From an ARS Technica article today:
“But zooming out from the specifics of Herzberg’s crash, the more fundamental point is this: conventional car crashes killed 37,461 in the United States in 2016, which works out to 1.25 deaths per 100 million miles driven. Uber announced that it had driven 2 million miles by December 2017, and is probably up to around 3 million miles today. If you do the math, that means that Uber’s cars have killed people at roughly 25 times the rate of a typical human-driven car in the United States.”
Single data point and all that, but still not a good sign for Uber.
I imagine there are logs of the active LIDAR and IR systems that the navigation computer uses, which will put the grainy dashcam image into perspective.
I don’t assume the failure here is so basic, nor so easily addressed.
But that “general intelligence” is what gives humans the ability to do things like…
notice I’m driving in a school zone and should slow down a bit and watch for children doing unexpected things. I don’t even need a sign to tell me I’m in a school zone; I can sense that from other cues.
I’m sure the response will be: the AI will have the school zone geolocated and will know to slow down and be extra cautious inside these boundaries. But what about 10 minutes after the school day ends? Now the “school zone”, the area where children are, is expanding as they head home. I pick up on this by seeing groups of children walking with backpacks, from the presence of crossing guards, and my general knowledge of school timings.
pick up visual cues from other drivers; where they’re looking and how they’re looking tells me plenty about their immediate intention. Or a quick wiggle of the finger or wave of the hand. Or they’re looking down at their f#$%ing phone so I give them extra attention.
I doubt AI will ever have the ability to sense these cues, regardless of the improvements in cameras and face-detection. These are really hard problems. The really hard problems that are holding back “general AI”.
I’m sure we can go back and forth with examples. My point is a “general intelligence” has many abilities, many which don’t seem driving-related. But I think these abilities are essential to driving in ways that aren’t obvious, because we take them for granted. Take our uncanny ability to know when we’re looking at another person eye-to-eye, even across an intersection and thru two windshields.
I think you’re mistaking fast reaction time and better seeing (IR and LIDAR and cameras at each corner, etc) as “intelligence” (you said “smarter”); but that’s like saying I’m smarter if I increase my arm strength or wear my glasses (I need them for seeing things at distance, but not so badly that my driving is affected). These are “sensors” and “actuators”; they feed info to my brain or respond to my brain’s decisions. Better sensors and faster/more responsive actuators are important, but they’re not “intelligence”.
In the Uber example, I’d wager that a human would have picked up small flashes from the bike reflectors or the bike itself, far enough away to put the driver in “something’s on the side of the road, so I better be careful…” mode. I’ve watched the video a few times and have to admit I can’t see such small cues, but I’m assuming I’d see them IRL.
It sure looks to me that the safety driver was looking at a f$%^ing phone…
How is that relevant? It’s not like she was in a stoplight crosswalk and failed to cross during the allotted time. She was jaywalking and not even looking to see if a car was coming!
Lidar aside, darkly clothed people even on lit streets can be very hard to spot, I’ve had many close calls on our narrow local streets, where fortunately I was obeying the 25 mph speed limit.
As for effectiveness of the “safety driver” in these tests, in my experience if you’re not actually driving you have a much lower situational awareness. My wife will often gasp and brace herself if a merging or lane changing car is nearby, to my endless annoyance. She explains that she sees it much later than I do and it startles her, even though she’s just sitting there not doing anything else. She’s an awful passenger even though I’ve never had a real crash, we went through a period where I insisted she drive if we were together.
She could have stopped? Waited for all the cars to pass? She NEVER looks up, not even a turn, up until video cuts out. Reading the reports about her situation, it is tragic. Homeless with potential substance abuse. I stated earlier, I want to see what toxicology says. I love some substances, but I don’t use then cross 5 lane roads at night.
“But zooming out from the specifics of Herzberg’s crash, the more fundamental point is this: conventional car crashes killed 37,461 in the United States in 2016, which works out to 1.18 deaths per 100 million miles driven. Uber announced that it had driven 2 million miles by December 2017 and is probably up to around 3 million miles today. If you do the math, that means that Uber’s cars have killed people at roughly 25 times the rate of a typical human-driven car in the United States.”
“Indeed, it’s entirely possible to imagine a self-driving car system that always follows the letter of the law—and hence never does anything that would lead to legal finding of fault—but is nevertheless way more dangerous than the average human driver. Indeed, such a system might behave a lot like Uber’s cars do today”
It sure looks to me that the safety driver was looking at a f$%^ing phone…
Which essentially invalidates your entire argument. Your argument is essentially, “I’m a perfect driver, so we’ll say all human drivers are perfect. So, I’ll put that up against the weakest strawman of an AI I can create. Surprise, human driver wins!”
Then you end the whole diatribe with, “the human was looking at its phone.” Too funny.
Looking at the road on Google Maps, there is a slight bend in the northbound lanes that’s shrouded by trees, bushes, and an overpass just south of the location where the crash occurred (based on @beschizza’s photo and the contour of the road I can make out in the footage). Eyeballing the distance to the car in the left lane up ahead of the Uber car, I’d venture a guess that she started crossing just after that car passed by. Given that she didn’t move in front of the other car, it’s possible that she did check the road before crossing, but that the Uber vehicle (which I will reiterate was speeding in autonomous driving mode) was obscured when she looked. Yes, it’s a bad idea to jaywalk and not constantly monitor oncoming lanes, but as has been said repeatedly in this thread on behalf of the inattentive safety driver, nobody is perfect.
Incidentally, that “please don’t use this path” path looks even worse when you look at the location from overhead:
The paved path in the median closest to the collision curves directly into what looks like a desire path around the edge of a park on the right-hand side, and there’s a bus stop in the southbound lanes that looks like a fancy crosswalk - especially since it’s precisely lined up with what is in fact not a bus stop on the exact opposite side of the road, across the median which has pedestrian-friendly grooves cut through the bushes at that location.
If nobody has been killed crossing this street before, I will be gobsmacked. It’s a failure of pedestrian-friendly infrastructure at pretty much every single possible opportunity. That said, in this incident Uber didn’t even fail at Autonomous Driving 101, it failed at Remedial LIDAR For People Who Failed Autonomous Driving 101.
(Oh god, I followed the foot path around the theater to the left and it gets even worse, with a level-grade rail crossing that’s angled to imply that you can also cross the street at that point as well, even though you once again very plainly can/should not. There’s even a curb round-down to flatten the sidewalk, and no textured marking to warn visually-impaired people that they’re about to wander into traffic!)
I don’t think we have any reason to assume she wasn’t aware of the car. Cars make a ton (or two) of noise. The car had lights, she could have seen it coming. If her peripheral vision was working, even if she hadn’t looked directly at it before, she would have known where the car was.
What do you do if you’re in the street and a car is coming? Get off to the side, which she was trying to do. She may have even stepped off, gone a little way, then turned back after seeing traffic.
How much did the bike weigh? Did it contain all of or most of her worldly possessions? I read she was homeless and living at the campground – did she leave her stuff or carry it with her? How strong was she? Being homeless doesn’t come with a gym membership, it’s a rough life. Sure she gets exercise just getting around, but did she ever get the stress-free rest and relaxation needed to recover, turning exercise into strength? Or did it just grind her body down, day after day?
And speaking of rest, how long had it been since she’d gotten a good night’s sleep? Did she ever get the restful sleep that brought her awake and alert in the morning? Or did she live in a state of perpetual stress, just falling over each night when she was too exhausted to continue, more unconscious than asleep? To really let go and rest you need to feel safe and secure … did she ever feel that way, as a homeless person? We know that just a few nights without rest will leave someone so impaired they may as well be drunk. Was that her mental state?
My point is that we may be looking at a person who got in trouble crossing the road, trying their best to get out of trouble, knowing they were in trouble, making best speed to get off the road, and not making it.
Sure, but it’s the not looking at oncoming traffic at all that has some of us baffled. In your long litany of reasons she might have been impaired physically and cognitively, you studiously left out the most likely reason by far. But, after all, this is BBS.
No, I felt that was adequately covered by @Al_Estok . It was the other ones, the ones not hammered like they were White House talking points, that need to be considered too.
Thank you for the aerials, I am not a traffic engineer, but that geometric design is fucking horrendous for pedestrians. So many cues to places to get hit. I agree, the lidar failure is pretty glaring, I wonder what the point cloud of dark cotton looks like, maybe a valuable lesson here for the designers. At the cost of a woman leading a rough life, loved by her peers, and unfortunately, her biggest mark on the internet is going to be this clusterfuck.
I have been reading all the interviews on this, she really had a rough, heartbreaking story. Still think if it had been a regular dumb schmuck driving along the same path, the same outcome would have happened.
And I’m pretty sure the pedestrian received a more than sufficient punishment for their actions. That’s the difference between a reckless pedestrian and a reckless driver. A reckless pedestrian puts themselves in danger. A reckless driver puts others in danger. If you want an equal burden to fall on both parties to ensure care in the public right of way the legal burden needs to fall far heavier on the person operating a vehicle with a strong possibility of killing someone else.
The video establishes that Herzberg carelessly walked into a dark street at night
…”
No, it doesn’t establish that one bit. Herzberg was halfway across the right lane. She is walkng from the left to the right. She was already most of the way across the road.
The video also does not establish the speed of Uber’s vehicle, which looks quite fast. And there’s also the back up driver carelessly not paying attention to her responsibilities as a back up driver.
It seems very plausible that Ms. Herzberg entered the road when it as safe but a car traveling fast arrived and, as the video does establish, its Lidar and back up driver both failed.
True, there are also other sensors, but since they mentioned darkness I thought that also the color camera may be used to navigate and, maybe, in such a situation with strong lights and darkness also IR cameras may be disturbed (IDK how much infrared light is emitted by street lights).
Anyway, if the car was not able to detect a big obstacle close to the road or even inside it (the woman was moving but did not appear out of nowhere) it should have slowed down instead of driving close to the speed limit.
I agree, the car should have easily avoided the collision, or should have warned the backup driver of being unable to handle the situation, while slowing down,
It does not seem a complex situations, so i am curious to hear what has gone wrong.
I am pretty sure that there are a lot of engineers losing sleep because of this.
The good thing is, once this is fixed, is fixed, it will not repeat again (at least not in the same way).
Humans repeat their errors.
