Self Driving Uber Fatality - Thread III 4

[drawoh]

This thread is continued from thread815-437388.

--
JHG

 

[GregLocock]

Well, I guess we'll find out from the final NTSB report. But all the youtubes I've seen look as though they predict a single future path for each object.

Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376

http://eng-tips.com/market.cfm?

 

[VEBill]

Per the traffic laws, you are not supposed to pull out into the intersection until it is clear to make the turn.

The rules vary. In every jurisdiction that I'm familiar with (Canada), those turning left may enter the intersection and wait for opposing traffic to clear. When the signal lights change, the Right of Way remains with them, as they are already in the intersection. In other words, those given a green light must yield to those that are already in the intersection, giving them time to complete their left turn. Typically one or two vehicles will do so with each cycle.

"The vehicle already in the intersection has the right of way ahead of any car that has not yet entered the intersection."
Ref: Rules of the Road, Nova Scotia Driver's Handbook, Chapter 2 [PDF], page 44.

It appears that Colorado hasn't clearly defined the rules. Appears muddled.
Ref: The Denver Channel's view

Autonomous Vehicles (as well as humans) would need to be aware of the local rules and habits. Assuming that the rules are well defined.

 

[LionelHutz]

At 1.3 seconds before impact, the self-driving system determined that emergency braking was needed to mitigate a collision. According to Uber emergency braking maneuvers are not enabled while the vehicle is under computer control to reduce the potential for erratic vehicle behavior. The vehicle operator is relied on to intervene and take action. The system is not designed to alert the operator.

This part makes no sense. I can understand that the programming might be attempting to limit the amount of false emergency braking. But, if the system DID determine it should be emergency braking, they why wasn't it emergency braking? It's rather hard to believe the default programmed behavior is to just drive into something it knows it must avoid, which makes me believe this report is not properly giving the facts.

I wonder if they meant the Volvo emergency braking system determined it should be braking at that point???

The vehicle operator said in an NTSB interview that she had been monitoring the self-driving interface and that while her personal and business phones were in the vehicle neither were in use until after the crash.

This tells me the setup of the vehicle was also at fault if the interface was on the console and the operators are expected to watch it instead of using the large transparent safety devices installed in the car. That makes the operators also rely on the computer systems instead of being a backup to them.

 

[LionelHutz]

OK, at least, the sensing system performed exactly designed, including detecting the pedestrian at over 500-ft distance, as expected.

The report didn't say it detected the pedestrian crossing the street at 500ft, it just that started detecting what was possibly a physical object at that distance.

The rest is a complete fubar; it basically allowed nearly 5 seconds to lapse before figuring out that it couldn't brake and needed to warn the driver. No wonder Uber is shutting down that operation; the people that came up with that logic should never be allowed to work on anything safety related again. This is one of the few times I think licensing such engineers ought to be a requirement.

The article says the classification went from object to car and finally to bicycle. The expectation of what each object type might do is different. The unknown object might be expected to stay in place safely off to the side of the road. The car might be expected to travel in a parallel direction since there were no side streets or driveways or parking lots entrances in that area. The final determination that there was a bicycle in profile view could finally change the expectation to it possibly crossing paths with the car. By the report, it would appear that the proper determination of what the object was and how it was travelling occurred less than 2 seconds before the collision which is way too late to be getting it right.

 

[drawoh]

LionelHutz,

I get 43mph for six seconds as 115m, or 380ft. If a LiDAR could scan ahead 115m and identify you as LionelHutz, poster on Eng-Tips.com, and work out whether to run you over or not, it would be remarkable. Any object identified by a robot must not be run into. The space the object occupies must not be passed through. The robot must avoid the object by some minimum distance because it may do something unpredictable, and there may be something behind the object that is moving. Somebody mentioned the case of a vehicle running through a puddle and kicking up a splash that was detected by the following LiDAR, causing the robot to take evasive action. Any algorithm that identifies objects that can be run through, must be absolutely reliable.

I was on the highway a couple of days ago when I approached a truck on the inside shoulder. The passenger door opened and a guy got out and signaled me to move across a lane. All I did was slow down and move to the outside edge of the lane. I did not have time to check my blind spot and mirrors, especially since I was no longer matching the speed of traffic. Robots must cope with this.

Uber's decision to not take evasive action meant that the driver was in control, and that she should have been watching forward, gripping the wheel and holding her foot over the accelerator and brake pedals. If Uber wanted a computer fiddled with, they should have provided a passenger.

--
JHG

 

[VEBill]

Uber's decision....the driver was in control, and that she should have been watching forward, gripping the wheel...

Hold the steering wheel to maintain possible back-up control, but don't apply any pressure else...

The operator can transition from computer control to manual control by providing input to the steering wheel, brake pedal, accelerator pedal, a disengage button, or a disable button.

Reminds me of:

https://en.wikipedia.org/wiki/Aeroflot_Flight_593#Accident

 

[drawoh]

At 1.3 seconds before impact, the self-driving system determined that an emergency braking maneuver was needed to mitigate a collision...

I have just crunched this number. At 43mph, a panic stop at 0.9G (achievable?) takes 2.2s, and 21m (69ft) of road. There is nothing the human could have done even with infinitesimal reaction and evaluation time.

--
JHG

 

[IRstuff]

"The article says the classification went from object to car and finally to bicycle. The expectation of what each object type might do is different. "

That's not really relevant; a moving bush, or whatever, moving into your lane is problematic, particularly when the object is in a traffic lane, and isn't going with the flow of traffic. Even in a most basic use-case, an unknown object in an adjacent lane ought to demand a decrease in speed, particularly when the object itself presents a blind spot to the driver.

The bottom line is that the fundamental target processing was flawed; the difficulty in classification itself is problematic, and should have caused at least a slow-down, all by itself. It's often typical of people designing these sorts of things to ignore what the system did in the past frame(s) and only look at the current frame as the end-all and be-all for decision making. Algorithms exist for doing this sort of thing.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert!

https://www.youtube.com/watch?v=BKorP55Aqvg

faq731-376 forum1529 Entire Forum list

http://www.eng-tips.com/forumlist.cfm

 

[GregLocock]

drawoh I was quoting the report. It says mitigate, not prevent. If it had started an entirely achievable 0.9g stop at 1.3 seconds it would have hit her at less than 15 mph (in my head), a far cry from 43 mph. Admittedly I think it should have decided it needed to stop or match speeds long before then.

My guess is that AEB was not included in uber's software because the intention was to have a full time driver and an observer. Thus the AEB could be left out. Later on they eliminated the driver as a cost saving but didn't add AEB because it hadn't been needed up to that point.



Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376

http://eng-tips.com/market.cfm?

 

[IRstuff]

At any time in the prior 5 seconds, a reduction in speed might have limited the damage to non-fatal broken bones. There were numerous times during that interval that the system should have concluded that the risk of not reducing speed was too high.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert!

https://www.youtube.com/watch?v=BKorP55Aqvg

faq731-376 forum1529 Entire Forum list

http://www.eng-tips.com/forumlist.cfm

 

[GregLocock]

"The expectation of what each object type might do is different. "

6 seconds out object 1010b type unknown closing velocity 43 mph range 380 ft impact in 6 seconds
5 seconds out object 1010b type A closing velocity 43 mph range 315 ft impact in 5 seconds
4 seconds out object 1010b type B closing velocity 43 mph range 250 ft impact in 4 seconds

etc


No, I don't think you need to classify the object in that case to figure out that you are on a collision course. Classifying the object is less important than avoiding a collision. You'd have taken your foot off the throttle wouldn't you?

Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376

http://eng-tips.com/market.cfm?

 

[LionelHutz]

....I get 43mph for six seconds as 115m, or 380ft. If a LiDAR could scan ahead 115m and identify you as LionelHutz, poster on Eng-Tips.com, and work out whether to run you over or not, it would be remarkable. Any object identified by a robot must not be run into. The space the object occupies must not be passed through. The robot must avoid the object by some minimum distance because it may do something unpredictable, and there may be something behind the object that is moving.....

While shaking my head at the stupidity of this argument....the object wasn't anywhere close to being in the path of the Uber 6 seconds before the impact. She was probably still 2 to 3 car lanes away from the path of the Uber when she was first detected. Sorry, but your royally failed at making any kind of serious point here.

I have just crunched this number. At 43mph, a panic stop at 0.9G (achievable?) takes 2.2s, and 21m (69ft) of road. There is nothing the human could have done even with infinitesimal reaction and evaluation time.

Sure, the accident was not avoidable due to the driver not looking out the windshield or being ready to take over. But, the accident would have been very easy to avoid for a driver actually paying attention to the road out of the windshield who was ready to start driving the car.


IRstuff

As I ALREADY posted, from what the report is saying the correct object detection/determination occurred far too late.

But, when the first initial detection occurs, there really is no way to determine the rate or direction of travel yet, is there? The Uber wouldn't even yet know the object was moving, would it?

Once the Uber first decided this object might be another car, it likely expected this object to be either travelling towards it in an adjacent lane or travelling away from it in an adjacent lane. Wouldn't this be logical of the Uber when it detected what it believed was the front end or back end of a car in an adjacent lane? Would it be logical of the Uber to expect this object would cross it's path at a 90* angle or would it be logical of the Uber to expect this object was changing lanes?

Just to make sure it's clear for those that seem to have forgotten. I will point out that she WAS NOT in the lane the car was travelling in from the point of first detection. From the data I'm seeing, she was NOT in the path of the Uber or even in the lane the Uber was travelling in for around 4.5 of those seconds.

 

[GregLocock]

She was on a constant bearing from the car. Watchkeeping 101.

Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376

http://eng-tips.com/market.cfm?

 

[IRstuff]

Assuming your 3 lanes in 6 seconds, that's 3.4 mph; if the car can't do THAT math, we're in way big trouble.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert!

https://www.youtube.com/watch?v=BKorP55Aqvg

faq731-376 forum1529 Entire Forum list

http://www.eng-tips.com/forumlist.cfm

 

[drawoh]

LionelHutz,

LiDAR needs one scan to see the object. It needs two scans to work out the object's velocity, and three to work out the acceleration. If the scanner runs at 10Hz, this take 200ms. This all takes place at around 100m range, which the robot can manage with a fairly gentle stop, or a lane change.

What we think about robots honking horns at us?

--
JHG

 

[LionelHutz]

There's going to be a lot of automated cars braking to a stop for every pedestrian walking up to a street corner in the future then.....

The final report should be interesting. There could have been a conflict between the classification and direction detection due to the classification of the object having a programmed expectation saying that object type shouldn't or couldn't be travelling perpendicular to it's path. Another possibility is the detection first thinking the object was only changing lanes or pulling out of a driveway into an adjacent lane.

 

[IRstuff]

"There's going to be a lot of automated cars braking to a stop for every pedestrian walking up to a street corner in the future then"

In most states, that's pretty much a requirement, given that pedestrians have the right of way; if they get there first, then the car damn well better stop. For intersections with lights, the car should already have the capability to decipher the lights and make valid decisions.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert!

https://www.youtube.com/watch?v=BKorP55Aqvg

faq731-376 forum1529 Entire Forum list

http://www.eng-tips.com/forumlist.cfm

 

[LionelHutz]

For intersections with lights, the car should already have the capability to decipher the lights and make valid decisions.

So, at green lights just keep driving and take a chance at running over unknown objects that are nearby? Otherwise, come to a full stop for any object detected that is within 3-4 car lanes to either side until 100% certain that object won't end up in the travel path? That's an easy enough rule set to follow....

 

[IRstuff]

My point was not to describe the entire solution, but to simply point out that basic rules were already being violated. If Uber's internal review have revealed an unavoidable accident, the outcome possibly would be different than if they found that basic safety rules were violated.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert!

https://www.youtube.com/watch?v=BKorP55Aqvg

faq731-376 forum1529 Entire Forum list

http://www.eng-tips.com/forumlist.cfm

 

[LionelHutz]

Road rules must be different where you are than here where I live. Here, drivers only have to stop for pedestrians at pedestrian crossovers, which are defined by their special signs and markings. They don't have to stop at crosswalks, which are basically defined as the strip of road going between 2 sidewalks. They certainly don't have to stop to allow pedestrians to randomly cross the road jaywalking style.