Eng-Tips is the largest engineering community on the Internet

Intelligent Work Forums for Engineering Professionals

Register Log in

Self Driving Uber Fatality - Thread IV 7

Status
Not open for further replies.
Replies continue below

Recommended for you

Sort by date Sort by votes
The preliminary NTSB report is available Not much is new

I've read a couple of news articles, speculating on Uber's processor failing to consistently detect the pedestrian as such, and not have a "jaywalking" detection.

If those comments are true, the Uber missed a more fundamental requirement, which is to not collide with ANYTHING, regardless of whether it's identifiable, and regardless of whether it's in a crosswalk or not. Another failure was the software's inability to string the detections into a single coherent story; instead, it apparently treated each detection as a completely standalone event, and the fact that the previous detected object disappeared and a new one appeared along the same trajectory didn't have any processing to link these events into a basic "unknown, inconsistent, object" on a collision course.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert! faq731-376 forum1529 Entire Forum list
 
Upvote 0 Downvote
Constancy of existence is one of the first things that infants learn.

I think I've mentioned earlier that one mystery of intelligence is the ability to create an internal simulation of the outside world and that so far no one understands the mechanism that animals use to do that, much less how to make machinery that can mimic that ability in a general purpose manner. Soon after that happens an AI will be able to simulate itself, a key step to self-awareness. My guess is that it is limited because no one has been able to build a self-organizing network with even a tiny fraction of those available in animal brains that are necessary to do the mass-parallel process pattern matching required.
 
Upvote 0 Downvote
Sometimes, colliding with something is the best option, though- say, when something like a plastic bag goes blowing across the road.
 
Upvote 0 Downvote
That's if you can tell it's that; but this was a big object, that was occasionally detected as a pedestrian and attached to the ground. Regardless of that; the software didn't even attempt to consider the object sequence to be an issue at all, until the victim was essentially 1.3 second from impact, i.e., she was completely within the path of the car and collision was pretty much unavoidable.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert! faq731-376 forum1529 Entire Forum list
 
Upvote 0 Downvote
In the normal course of driving pedestrians are far more often on a collision course than plastic bags.

Surely the primary priority is dealing with avoiding hitting objects whatever they might be.

Challenges of plastic bags and birds are certainly present but the default should be avoiding collisions.
 
Upvote 0 Downvote
I run over plastic bags on my commute at least once a week, probably more. Have not come close to running over a pedestrian yet but I only encounter pedestrians within my neighborhood, less than 5% of my drive. But I drive more slowly and cautiously when they are present. The Uber should have slowed down when it first started to detect objects. The fact that it detected objects and discarded the detections leads me to believe they get a lot of false detections and just treat them as noise. Unfortunately this time they were not noise.

----------------------------------------

The Help for this program was created in Windows Help format, which depends on a feature that isn't included in this version of Windows.
 
Upvote 0 Downvote
I worked on a collision avoidance system over 25 years ago, and we dealt with false detections by looking at spatial correlations, which weeded out almost everything that was truly false. They had multiple detections correlated along a trajectory; it should have been relatively easy to figure that out.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert! faq731-376 forum1529 Entire Forum list
 
Upvote 0 Downvote
dgallup said:
I run over plastic bags on my commute at least once a week, probably more. Have not come close to running over a pedestrian yet
Click to expand...
Fair call. My implicit assumption that my driving is representative of all is far from ideal. Some people might often encounter plastic bags, others tumbleweeds, others snow drifts. (I've never come close to running over a pedestrian, but having to slow to give way to them is a daily requirement.)

This variation highlights the difficulties faced by object detection. Needless to say, pedestrians, cyclists and other vulnerable road users need to be protected and they are more difficult for the AI to recognise and predict than other vehicles.

dgallup said:
The fact that it detected objects and discarded the detections leads me to believe they get a lot of false detections and just treat them as noise. Unfortunately this time they were not noise.
Click to expand...
It almost seems like it has requires the system to confirm the object before it reacts to it. A cautious approach would be begin reaction until unknown object is identified and a collision is ruled out.

It is easy to be armchair experts on this. I don't doubt the challenges are significant.
 
Upvote 0 Downvote
My armchair expert opinion is that whoever is (was) in charge at the system level was way out of their depth.

"Schiefgehen wird, was schiefgehen kann" - das Murphygesetz
 
Upvote 0 Downvote
I completely agree. We had, nearly 20 years ago, a somewhat similar case. We were supposed to be tracking a target using an existing tracker in a new hardware configuration. In the first field test, the tracker was misbehaving, so I asked why that was; "I re-acquire the target every frame." OK, that's bad, but not catastrophic, so what about the track history, "Oh, I don't use one." ARRGGGH.

I can see that his descendant obviously worked on the Uber obstacle processor; "Sure, there's a new detection after the previous detection, but I don't bother the processor with keeping track of things like that."

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert! faq731-376 forum1529 Entire Forum list
 
Upvote 0 Downvote
I've nearly hit a pedestrian. I was turning left, at night, across 4 or 5 lanes. There aren't many pedestrians in this area, and I didn't see anyone waiting to cross. It was easy for the man to avoid, but he was still angry with me. There were 6222 pedestrian deaths in 2018. In 6221 of those cases, the driver was a human; in one case the driver was software.

 
Upvote 0 Downvote
"A.I. is hard.'

"A.I. outdoors is really hard."


The first is a very famous quote, and the word 'hard' is a hilarious understatement.

The second is my extension. The problem with 'outdoors' is that it is monumentally unpredictable.

The issue with making comparisons to human drivers is that autonomous vehicles will tend to concentrate the liability onto the manufacturer. If (for example) Ford has to pay wrongful death compensation of hundreds of victims per month, they won't last a year. AVs have to be orders of magnitude safer than human drivers, unless someone figures out a whole new liability insurance scheme.

They're not going to be sufficiently safe if they're still crashing into things like a blind drunk.

 
Upvote 0 Downvote
IRstuff said:
I've read a couple of news articles, speculating on Uber's processor failing to consistently detect the pedestrian as such, and not have a "jaywalking" detection.
Click to expand...

The processor observed the object at 6[ ]seconds, and identified it as a vehicle, and then a person with a bicycle, as it closed the distance. It recognized the impending collision. The failure was that all the systems that could have taken evasive action, including the driver, were disabled.

If you are driving behind a robot controlled vehicle, I would expect you to see different weird, erratic behaviour that what you would get from a human driver. As you point out, robots must be programmed to not smash into anything, ever. If you do not have an algorithm that reliably identifies empty plastic bags or puffs of smoke, the robot will have to slow down or take some other evasive action. This will be annoying as hell if you are driving behind them. It is presently annoying as hell to be tail-gated as you drive 20kph over the speed limit.

--
JHG
 
Upvote 0 Downvote
It recognized the impending collision.
Click to expand...

I don't think so; unless we're talking about different things; the Uber processor did not recognize the impending collision until 1.3 seconds before impact. The reason is that the Uber processor didn't consider the individual detections as a coherent whole; that was what I was referring to; the processor should have had an over-watch algorithm that would have simply considered the detection as objects and strung them together into a single object moving on a collision course. But, because the processor wasn't programmed to worry about disappearing targets, it simply say a disconnected string of stationary objects, rather than one moving object.

The car's original equipment potentially could have lessened the impact; we'll never know for sure.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert! faq731-376 forum1529 Entire Forum list
 
Upvote 0 Downvote
"There were 6222 pedestrian deaths in 2018. In 6221 of those cases, the driver was a human; in one case the driver was software."

How many human drivers on the road in 2018 vs. how many self driving cars?
 
Upvote 0 Downvote
So far the Teslas have been content with killing their owners.

----------------------------------------

The Help for this program was created in Windows Help format, which depends on a feature that isn't included in this version of Windows.
 
Upvote 0 Downvote
Status
Not open for further replies.

Similar threads

IRstuff
  • Locked
  • Question
Self-Driving Cars and Taxis Are Here! 5
2
Replies
31
Views
40
Luceid
Luceid
drawoh
  • Locked
  • Question
Self Driving Uber Fatality - Thread V
Replies
11
Views
12
drawoh
drawoh
JohnRBaker
  • Locked
  • News
USS Bonhomme Richard fire, reopened thread...
2
Replies
24
Views
25
cranky108
cranky108
thebard3
  • Locked
  • Question
thread815-457402 Dive boat fire h 2
Replies
1
Views
8
drwebb
drwebb
drawoh
  • Locked
  • Question
Self Driving Uber Fatality - Thread III 4
5 6 7
Replies
121
Views
122
Spartan5
Spartan5
Back
Top