Slashdot Mirror
Police Release First Video From Inside the Uber Self-Driving Car That Killed a Pedestrian (recode.net)
An anonymous reader quotes a report from Recode: Three days after an Uber self-driving vehicle fatally crashed into a pedestrian in Tempe, Ariz., police have released video footage of what the vehicle saw with its cameras moments before running the woman over, and what happened inside the vehicle, where an operator was at the wheel. The video footage does not conclusively show who is at fault. However, it seems to confirm initial reports from the Tempe police that Herzberg appeared suddenly. It also showed the vehicle operator behind the wheel intermittently looking down while the car was driving itself.
... however the test driver did not really pay attention.
Being test driver is obviously a fucked up job. 99% is killing time and 1% is killing time.
In Germany there is not one test driver but 3 ... one who would react if something goes wrong and 2 to write protocols about notable stuff.
In this case it is notable that the lights are configured incorrect. They barely shine 15 yards ahead, that is definitely wrong, and a driver or the automatic driving system should adjust speed to about 1/3rd of what it was driving.
Cost free eBook I read (by iBook/Kobo/Amazon/ObookO/Gutenberg etc.): "The Green Odyssey" by Philip Jose Farmer.
The pedestrian didn't seem to notice the car. The car appears to be a Ford Fusion (probably hybrid). If it was in charge-sustaining mode, the car might have been very difficult to hear.
The pedestrian was wearing a yellow hat and pushing a pink bicycle. Her shirt was dark though. She was visible to the camera for only about .77 seconds prior to impact. A human being would take 0.5-2 seconds to react to the object in the road once it became visible. Depending on the human, the pedestrian might have been visible for a couple of seconds longer than we see her in the footage, but the safety driver appeared to be distracted.
The reaction time of the autonomous car should be milliseconds. Assuming that the dashed lane markers are fairly evenly spaced, the car doesn't appear to have decelerated at all from my perspective. According to the police, the car was traveling 38 MPH, or roughly 61 km/h. On dry pavement with decent tires, the stopping distance in meters without accounting for any reaction time should be about (s^2)/(250*.8) with s = speed in km/hr... so, about 18 meters, or to be generous, 60 feet.
See https://korkortonline.se/en/th... .
Judging from the aerial layer on Google maps, the distance between the beginning of a lane marker and the beginning of a subsequent lane marker is 30 feet or so. From this, I think the first time you see the victim in the video she's about 43 feet away (.77 seconds at 38 MPH).
Here's the thing though... the LIDAR should have seen this in time to at least swerve to avoid. The LIDAR should also have seen the victim before the victim was visible in the headlights. In my state, the driver has the responsibility to swerve to avoid even if there isn't enough time to stop. It's obvious that there was nobody in the left lane (even in the blind spot, which isn't blind with LIDAR).
This really seems like an example of where an autonomous car could have saved a life that would have been lost due to a human driver's natural limitations, but it failed to do so. The car should have been able to see hundreds of feet, and the car should have had practically zero reaction time. Just as you would be lenient in judging and older driver for longer reaction times, I think we should hold the autonomous car to a higher standard.
This thing was a test vehicle. The debug-level logging of the incident should be made public so that if there was a bug that killed this woman, the truth will be known.
We don't know that it didn't break
We do know that it didn't brake. There was no change in orientation of the camera that would have resulted from even slight braking. Braking force results in the front of a car dropping and the rear rising. This happens in all cars, even those with active suspension (although with A.S., the level is restored pretty quickly and the dip isn't nearly as noticeable). This camera did not change orientation at all; ergo the vehicle did not apply the brakes.
I wish I had a good sig, but all the good ones are copyrighted
Level 3 shouldn't exist; it's too dangerous. Even Level 4 probably shouldn't exist, in that it can leave people stranded when things go wrong (but it's supposed to always avoid accidents / unsafe situations and pull over safely when it can't handle them). The levels should be 1, 2, 5.
I'm sorry to have to bring reality into your delusion, but that is not how engineering works. You can't just go from level 2 to level 5 without many millions of miles of real world experience.
The best example of how this process works is the airline industry. In the airline industry, all kinds of new auto piloting features have been added over the years. They were mostly good, but flaws have been found that have caused crashes. Without those crashes, the flaws would never have been found because they are so subtle. That is why the NTSB investigates crashes, and very rarely are criminal charges brought. The reason is simple. If you start bringing criminal charges, then people stop co-operating, and the system that we have now that results in constantly improving safety doesn't work. There isn't anyone that will argue anymore that what we have now for airline safety is far better than it would have been if we stayed solely with 1950's technology in the cockpit, and you can't get where we are now any other way than the way we did: Billions of flight miles of testing with live passengers.
The NTSB seems to be taking the same approach with self driving cars, and I applaud their approach. The NTSB doesn't (and nothing else can) guarantee that there wont be accidents like this. The assurance you do get is that every year, the technology will improve and the danger will be less than before. The systems will always get safer because of the way the NTSB works.
It was long ago determined by our good friends at NASA and the NTSB that the single most dangerous piece of equipment in power air flight is the pilot. This part cannot be significantly improved over its current state, so it has been systematically replaced in the cockpit over the last 50 years. Today, the Pilots are largely just there in case something goes truly wrong, but we are quickly approaching the time when the presence of a pilot will not significantly improve the odds of surviving any given flight.
Applying that same principle to over the road travel is a no-brainer, and given that there are some tens of thousands of traffic deaths in any given year, self driving cars would need to kill on the order of thousands people for every year sooner that they bring about an end to humans in the driver seat.
So far self driving cars have killed less than 10 people that I know of, and if it takes them 10 more years to get it right without killing anyone, then it would be worth the trade if they had to kill 10,000 people to get it right in 9 years instead of 10.
I wish I had a good sig, but all the good ones are copyrighted
No human driver could have seen that woman in time to stop
Had the headlamps been aimed properly, they could have. In the video, when the car is traveling at 38 mph (56 feet/second), it takes about 1.5 seconds between the time the pedestrian came into view and when the collision occurred. That means that the headlamps are only lighting up an area 84 feet in front of the vehicle. If the vehicle's headlamps are about 2 feet off the ground, then when they're properly aiimed, they should be lighting up an area about 285 feet in front of the car (VOL headlamps where the left half of the horizontal beam cutoff is 2.1 inches below headlamp height at a distance of 25 feet from the front of the vehicle).
If the pedestrian was visible at 285 feet, it would have taken 5 seconds from the time the pedestrian came into view till when a collision could occur. That would have given the driver a second to react and 4 more seconds to slow down and/or change direction to avoid a collision.
The dynamic range of human eyes is much greater than a camera. It was not pitch black outside to a human. I have lived in Tempe and the ambient light of the city would be enough for at least minimal night vision to apply. This is the reason why you drive in a darkened vehicle without your dome lights at night, for your night vision to be effective. Texting on your phone in a part of town where there are a lot of people roaming the streets (such as south Scottsdale Rd) is simply a negligent thing to do.
The typical response/reaction time is 0.3s
That's only true when you're anticipating an event, with max attention, and have already mentally prepared your response. Response in normal driving circumstances is much slower:
http://www.croberts.com/respon...
McGee et. al. (1) reported that perception time is the sum of eye movement time, fixation on the hazard time delay, recognition time delay and muscle response delay time. They found that for the 85th percentile of drivers, eye movement delay was 0.09 seconds, fixation delay time was 0.20 seconds, recognition delay time was 0.50 seconds, decision time 0.85 seconds, muscle response delay was 0.31 seconds and brake reaction time was 1.24 seconds. The sum total of these times, the response time, was 3.19 seconds. The 85th percentile is often chosen as the upper bound for design analyses