Uber self-driving crash: Footage shows moment before impact [video](bbc.com)
bbc.com
Uber self-driving crash: Footage shows moment before impact [video]
http://www.bbc.com/news/world-us-canada-43497364
14 comments
> We don't yet know why the driver was looking down and to the right.
Looked like they were playing on a phone to me.
Looked like they were playing on a phone to me.
I'm not sure how closely we should analyze this footage. Its relatively low-quality dashcam footage, and I highly doubt its representative of what the car "sees" and what a human would see in similar conditions.
That being said, Uber's LIDAR should've seen the woman in the road, right? Does the range not go beyond the lane ahead?
That being said, Uber's LIDAR should've seen the woman in the road, right? Does the range not go beyond the lane ahead?
Published vs. real-world LIDAR ranges are radically different. Even if the LIDAR sensor is looking 360deg around the car, the Uber software may only be looking at a small cone in front of the car.
The challenge is: what do you do when something is moving toward the vehicle's path? You could track every object in front of the car, but since people walk toward the road, then stop, you can't just apply full brakes every time something is moving toward the road.
It's really, really hard to get this right: we humans can look at people standing next to a road, and, based on movement and body language, predict whether pedestrians will walk into a road.
And since there will be cases where pedestrians don't look and walk in front of a vehicle going 40mph, autonomous technology will be limited by the physics of stopping a vehicle with full brake applied.
It's impossible to figure out what the Uber vehicle should have, or could have, done autonomously without looking at the code, and all the data.
The challenge is: what do you do when something is moving toward the vehicle's path? You could track every object in front of the car, but since people walk toward the road, then stop, you can't just apply full brakes every time something is moving toward the road.
It's really, really hard to get this right: we humans can look at people standing next to a road, and, based on movement and body language, predict whether pedestrians will walk into a road.
And since there will be cases where pedestrians don't look and walk in front of a vehicle going 40mph, autonomous technology will be limited by the physics of stopping a vehicle with full brake applied.
It's impossible to figure out what the Uber vehicle should have, or could have, done autonomously without looking at the code, and all the data.
Also challenging: looking ahead around the corner. You have to be able to predict somewhat where the vehicle is going, so you have to have an accurate map of the road, and know accurately where you are.
I can't be alone in thinking that the woman walking the bicycle didn't seem to be in a crosswalk...
Like, ok so autonomous driving isn't perfect, and yeah the backup driver maybe should've been paying attention to the road instead of whatever she was doing. Fine. But I can't be alone in thinking it looks at least partly the woman's own fault. Right? Walking your bike, in the dark, across a road without looking when you don't seem to have right of way is just not a great idea. Plenty of human drivers would've probably hit her too, albeit possibly slower while slamming on the brakes.
I don't mean to blame the victim here entirely, obviously I don't think she deserved to die, I just can't help but think that Uber isn't entirely at fault here. Maybe Uber can make a car that's statistically a safer driver than human drivers, but you can't out-engineer the human ability to invent ways to get around rules other humans put in place.
Like, ok so autonomous driving isn't perfect, and yeah the backup driver maybe should've been paying attention to the road instead of whatever she was doing. Fine. But I can't be alone in thinking it looks at least partly the woman's own fault. Right? Walking your bike, in the dark, across a road without looking when you don't seem to have right of way is just not a great idea. Plenty of human drivers would've probably hit her too, albeit possibly slower while slamming on the brakes.
I don't mean to blame the victim here entirely, obviously I don't think she deserved to die, I just can't help but think that Uber isn't entirely at fault here. Maybe Uber can make a car that's statistically a safer driver than human drivers, but you can't out-engineer the human ability to invent ways to get around rules other humans put in place.
The problem is that video is deceptively dark. Video cameras don't pick up nearly the same amount of light as humans eyes do and there are even photos circulating around showing it significantly brighter in this area (to the point where a human could see the entire area really, really easily).
Also some areas give pedestrians the right away regardless of where they are. Granted there are plenty of exclusions to that too and this may fall into that but regardless of who is at fault this looks avoidable if a human was driving, IMO.
Regardless it's pretty insane that their software didn't pick up this person at all. That's terrifying, IMO. That shows me it's not ready for a driver to be using it and not paying attention like this driver (they kept playing on their phone, looking down and not paying attention). People, deer, lots of crazy things are going to jump out into the road and you need to be able to handle that scenario. This person didn't even jump out, they were already in the road and were likely very visible.
Also some areas give pedestrians the right away regardless of where they are. Granted there are plenty of exclusions to that too and this may fall into that but regardless of who is at fault this looks avoidable if a human was driving, IMO.
Regardless it's pretty insane that their software didn't pick up this person at all. That's terrifying, IMO. That shows me it's not ready for a driver to be using it and not paying attention like this driver (they kept playing on their phone, looking down and not paying attention). People, deer, lots of crazy things are going to jump out into the road and you need to be able to handle that scenario. This person didn't even jump out, they were already in the road and were likely very visible.
Uber is a well-funded corporation with a long history of disregard for the law which in order to beat competitors has chosen to conduct real world road tests of many instances of an "autonomous driving" system apparently inferior to heavily intoxicated humans at reacting to unexpected pedestrians in certain conditions.
The woman was a homeless person who is now dead.
The above may influence people's emphasis on whose actions warrant the most scrutiny.
The woman was a homeless person who is now dead.
The above may influence people's emphasis on whose actions warrant the most scrutiny.
I'm less interested in blame than I am in the engineering whys. Shouldn't LIDAR have seen her, even if I wouldn't blame a human for hitting her in the same situation?
Sure, but this particular problem seems solvable. A self-driving car can have much better sensors than a human. I.e. maybe we should expect self-driving cars to see in total darkness. That's something we, the public, could decide is a minimum requirement.
I would argue these are not self driving cars. I would expect a self driving car to have multiple compute nodes in the trunk that use something more advanced than a gaming engine, a voting system and numerous sensors. LIDAR, infrared, thermal, sonar, other.
An autonomous car that expected to operate in locations and conditions that humans often fail in should be advanced enough to know everything around it and then some.
An autonomous car that expected to operate in locations and conditions that humans often fail in should be advanced enough to know everything around it and then some.
That's a software quality issue, not one of having the right hardware. We went to the moon using pretty weak hardware and clever software.
I think it is both. Human lives are at risk, so I expect solid battle hardened software and hardware.
What this raises for me is -- shouldn't self-driving car companies be using ex-test pilots during autonomous driving? Of course, I'm assuming that the driver wasn't one.
That's the personnel issue. On a technical front, what was she looking at? Her own phone? The company-designed UI for the self-driving car? Does Uber use LIDAR, if so, did the victim show up on it -- could there be an obstacle alarm? What about IR cameras? This was obviously a very difficult visibility situation. If we really want self-driving cars to be better than humans, can we get away with just visible-light cameras, or do we need better sensors?
We are all far to blase about the risks of driving, and driving distracted, from constant familiarity. It seems that pilots, and especially test pilots might have the experience and discipline to stay alert during long autonomous drives.
One might argue (certainly moving forward) that using regular drivers vs. test-pilots is negligent.