilot on is much safer than driving with it off. If you're paying attention, then you should be able to counter any potential bad behavior from autopilot. If you're not paying attention, then in most cases, autopilot will continue to hold course properly and keep you from drifting out of your lane or rear-ending stopping traffic. If you're not paying attention and autopilot crashes, two things may have failed to create that situation, but your failure is still the primary issue.

 

[cbdream99]

n summary, if you have an AP2 vehicle with Autopilot engaged and Autopilot is detecting wider-than-normal lane markings, be extra alert and be ready to take control. It goes without saying that you should always be in this state when using Autopilot, but pay extra attention to wide-lane scenarios until you know how the vehicle will react.

That's my experience too, how the system react depends on speed, line of sight, which lane marking is detected first and how many events are processing at the moment.

When I am traveling 45mph on a slight curved 2 cars one direction road, after crossing a junction that has no markings and just after the junction, the middle markings has not started but the right lane markings are solid (I am on the left lane and the car obviously see the solid right lane markings), so the car immediately swerve to the right to center itself and once it detected the middle lane again, it goes back to the left lane. If I reduce the speed by 10 miles then the sudden swerve will not be as noticeable I can't image what will happen if there is a car on my right shoulder and I wouldn't even try to engage AP if there is heavy traffic in that section of the road just try to 'reproduce' the problem.

 

[mongo]

Regarding the recent blog post:
The distance from the start of the lane split to the barrier is 200 meters (per Google maps).
Tesla reports unobstructed view for about 150 meters (roughly from just before the overhead sign to the collapsed barrier, 6+ car lengths into the split).
Tesla also reports following distance set to 1.
Given the missing 50 meters of visibility, a potential senario is that there was a vehicle in front that AP was tracking until that vehicle rapidly exited the gore point (changing lanes to 101-S seems popular), whereupon AP continued tracking the most reasonable set of lines based on vehicle path and position. (Note there is a difference between where the car is and where the camera is looking, so it would likely be processing lane data from further down the path where it is wider)

With the .4 "wide lane" software, it could end up tracking the wedge not-a-lane section right into the barrier. Apologizes to a previous poster, the impact point from the crash does align with being centered in the gore lines. The right hand side being the widening one would also provide a small amount of left to right moment. Additionally (based on news footage), the pack may have achieved a high level of pentration through the lower right side of the metal structure inducing the (likely counterclockwise) spin into the 101-S lanes.

Regarding AEB: the car was likely half out of the path to the barrier, so driver intervention could have prevented impact, and possibly suppressed AEB response. It is a difficult distinction between clearing a construction barrier with nearly no shoulder and clipping it. The track structure (many close spaced horizontal tubes) on the barrier also likely produced a noisy radar return. This is a situation that vision (or other higher resolution sensor) would be helpful in.

 

[Krugerrand]

That's assuming if the driver continue to pay full attention when using the current autopilot system. But like I said before, many driver won't pay full attention because hands off and legs off driving encourage driver NOT to pay attention.

No it doesn’t. If you’re not paying attention with AP on, you were already someone not paying attention without it; you’ve just simply given yourself an excuse now.

 

[22522]

I have experience now with the new "wide-lane" feature.

I guess I'd prefer AP2 to "know" that it has extra lane space (for emergency maneuvers), but continue to stay a reasonable distance from the "primary" lane paint instead of constantly re-centering for lanes that are perceived to be very wide.

Your statement is very kind and gracious.

I am disappointed that the people in Tesla that believe in the existence of wide lanes continue to have any sway.

Maybe your kind and gracious approach will work, although it has not in the past. I will add a less kind and gracious approach here:

1) Many roads have crumbling right shoulders where the right side line is gone. Or snow or dirt drifts where it is covered. There is a primary center line that separates traffic. Register off of that. Alarm if the lane becomes wider than normal! That might mean that you are in an intersection, where traffic has eroded the right lane marker.

2) The google Lexus suv fender bender with the bus was built on the fiction of a wide lane - with sand bags on the right edge. A lane centering algorithm that weighs the right edge anywhere near as much as the primary marker on the left causes steering inputs that are not well predicted by other drivers.

3) In this case, a wide lane alert may have helped the car know it was tracking the wrong left line. Although the real left line was so worn, it may have been processed as a dashed line for a left exit. Does the software do an FFT (looking for fixed frequency) on dashed line spacing to sort worn lines from deliberate man made dashed lines?

 

[Az_Rael]

Why do people think they need make AP do things it wasn't intended to do? Why would you think it's a good idea to try to make the system change lanes into a shoulder?

Well, the shoulder thing is interesting as EAP will eventually have fully auto lane changes - I.e no driver input required to change lanes. So it will be a feature, it just isn’t one yet.

 

[MikeQ]

Tesla says autopilot was on during deadly Mountain View crash

“Tesla emphasized this crash was so severe because the safety barrier "had been crushed in a prior accident." That happened, as Dan Noyes reported this week, because an alleged drunk driver in a Prius hit the smart cushion at 70 miles an hour -- he walked away with minor injuries.”

So basically Autopilot is no better than a drunk driver. I’d say that’s accurate at times.

It also suggest that had the barrier been fixed, we probably wouldn't be talking about this right now.

 

[AWDtsla]

I don't understand AP. From all the posts I've read, it's frontal detection seems to be dismal. Calling it AP is a scam. It should be called BP (Beta Pilot) or GP (Guinea Pigs). Eon Musk is effectively crowdsourcing the testing of AP. This product should never have been sold.

Why would I want a product that is effectively the equivalent of teaching a teen to drive. Keep hands on the wheel, watch the road, correct errors. That's what a parent does when teaching a teen to drive!

I will never buy the AP product for my M3. And yes, I am an engineer, but the kind that addresses ALL scenarios.

It's just another tool to use while driving. Don't expect it to do things it can't. Half of that is Tesla's fault, the other half is the fanbase. Again, putting your hand on the wheel does nothing if your expectations are wrong.

 

[Snowstorm]

Why did this car crash into it and not all of the previous AP drives that goes past this point? I would assume dozens of AP drives would go past this point dairly. What special condition exists for this drive? Or is AP somehow non deterministic in behavior? What firmware was he on?

I surely hope Tesla doesn't further nanny the system as it is today. I remember that nags were almost not a thing before the Florida crash and now happens all the time unless you tug on he steering wheel. Hope this doesn't cause further disablement of the system.

 

[Krazaak]

Well, the shoulder thing is interesting as EAP will eventually have fully auto lane changes - I.e no driver input required to change lanes. So it will be a feature, it just isn’t one yet.

I don't remember which update improved it, but the recognition of adjacent lanes has dramatically improved since I got my car last July. With AP engaged, there's a visual cue if the car think there's an adjacent lane. In the left lane on the highway with an open median, it would often indicate an open lane in the grass to my left. Within the last few months, I've noticed it has been close to 100% accurate in determining adjacent lanes. I assume they're reading solid vs. dashed lines. I can confirm that AP will not change lanes in a direction that it does not believe a lane exists, because I tried to change into an exit lane before the display picked the lane up.

 

[scottf200]

Why did this car crash into it and not all of the previous AP drives that goes past this point? I would assume dozens of AP drives would go past this point dairly. What special condition exists for this drive?

** Driver's make choices in how to use the driver assistant feature **

Driver turned it on ** moments ** before the accident and set the following distance to the minimum of 1 (1 thru 7).

READ the last four paragraphs. Obviously a very sad situation for the family and everyone involved.

 

[Krazaak]

Why did this car crash into it and not all of the previous AP drives that goes past this point? I would assume dozens of AP drives would go past this point dairly. What special condition exists for this drive? Or is AP somehow non deterministic in behavior? What firmware was he on?

I surely hope Tesla doesn't further nanny the system as it is today. I remember that nags were almost not a thing before the Florida crash and now happens all the time unless you tug on he steering wheel. Hope this doesn't cause further disablement of the system.

Tesla made the point that the driver had the TACC follow distance set to 1. I don't know if that was relevant to the issue, because I've not seen any discussion regarding traffic in front of the driver before the crash. I do know that closely following another vehicle impedes your car's ability to identify and track lane markings, especially if that vehicle is riding the margin of the lane.

It's possible that if he was following a car closely that was hugging the right side of the lane at the split, that the car would not identify the lane split and continue straight into the obstacle, thinking it was properly staying in lane.

 

[Reciprocity]

Well, the shoulder thing is interesting as EAP will eventually have fully auto lane changes - I.e no driver input required to change lanes. So it will be a feature, it just isn’t one yet.

It will require high def 3d Maps which are a feature that is beyond AP. See lvl5 lvl5, Inc. | HD Maps and Computer Vision Software for Autonomy for more info. 2 former Tesla employees. Every drivable lane will be stored in local mall tiles and updated frequently.

 

[scottf200]

It's possible that if he was following a car closely that was hugging the right side of the lane at the split, that the car would not identify the lane split and continue straight into the obstacle, thinking it was properly staying in lane.

But what do you think ** moments ** before the accident AP was turned on means ?!?

 

[trmmcd]

I use AP1 extensively - probably 20K or more miles of driving utilizing AP. As many active users have suggested - it can be a safer form of driving when used correctly. In Tesla's account of this terrible accident - they indicate that the driver experienced multiple visual alerts and one audible alert during this driving session. In all of my driving I frequently experience the flashing visible alert nag - as that happens after only a few seconds - but after the first week or so of use I have never been inattentive long enough to trigger the audible alert. I don't know the specific length of time before the audible alert sounds - but that suggests to me significant inattention to the operation of the car. As I understand the current methodology - three audible nags shuts down the AP capability until the car is parked. Based on my experience - successfully driving thousands of miles without an audible alert I would fully support one audible alert triggering this shut down. I have assumed that the AP2 audible alert triggering is similar to the AP1 audible alert algorithm.

While there are clearly many contributing factors to this accident, I reluctantly have to admit that AP/user interaction seems to have been an important factor. My hope is that there are some readily implementable refinements to AP alerting that allow us to continue to reap all of the safety benefits of this technology while further discouraging inappropriate use.

 

[Xenoilphobe]

View attachment 288884 A typical road marking of a european/Norwegian highway split.
Oslo
Google Maps

Do you guys use rumble strips to notify the driver?

 

[mongo]

The google Lexus suv fender bender with the bus was built on the fiction of a wide lane - with sand bags on the right edge. A lane centering algorithm that weighs the right edge anywhere near as much as the primary marker on the left causes steering inputs that are not well predicted by other drivers.

That instance was not caused by centering in a lane for the sake of being in the center if a lane.

The Google fender bender was based on the car being programmed to act human and use the curb edge of a wide single lane to make a right turn without blocking through traffic. The car was going to make a right turn, then had to move back into the center of the lane to avoid the sand bags. It assumed (wrongly) that bus would yield.

 

[Barklikeadog]

It also suggest that had the barrier been fixed, we probably wouldn't be talking about this right now.

The story would be different....
The driver would be alive and claim that AP wrecked his car.
Tesla would say that the driver was at fault and AP worked fine.
Tesla makes public the drivers actions.
The driver tells his story.
Tesla wins.
Next day Tesla issues an AP update for that split in the road
Even though they said AP functioned properly.