Autopilot: Crashed onto a stopped Switzerland Van

[Tam]

Tesla Model S driver crashes into a van while on Autopilot [Video]

I hope this thread is for educational purpose for those who are interested in Autopilot.

It looks like the owner is posting the youtube publicly again.

I may be wrong but this is my interpretation:

This incidence involved 3 vehicles:

1-Tesla with the system locked onto Black Wagon that's leading the way
2-Black Wagon that leads Tesla the way.
3-Stopped Schubiger Van further ahead that the system never locked on because it already locked onto the Black Wagon

The chronology:

Tesla is happily following the Black Wagon that's leading the way. The system locks onto it tightly and wouldn't let go until the collision in this case.

The Black Wagon is happily leading the way and saw the Stopped Schubiger Van further ahead.

The Black Wagon carefully slowed down.

The Tesla also carefully slowed down.

The purpose of the Black Wagon to slow down was to straddle the two lanes in order to pass the Stopped Schubiger Van further ahead.

The purpose of the Tesla in slowing down was to adjust the constant distance to the car that it has locked onto which is the Black Wagon as it is designed.

Once the Black Wagon has passed and cleared the Stopped Schubiger Van, it speeds up to get back to the correct lane.

The Tesla noticed that the Black Wagon speeds up, so it faithfully speeds up as designed to keep a constant distance to the car that it has locked on which is the the Black Wagon.

And in so doing, Tesla collided with the Stopped Schubiger Van which it has never locked on before.

It is not designed to lock on the Stopped Schubiger Van but it is designed to keep following the leader which is the Black Wagon so there's no reason for it to keep distance for this case.

The owner seems to indicate that manual brake was applied but it's too late because of trusting the system that works 1,000 times before so there's no reason to distrust it in this case.

If so, manual brake would disable the Autopilot.

Thus, technically, this is not an Autopilot crash because at the moment of the crash, it was manually disabled (again by the act of manual braking.)

Manually disabled crash was also described in Lebec, CA because the driver also manually applied the brake.

Whether a crash was under "Autopilot" or not when the collision occured (due to manual applying the brake in each case) is another academic and legal discussion but the fact still is: Autopilot was initially enabled and trusted to work in each case.

So what has gone wrong in this case?

According to Tesla design: Nothing!

That is what it is designed to do in this case.

And it faithfully executed according to its design in this collision perfectly.

The manually clearly states that it:

“Warning: Traffic-Aware Cruise Control can not detect all objects and may not brake/decelerate for stationary vehicles, especially in situations when you are driving over 50 mph (80 km/h) and a vehicle you are following moves out of your driving path and a stationary vehicle or object, bicycle, or pedestrian is in front of you instead. Always pay attention to the road ahead and stay prepared to take immediate corrective action. Depending on Traffic-Aware Cruise Control to avoid a collision can result in serious injury or death. In addition, Traffic-Aware Cruise Control may react to vehicles or objects that either do not exist or are not in the lane of travel, causing Model S to slow down unnecessarily or inappropriately.”

This kind of collision can be replicated consistently according to the manual.

So it is up to a driver to read the manual and prepare in such as this scenario.


About the driver's points:

1. The TACC, active cruise control did not brake as it normally does
2. The automatic braking system (AEB) did not make an emergency brake
3. The forward collision warning turned on way too late, it was set to normal warning distance
4. The TACC actually was speeding up just before I did hit the brakes

I think we covered them all in here and in the other case in Lebec, CA, so there's no need to repeat unless there's still a repetition need.

 

[TS14PCC]

I'm no Tesla fan boy but this was very clearly a driver fault. The equivalent of following your sat navs directions and driving off a cliff.

 

[wdolson]

I think this has been discussed elsewhere on the forum, but AP has a major limitation that Tesla talks about in the AP documentation, but most people don't pay enough attention. AP "see" anything that is moving and when something it knows can move stops, it knows that space has a moveable thing in it. However if AP is presented with something it hasn't seen before that is already stopped, it can't figure out if it's another car or if the road suddenly goes up a steep hill and it doesn't try to stop.

All the accidents I've heard about with AP have happened when the car the Tesla was following zipped around a stopped car and AP didn't see the stopped car. That's something the driver needs to aware can happen and be alert for it.

 

[HyperMiler]

The issue mainly seems to be that the software algorithm is not sophisticated enough yet to combine the information of "stationary object in flight path" and "lead vehicle performs evasive maneuver (brake + steer)" into "prepare for evasive maneuver by a) following lead vehicle if inside lane is clear or b) alert driver or c) = a) + b)".

Nominally, according to the manual this is of course still correct and all that. Notwithstanding driver's responsibilities, to completely fail to initiate a warning, evading, or even concluding an emergency breaking maneuver in such a dangerous situation by something calling itself AP is a bug, not a feature. The information is all there, so this really is a programming and philosophy issue, not insurmountable.

The "driver touched brake so it's not our fault" is a pretty lame excuse for this outright failure to perform. Tesla can do better than that, and I am sure they will, eventually.

 

[Tam]

...the software algorithm is not sophisticated enough yet...

Agreed. It is desirable that the system should know how to avoid collision in such as this case.

However, it is still in its infancy. It's rough enough to make sure it can follow a leading moving car in front and stay dead center inside a lane with no deviation even when there's a Stopped Schubiger Van in this case.

I believe the system will get better and some day, it will be able to steer itself away from the center of the lane to pass the Stopped Schubiger Van without colliding to anything but not with this current stage, again per the manual.

I hope the Automatic Emergency Brake will be redesigned to brake to a stop to avoid collision in future, not the current design of just reducing the force of collision.

But that's the way it is right now, so babysitting your Autopilot is still a requirement.

 

[Vitold]

Mobileye did not detect the van in front and did not pay attention to hazard lights. Plus, Vans coloring may have worked like a camouflage to Mobileye's machine vision

I believe AEB is using ultrasonic sensors therefore it has low reaction time and is unable to stop in time at higher speeds.

 

[wdolson]

When your moving at any speed over a crawl, the ultrasonic sensors are useless at detecting anything in front of you. They only have a range of about 16 feet max. This is adequate to know if a car is in the next lane in your blind spot, but not for forward view. For forward sensing on the road, it uses the camera and the radar. The camera today only detects lane markers and reads the speed limit signs. The radar is the primary sensor for detecting cars in front of you.

The problem is the radar can detect things that are moving from a doppler signature. The radar detects the road surface by knowing how fast the car is going and anything moving at the exact opposite velocity as the car is considered road. For example if you're doing 60 mph, the road is going -60 mph as far as the radar is concerned. Anything that is not moving at -60 mph is assumed to be another car. The radar can tell the speed of that other object and if it is directly in front it sets it's speed to keep x distance from that object.

The radar does not see like a human. In fact the optical camera probably doesn't see like we do either, it's probably tuned to do its job and not see much else. The radar sees and works out everything based on its relative position in space and the differences in velocity. When it sees a car for the first time, it that car isn't moving, it sees road, not a car. The van wasn't camoflauged because of its color or shape, it was camouflaged because it wasn't moving the first time the car saw it. If the van started moving right after the car saw it, it would have figured out it was a vehicle and set a distance, but it being stationary, the radar couldn't detect it was not clear road.

To solve this problem, Tesla needs to add new sensors that aren't there. The current sensor suite is not capable of gathering the data needed. Doppler radar can only tell you within it's range and cone if there is something moving and how fast it's moving, that is the only data you get. It's fixed and doesn't move like aircraft radars that sweep a pattern. Other types of radar could tell you the angle of the object with respect to you, which you could use to figure out that the rear end of that vehicle is the wrong angle to be the road going up a hill, but again, that's a different type of sensor.

The schematic of the AP 2.0 system that has been floating around is a 3 camera system with an ultra wide angle, and two narrower views. With that system you can gather a lot more information about your environment including pedestrians crossing in front of the car, traffic lights, and detecting stopped cars in front of you.

 

[Vitold]

The schematic of the AP 2.0 system that has been floating around is a 3 camera system with an ultra wide angle, and two narrower views. With that system you can gather a lot more information about your environment including pedestrians crossing in front of the car, traffic lights, and detecting stopped cars in front of you.

I doubt that more cameras would help in this situation. The car (obstacle) is right in front of Tesla and for a long time. Tesla needs either better radar or more of it.

 

[electracity]

I doubt that more cameras would help in this situation. The car (obstacle) is right in front of Tesla and for a long time. Tesla needs either better radar or more of it.

"More cameras" probably means stereoscopic range finding. The problem with one camera is the need for motion to distinguish foreground from background.

I don't understand why the radar didn't see a blocked path.

 

[bhzmark]

That looks more like it was just tacc without full auto steer autopilot working and tacc assume that the driver would drive around the car parked on the side of the road like TACC has to assume all the time when it is engaged driving past parked cars.

 

[Bladerunner]

I'm no Tesla fan boy but this was very clearly a driver fault. The equivalent of following your sat navs directions and driving off a cliff.

The car was not on AP, driver was just using TACC (driver participated in the discussion over at elektrek.com).

Most probably the car was tracking the car that merged, "assuming" the driver would follow that path.

So no AP involved, just TACC which obviously has no situational awareness.

Not sure whether such a difference between AP and TACC is clearly evident for a driver at all times, I think it's a usability issue.

 

[kort677]

there is no problem with the car, the problem was operator error. you need to be able to assume control of the car in an instant. one poster noted that this is the equivalent to driving off of a cliff because the nav directed you there.

 

[Vitold]

there is no problem with the car, the problem was operator error. you need to be able to assume control of the car in an instant. one poster noted that this is the equivalent to driving off of a cliff because the nav directed you there.

With current version almost anything that happens is drivers fault so it is a silly argument to make.

Some make argument that system does not give enough time to react. On one hand ne would expect that AEB should save the day but current version clearly does not offer that (which puts Tesla behind other manufacturers).

Still, this situation is obvious enough that driver should have reacted.

 

[xG35]

Here is how the real world is going to work...

Car manufacturers (regardless of which one) will never be liable for any kind of damage while the car is in motion; regardless of what kind autonomous driving you have.

HOWEVER, the day they can claim that... all auto insurance company days will be limited and car manufacturer will be underwriting those as part of the car purchase.

 

[Ashkenaz]

I have only driven the Tesla once, during my test drive. I'm waiting while mine is in production, and I'm very excited about it. One of the main attractions for me was the autopilot. As a professional pilot, the Tesla A/P was a real draw for me, and as someone who uses an airplane autopilot, I'm assuming that it requires significant monitoring by the driver, and I'm hoping that these two accidents were "driver error," that being said, I will be very cautious using mine - but I would love to know more how the system actually works, the hardware particularly. I understand the doppler radar... is that what it is really using?

 

[thimel]

I have only driven the Tesla once, during my test drive. I'm waiting while mine is in production, and I'm very excited about it. One of the main attractions for me was the autopilot. As a professional pilot, the Tesla A/P was a real draw for me, and as someone who uses an airplane autopilot, I'm assuming that it requires significant monitoring by the driver, and I'm hoping that these two accidents were "driver error," that being said, I will be very cautious using mine - but I would love to know more how the system actually works, the hardware particularly. I understand the doppler radar... is that what it is really using?

It also uses a camera mounted near the rear view mirror. This detects the lines and speed limit signs and also the cars. It uses the ultrasonic distance sensors also. These move the car off center in the lane if another vehicle or wall is near the line.

 

[Tam]

;;;So no AP involved, just TACC which obviously has no situational awareness...

"Autopilot" is a generic term that includes a host of features including TACC.

AutoSteer wouldn't help in this case because its current design is to keep it dead center of a lane while an appropriate action would be to steer away from the center and avoid the Stopped Schubiger Van.

AutoSteer is not currently designed to steer away from the Stopped Schubiger Van in contrast with the appropriate action that was manually taken by the locked-on leading Black Wagon.

Without timely intervention from a driver, whether TACC, AutoSteer... were on or off, the collision would still happen.

If AutoSteer was off, then there are more blames to the driver because the car didn't steer itself to the center of the lane with the Stopped Schubiger Van, the driver did all the steering without any attempt to steer away from the center of the lane to avoid the collion!

 

[wdolson]

I have only driven the Tesla once, during my test drive. I'm waiting while mine is in production, and I'm very excited about it. One of the main attractions for me was the autopilot. As a professional pilot, the Tesla A/P was a real draw for me, and as someone who uses an airplane autopilot, I'm assuming that it requires significant monitoring by the driver, and I'm hoping that these two accidents were "driver error," that being said, I will be very cautious using mine - but I would love to know more how the system actually works, the hardware particularly. I understand the doppler radar... is that what it is really using?

Tesla's AP 1.0 uses the Mobileye EyeQ3 which just doesn't have the I/O to handle more than what Tesla is putting into it. The EyeQ4 which can handle more I/O devices is out now, but I don't believe anybody has put it into a production car yet. Tesla test mules with more AP hardware have been spotted and there is speculation that they are testing the EyeQ4. The schematic that was leaked also shows extra cameras.

AP 1 is a big advance, but it's like the aircraft of WW I. Definitely better than what was flying only a few years before, but a long ways to go before it was anything close to a mature technology.

wk057, the white hat hacker who posts here on this forum and who has discovered many things about Teslas (as well as informed Tesla of bugs and security holes he found before they became a problem) has discovered how to equip a pre-AP Model S with a functional AP 1. It cost about $10K in parts and labor. If someone can do it in his garage, Tesla should be able to offer an upgrade package at some point which might cost less than what wk057 did.

The classic Model Ss would need the most work to upgrade, but AP 1 Teslas might be upgradable for something less, say $2500-$5000. There are rumors the ones being built today have the wiring for AP 2, but no hardware installed yet. That could be an even cheaper upgrade.

If they do offer that, it probably won't be until the AP hardware is mostly mature and all cars will be able to upgrade to the same level.

If I was Tesla, I'd be thinking about it. They could make a small profit on the upgrade kits and it would make the owners of older Teslas quite happy. About the only tech the owners of older Teslas want are AP, dual motor (for the original rear wheel drive cars), and a bigger battery. Tesla has already said the 85/90s can upgrade to larger batteries when they become available. I don't think anyone upgraded to a 90 KWh battery because it really wasn't worth the cost for a small bump in range, but when the batteries get up around 120 Kwh people will probably start upgrading those old 85 and 90 packs to the new batteries.

Dual motor is the toughest thing to put on an older car, but a classic with a battery upgrade and mature AP would be a popular car.

 

[bhzmark]

AutoSteer wouldn't help in this case because its current design is to keep it dead center of a lane while an appropriate action would be to steer away from the center and avoid the Stopped Schubiger Van.

AutoSteer is not currently designed to steer away from the Stopped Schubiger Van in contrast with the appropriate action that was manually taken by the locked-on leading Black Wagon.

Without timely intervention from a driver, whether TACC, AutoSteer... were on or off, the collision would still happen.

I think autosteer would have helped in the sense that autosteer and TACC can work together to remove the assumption that the car will steer around the blocked lane. Together autosteer knows that it has to stay within the lane, and TACC knows that the lane is blocked -- together they will produce the result that the car must stop.

But wIth only TACC on, TACC will usually assume that the driver doing the steering will follow the car in front of it, even (especially?) when the followed car reveals a stopped obstacle.

Note that the driver says in the comments to the Youtube video that this just TACC -- only the adaptive cruise control -- not the autopilot.

So the text for the link and the title of this thread is wrong.

I think if he was using full autosteer with TACC (i.e., AP), his car would have slowed sooner, and perhaps stopped behind the truck.

 

[ecarfan]

Note that the driver says in the comments to the Youtube video that this just TACC -- only the adaptive cruise control -- not the autopilot.

I read the YouTube comments shown at

It appears the Tesla driver's name is Chris Thoman. In the comments he posted he does not specifically say he was only using TACC. He states that TACC was on but does not state that AutoSteer was off, as near as I can tell. But it doesn't matter if it was on or off, AutoSteer will not steer away from an obstacle.

He says "I do understand how the system works and I red (sic) the manual twice" and also "Autopilot and TACC is made specifically for these kind of traffic situations".

Yet in another comment he states "At least the ultrasonic sensors should have seen the object right ahead. They reach 20m if I recall correctly".

That shows that he does not fully understand how the current Tesla ultrasonic sensors work, since Tesla clearly states that they only have a a 16 ft (about 3 meter) range, and in any case TACC relies on the forward facing radar and not on the ultrasonic sensors to maintain distance and detect other vehicles in front of the car.

Unfortunately, many Tesla drivers do not understand the capabilities and limitations of TACC and AP. I do not place all the blame for that on them. Tesla deserves much of the blame, in my opinion, because Tesla does not appear to spend enough time educating owners when they buy their car. But it also seems clear that some Tesla drivers do not read the manual or do not absorb the information in the manual, since they sometimes put too much trust in TACC and AP.

In these early days of partially self-driving cars, owners and manufacturers have a lot to learn.