Model X Crash Fatality and Fire on 3/22/18

Maevra

Top-Contributor
Joined
Oct 24, 2017
Messages
908
Location
Austin, TX
Tesla Owner
Model 3
Country
Country
#21
Here is another streetview picture showing someone drifting left into the "gore area":

View attachment 6825
Based on this photo I wonder how easy it would be for AP to misread the gore area as a legitimate lane? 150 meters away the lines are pretty close together and it's only towards the last 20 meters that the lanes really diverge and the space widens.

ETA:
This ABC7 video has a good view of the road leading up to the gore point (watch minute 1:58) and the difference in color on the asphalt and concrete and where it all meets the white lines.

HOWEVER, ABC7 made that video assuming the driver was in the 85-S lane to get to work (Apple HQ) but this info suggests the driver may have been in the 101-S lane. Totally changes the viewpoint depending on which lane they were in.
 
Last edited:

Joaquin

Well-Known Member
Joined
Jan 15, 2018
Messages
326
Location
Sunnyvale, CA
Country
Country
#22
That whole divider area is just wrong... at least the asphalt between the white lines should be rugged or something to clearly have feedback that you are NOT supposed to drive/cross there!

But anyway, no bueno if the autopilot was engaged. The news will take that out of context and blame the accident to that :(
 

RolfS

Active Member
Joined
Oct 2, 2017
Messages
33
Location
Santa Clarita, CA
Tesla Owner
Model 3
Country
Country
#23
My thoughts related to this accident and autonomous driving issues:


1. The name “Auto Pilot” is misleading. I was talking to a person at the LA auto show, looking at a Tesla Model 3. He said that his wife wanted one, because of the “Auto Pilot”. She liked the self driving feature. I explained to him that is not what “Auto Pilot” is. Not even close. He was not aware of that. And this guy was an avid amateur racer, racing his Porsche. People are just not informed. Using the wrong words to name a feature can end up disastrous for someone. People don’t read their manuals carefully, or don’t understand what they are reading. They think the feature does more that it does. Calling it “Driver Assistant” would be a step in the right direction.

2. CalTrans was negligent in fixing the crash attenuator. There was an accident 11 days prior to this one with a Prius running into the gore point and they did not reset the crash attenuator. (the Prius has lane keeping feature option too. I don’t know if this one did — hmm.) Big mistake, CalTrans said they were delayed because of rain. Not sure that is a valid excuse. They should have put an electronic sign in front of it to direct traffic around or at least a whole set of barrels along the whole divider.

3. CalTrans needs to re-stripe the no drive lines. Put chevrons and rumble strips in that whole area. In general CalTrans with the increase of self driving cars needs to rethink their road markings to make sure that the stripping is repainted everywhere when they fade and that road signs are clear and readable.

4. It was said that the radar does not detect stationary objects. Not sure that is an accurate statement. The radar might or might not. There are some pretty sophisticated radar techniques out there. Here is one example of a proposed system (date is 2003 — so I assume progress has probably been made in the mean time):

https://www.adv-radio-sci.net/1/105/2003/ars-1-105-2003.pdf

If the Tesla radar is anything like that then I would assume it’s a software problem not a hardware problem. The other pure issue I have with radar and active systems is when they get many cars that all use the same active radar all running with different timing, how do they prevent interference from occurring and getting incorrect readings? Radar is based on doppler shift and pulse timing to get an accurate distance and speed differential reading. This is not going to work when everyone has them. You can’t make it a packet with an encoded code, because the reflected signal is not going to be that clear and also doppler shifted. You might be able to detect collisions like on the ethernet and WIFI, but its not a data packet and you measure the return signal not the original one, it is probably frequency shifted and smeared.

The advantage with radar is that it can see in rain and snow and fog, since the wave length is way too long for it to see rain drops. LIDAR can’t see in rain or fog, LIDAR and radar can’t read signs. LIDAR reads the laser bounce back of its infrared laser signal. It has good resolution, but it can’t read the writing on signs it can only see the shapes and the distance. It can’t read traffic lights either. Radar and LIDAR always has to be backed up by passive optical means.

5. Neural networks and deep learning. I know that Elon Musk has stated that we have to be very careful with AI so that it does not get misused. Back in the 1970’s as a young college student, I played around with neural networks. Wrote a Tic-Tac-Toe game using a neural network. Person against the computer. In this game two experts always end up in a draw. At first I had no problem beating the computer. But eventually the computer got better and better. It got to the point that when the game was about to end in a draw and it was the computer's turn it decided to not make the move. It got stuck in a loop. That was surprising to me. As it turned out it was because the back propagation weights on the neural paths where allowed to go to zero on all paths that ended in a draw, so it never took that path. The neuron was inhibited from firing. It was obvious once analyzed, and to change the back propagation to not allow the zero state for a draw. But it was kind of stubborn of it to not make the move at all, like a child refusing to take a bath.

Anyway, back then the computer power to process any extensive neural networks was pretty lacking. I assume today we have the processing power. The process is still in its infancy IMO, but I have not kept up in the technology so I could be wrong. I think it still has a long way to go. I’m not worried about AI taking over the world yet. These systems take millions of trials to get the neural networks to make the right decision on roads, signs, rules of the roads etc. It does not take humans that long to figure things out. If you are shown an object that you have never seen before and then are asked to identify that object in pictures, I bet you could get it right away. You might miss some but that is just because of being careless not that you don’t recognize the shape. But show that to a neural network, and it would have to be trained to learn to recognize that object with many trials. Same thing with a road, something new comes up that it has never seen before it might get flummoxed (well just not know what to make of it), and probably just ignore it. But it could be something real important, which a human can figure out right away and avoid, if necessary. There is still something very important missing in how our brains recognize things and they way these neural networks operate. Something really important!

It needs to be possible to train neural networks faster. They need an language interface (maybe they have) that goes something like this:

a. this is a picture of a new sign.
b. when you see this sign it means such and such, these are the rules — incorporate.

The neural network needs to incorporate this sign into its structure by making new paths, without any new training taking place. That is what humans do they read the drivers handbook (or should) and know what to do — well most do assuming they learned to drive in the first place.


6. Maps can help the neural network. For example: road splits, look for gore point, and make sure you’re not traveling on it. How many lanes, which ones split off. Okay, I hope you get the idea. You can add more on your own.

EDIT: fixed some typos. They always show up after I post.
 
Last edited:

garsh

Dis Member
Moderator
TOO Supporting Member
Joined
Apr 4, 2016
Messages
11,077
Location
Pittsburgh PA
Tesla Owner
Model 3
Country
Country
#24
Great post. I agree with most of it, but would like to point out the bits I disagree with. :)
1. The name “Auto Pilot” is misleading... Calling it “Driver Assistant” would be a step in the right direction.
I disagree with both parts of this statement. I think "AutoPilot" is quite descriptive - it does operate similarly to an airplane autopilot system. I disagree that renaming it to "driver assistant" would prevent people from believing it is self-driving. There are too many reviews of Tesla cars were people turn it on and then exclaim that "the car then drives itself". It's just not a problem easily solved by re-naming.
If the Tesla radar is anything like that then I would assume it’s a software problem not a hardware problem.
I agree. I think the problem is "false-negatives". A close-up coke can is hard to distinguish from a far-away barrier to radar. You don't want the car to panic-stop for a coke can on the highway. They need to incorporate the cameras and some image recognition to help augment what the radar sees.
 

RolfS

Active Member
Joined
Oct 2, 2017
Messages
33
Location
Santa Clarita, CA
Tesla Owner
Model 3
Country
Country
#27
I disagree with both parts of this statement. I think "AutoPilot" is quite descriptive - it does operate similarly to an airplane autopilot system. I disagree that renaming it to "driver assistant" would prevent people from believing it is self-driving.
I agree with you that Autopilot is very descriptive. Too much so. When I first started learning about Tesla, was before they had Autopilot. When they first announced it I thought, wow, that's amazing, the car can drive itself. And many comments and articles attributed more to it than what was there.

It's not even close to an aircraft autopilot, but the average layman does not know that. In todays aircraft autopilot you have incorporated the "Flight Director" (it might be called something else today - I don't always keep up with everything.) It's sort of like a route planner equivalent to mapping software is in cars. It can help you plan your whole route, including waypoints to avoid mountains and other obstacles particularly near landing and takeoff zones near airports where you might be at a low altitude. So they do a lot more than the Tesla Autopilot.

I'm not married to calling it "driver assistant" either, I just threw that out there as an example. The ordinary autopilot for aircraft is more like a fancy cruise control. I think maybe calling it "Teslapilot" would of made people less inclined to jump to the wrong conclusions and actually make them read about it. But, it's too late now the cat is out of the bag, so to speak. I think we are stuck with the term as is.

EDIT: Typos.
 
Last edited:

RolfS

Active Member
Joined
Oct 2, 2017
Messages
33
Location
Santa Clarita, CA
Tesla Owner
Model 3
Country
Country
#28
@TEG, very informative posts, I have been following you posts on telsamotorsclub.com on this subject. I am not a member and did not want to join at the last minute and be called a troll.

How do we get the various highway departments to incorporate these measures? They are needed for safer driving and for the future of autonomous cars, that aren't as smart as we are, but more vigilant if done correctly.

Thanks for your great posts here and there!
 

TEG

Active Member
Joined
Apr 2, 2016
Messages
56
Location
Silicon Valley
#29
I wish I had a good answer to that. Write your congress-person? Keep the dialog going in public forums?

It feels to me like we have grown more lax on upkeep of safety features on our highways.

Does it boil down to lack of funds? Management issues? Just too much traffic to keep up?

I don't know the answers, but felt compelled to point out what looks like improperly maintained equipment, poor road markings and other disappointing things like that. The local news has been making a fuss about this situation, and I bet we will get some extra attention on our local interchanges due to that. But does that mean they will cut corners somewhere else?