Blog Tesla Releases Data on Utah Autopilot Crash

I'm not sure why folks are so fixated on the fact that "no hands detected" is not perfect. After all, the car doesn't report "no hands present". It reports they weren't detected. The detector isn't perfect. What is? Sure, it would be nice if it were even more accurate, as long as that wasn't at some other unacceptable cost.

(Granted and agreed that the report from Tesla about this crash was at best sloppy and at worst duplicitous, they say "she had her hands off the wheel for more than one minute each time" when of course they should hedge that as "the car didn't detect her hands on the wheel for more than one minute each time".)

 

I think it is precisely the duplicity that causes the fixation. Repeatedly Tesla either states or encourages others to state something that they simply cannot establish given their system in order to castigate drivers and to influence public perception.

 

As plenty of others have said, Tesla's data on a driver "holding" the steering wheel is flawed. I sometimes wonder if there are more false positives and false negatives than good data for that? The first visual warning is easy to miss if your eyes are on the road and the flashing visual warning has taken me by surprise several times. Bad data is worse than no data. The problem I have with Tesla's response in this case is that they seem to suggest that their data is flawless.

 

One additional data point: I used AP off and on for ~700 miles from South Carolina driving to Maryland. Multiple disengagements at my request due to traffic interactions. Multiple warnings received that I didn't have my hands on the wheel; most of these were surprises to me because ***MY HANDS -- BOTH HANDS! -- WERE ON THE WHEEL AND I WAS ABLE TO CONTROL THE CAR IF I NEEDED TO DO SO***.

I speculate that there is some pressure-sensitive mechanism rather than a capacitive mechanism and so the car is unable to distinguish a low-pressure grip on the wheel from no grip on the wheel.

Incidentally, there WERE occasions on this recent drive of mine where I did NOT have my hands on the wheel for a long enough period that the car's warnings triggered. Those incidents were valid.

Short of additional insight provided by Tesla, I would have no way to distinguish a priori the warnings triggered when my hands were actually on the wheel from those triggered when my hands weren't. Therefore, my personal experiences on this drive lead me to doubt the accuracy of Tesla's reporting on drivers actually having their hands on the wheel.

Alan

 

It's known that the mechanism is a torque sensor. You can test this yourself by applying mild resistance to wheel motion, or mild deflection to wheel position. The car will "detect your hands" and be happy. You don't have to apply enough torque to disengage Autosteer, nowhere near.

Folks say they've tested defeating the hands-on-wheel mechanism by hanging a small weight on one side of the wheel. I haven't tried this myself and of course, don't encourage others to.

 

In my flying career, having to do something quickly usually meant I failed to do something earlier, and more smoothly. Exceptions? Sure. Getting upset through inverted by a heavy helicopter's tip vortices at low altitude required prompt attention (my daughter still remembers that one). And flying formation is a place no living aviator employs an autopilot. Eyes outside is a fine idea, though, and I heartily agree with you about that, as well as the all-but-intractable problem of driving distracted with your eyes on a little blue screen, instead of the road.
Robin

 

In the Model 3 you can easily get to the blue flashy screen level of warning even with your hands on the wheel due to the screen placement. The first “Hold the wheel” notice is located towards the bottom of the center screen where it is harder to catch if you are watching the road. I would think even in an S you could get to the 2nd level warning with your hands on the wheel and your eyes on the road if you aren’t checking your speedometer that often.

So I think trying to determine actual hands on wheel based on response time to the warning is probably not indicative.


I am very certain there is a level of variance as well on the wheel torque sensors. My AP is difficult to disengage with the steering wheel - the car ends up fighting me a bit and then releases and I end up jerking the wheel despite my best efforts. Our Model S doesn’t do that, and other people have reported varying levels of force required to disengage AP. I took my car in to Tesla to see if it was calibrated properly, but the local Service Center can’t change that setting - it comes that way from the factory. Because of that, I have to keep a heavier hand on the 3 than I do our S to avoid hand on wheel warnings.

 

Take a look at the antenna pattern on page 10 of http://www.ti.com/lit/ug/tidudq6/tidudq6.pdf. You will see that the vertical resolution is about 35 degrees and horizontally it is about 80 degrees. At a distance of 60 feet, this gives you a pixel that is 36 feet tall and 84 feet wide.That is a whole lot bigger than a fire truck (thus no "vertical step change" in that case, quite a gradual blob in fact). It would certainly encompass any overhead signs or signal poles. That's at 60 feet and of course it gets bigger in direct proportion to the distance. At 60 MPH, you will travel 60 feet in 0.682 seconds and the stopping distance for the model S at 60 MPH is 118 feet (maximum braking) so to avoid this collision you must completely recognize the danger at at least 118 feet. At that distance the pixels are 71 feet by 165 feet.

Consider another common scenario: stopped cars ahead in the lane to the right and to the left but none directly in front. Scan your radar across this (say 100 feet ahead) and you will conclude all lanes are blocked. Slam on the brakes?

Hopefully you can see the challenges with false alarms for stationary objects as far as radar goes because the world is full of stationary objects and the probability of very large metal ones in the antenna side lobes is quite high and the main lobe is also necessarily quite large. It is a fact that clutter is a problem for all radars. For moving objects which can be distinguished by Doppler shift it is much, much easier because the clutter goes away. Cameras have the potential to solve this and lidar certainly can. Tesla's camera system is good at recognizing most stopped vehicles, but far from 100% today, but it is improving.

 

On my AP1 MS I find the "effort" to disengage AP depends on its confidence. Normally it takes quite a lot of force, or a sudden tug, as you described but on a tightening bend, over a crest I can disengage AP with slight pressure from my little finger - which I think is useful feedback as to the car's "confidence" ... but I don't know if that is the same on AP2 MS or on M3?

 

That's why the "orange in the steering wheel" hack works. There is constant torque applied to the steering wheel, which keeps the nag alerts from triggering.
I'm not sure if that's how the APBuddy (god, what a disaster in waiting) works though.

 

Because fine print is never very carefully worded for any other product on the face of the planet. And Tesla is the only company on the planet to do it.

Strawman to go down the ‘who reads the fine print/manual/instructions’ et al. Irrelevant. Ignorance is not a valid defense. As an adult you are fully aware that stuff is always buried in the fine print and that manuals provide all sorts if goodies on operational, risk etc... fronts. If not, too stupid to live comes to mind.

It never ceases to amaze me how otherwise intelligent people will do everything in their power to not take responsibility for their own actions and come up with a myriad of excuses as to why someone else is responsible for them having put their underwear on inside out that morning. Indeed I just did that yesterday and I’m sure it was because the guy down the street walked his dog after 8pm.

 

Could that be the difference between "Car did not detect any resistance whatsoever for prolonged period" and "Car noticed some resistance, some of the time, but because of gaps issued some reminders" ?

But, yeah, I agree more likely mitigating lawyer-speak entering into the publicist's "copy" ...

FWIW I think that's the problem. Even hand pressure, e.g. driving an 10-to-2 probably does not create much/any rotational torque.

Where'as my driving with one hand at 4 O'Clock definitely does create rotational torque... but, now I think about it, I am NOT honouring the "keep both hands on the wheel"

 

True, I have noticed some variability in disengagement force in different situations. Around curves it is less than grabbing the wheel to avoid a wandering semi on a straightaway. Overall though, the force is still higher than our S it seems like. Maybe the smaller wheel diameter factors into that.

 

Presumably the implication is that if you have to keep switching the darn thing on and off then maybe current traffic conditions are not really appropriate for AP.

Unless of course the driver is desperate to finish typing a text but can't decide on a fourth emoji cos the car keeps interrupting with something boring about....you know..... driving.

 

Time to sing along : The Autopilot Song

 

If your hands are on the wheel and the road is flat and straight this can happen. Since the wheel is not turning you are not resisting any turning of the wheel by AP. Hence no torque is detected.

The other scenario is when there is subtle AP turning and the driver is "unconsciously" making the same adjustments. Again, no resistance to AP turning, no torque detected.

 

I still don't get why the cruze control part of Autopilot did not stop the car, or at least reduce the speed big times before the crash here. My beloved Tesla Model X 100D always stops when car ahead of me reduces speed or stops, just with intelligent cruze control, and with this huge metallic truck it should have seen it, or use its other emergency brake feature. Why nothing stopped it ? When she finally pressed the brake at last second she for sure dis-engaged the cruze control and may be the emergency brake too. May be she interfered with the automatic tools doing that ? I was told that for emergency brake to work you need to brake and then it takes control of maxi braking for you. If not nothing happens. Strange....and can be confusing. But cruze control should have helped far ahead of that crash. Is it the fact that the truck was stopped on the road far before she arrived ? I'm still puzzed on this one, using my beloved Tesla intelligent cruze control very extensively and never had any issue with it when arriving at a red light with people already stopped ahead of me. It always stopped my car, reminding the speed set and returning to it when cars ahead move again to that speed. At worst sometimes the car does not restart till I press the accelerator a bit. But for stopping it always stopped so far.

 

Since September 2016, as described in Upgrading Autopilot: Seeing the World in Radar, Tesla started using radar "as a primary control sensor without requiring the camera to confirm visual image recognition." This means Tesla depends highly on radar (and cameras are only used for lane markings detection) for implementing their TACC+AS, ie. Autopilot 1 and 2. Looking at the Bosch radar's data sheet (I couldn't find Continental ones), we can find the following wordings when an ACC application is proposed, "At speeds of up to 150 km/h (93 mph) and a maximum relative speed of up to 80 km/h (50 mph), the system automatically maintains a set distance from the vehicle ahead..." I certainly do not hope that Tesla radar engineer simply took the Bosch code as their AP (or even AEB) specification without improving it, or considered the vendor's proposed 50mph target as the radar's limitation.

No matter what the driver is still responsible.

 

I thought AP1 did not have a radar?

 

It does have radar. It has only one camera.