Given the incredibly small number of unmanned, autonomous vehicles on the road, and the fact that they stick to fairly uncomplicated urban and suburban terrains, the comparison would be meaningless. The numbers for autonomous vehicles would border on statistical insignificance.
You're being very tendentious rather than objective, rhetorical rather than dialectical.
Statistically speaking, I agree that comparison isn't helpful. At the same time we have to take into account it's still learning phase for self-driving. I'd be interested to learn from the history how well automobile was received by public a century ago, and whether the accident rate spiked early on.
While I agree with your inference that autonomous vehicles are still safer than human drivers, this incident throws doubt on a lot of my previous assumptions.
A human driver, in the scenario as reported, can be expected to take a certain amount of time to react to an unexpected event, such as a pedestrian walking out in front of it. An autonomous driver, IMO, doesn't have that luxury. That vehicle should be aware of every other object it's sharing a length of road with (sidewalks and medians included). It should be tracking every other object at all times, anticipating a change of course. It should know the proximity of other vehicles, and be able to calculate whether it has space to swerve if necessary, and whether the following vehicle has enough stopping distance in an emergency. It should know EXACTLY what it will do in response to any of the potential hazards it can detect. It should absolutely not be exceeding the speed limit under any circumstances.
I suspect this incident will highlight one of the areas where investigators are not currently up to speed with technology. You can't look at an incident like this as if it was a human driver using multiple dashcams, because "It was dark and she came from nowhere" doesn't cut it. For an autonomous vehicle accident, the video isn't enough - investigators need to see and understand all the sensor readings taken by the vehicle in the lead up to the accident. Did the car detect the pedestrian? If not, why not? Why didn't it notice the change of course? How long before impact did it identify that a collision was imminent? Why didn't it brake?
Unlike a human driver, there's no human factor to consider. Every decision that vehicle made was based on rules. So "The Uber was not at fault" isn't enough - every day, thousands of people are involved in accidents where they're not at fault. That doesn't mean they couldn't have avoided it, and part of being a good driver is allowing for other people doing dumb things. The investigation needs to highlight exactly what decisions that vehicle made which resulted in this tragedy, and what can be changed to make sure it isn't repeated.
I think accident investigators could certainly take some tips from them when it comes to autonomous vehicles. They need to understand exactly what data is collected by these vehicles, and how they can use that data as part of their investigation. There should also be a full root cause investigation into every single accident involving an autonomous vehicle. These protocols really need to be in place now, with training and funding of accident investigators written in as part of the operating license for each city, because if we wait a decade, we'll get a lot more resistance from the operators.
Figuring out how to investigate accidents involving autonomous vehicles (and this certainly won't be the last) is going to be important for law enforcement.
Figuring out? What's there to figure out? it's a computer that captures live footage of its surrounding events. I'm pretty sure it will store at least the last 3 minutes before an unknown event or collision. Similar to a black box in an airplane. It's a matter of requiring them legally to provide such footage.
Have you interacted with the police recently? I had an encounter with the police here in Germany a couple years back about some property of mine being sold on eBay. Handling anything digital was certainly not their comfort zone, neither with regard to training nor equipment.
I have a hard time imagining how the normal police men/women on the street will be able to handle the information overload the black box of an autonomous car represents for them.
I'd assume law enforcement would need to create specialized units for this.
Well, there's the chain of custody of the live footage to worry about and ensuring it hasn't been tampered with. And cops have to be trained in how to look at it, there are probably special tools. In the long run we're going to have an FAA equivalent for autonomous cars with standardized black boxes and formal accident investigation committees and so forth so hopefully every accident where an autonomous car is at fault can be the last of its type. But currently we're still figuring out how all this works which means both that we don't want all that structure yet and also that relevant authorities have to learn by doing.
Likely too big to fail? Human driver would try at least an emergency braking. Even if it’s too late.
I can remember Continental testing emergency braking assistant in 2011. There was braking and accelerating noise near their office every day for weeks. They tested with giant rectangle and dummy pedestrian and apparently it worked well. Did Uber turned this feature off?
It’s natural to hit the brakes. Sometimes that can kill you. Like driving on the highway during the rain and trying avoid fox or dog.
This failure mode (pedestrian jumping in front of the car) is quite often. I drive 10k miles early and encounter this at least once yearly. Last time it was an elder gentleman who felt off his bicycle on the street from pavement. Time before it was a boy with scooter. It happens often and the scenario is easy to test in closed area.
I don't buy the "jumping in front of the car" part. Pedestrians pushing bicycles laden with bags are not known for being especially nimble. Crossing without looking, sure, but assuming the pedestrian was crossing the road, she'd be standing about where the bicycle seat is. That gives half a bike's worth of space before she's actually in front of the car - let's say a quarter of a second if she's going particularly quickly.
I'll forgive a human driver for not noticing a pedestrian and reacting in a quarter of a second. But all the marketing videos have trained us to hold an autonomous car to a higher standard. Certainly a lack of lighting isn't an excuse. The car should have known she was there, identified her as a potential hazard, and been prepared to take appropriate action should her course change. Speeding, along with the fact that the car made no attempt to stop before the collision, is a "back to the drawing board" level of failure.
Suicidal people wouldn't typically try to take their bicycle and their shopping with them. But honestly, the reason she was there isn't really important. The reason the Uber couldn't avoid her is a far more important question in terms of where we go from here. It should have detected her as a potential hazard. It should have been prepared to stop. It shouldn't have been exceeding the speed limit.
A good driver will look back at what they did after an accident to see if they could have done anything to prevent it, even if they weren't at fault. Uber need to do exactly the same thing. Fortunately, unlike a human driver, they have significantly more data than an adrenaline-fuelled mind and some grainy dashcam footage to work with.
The recurring theme here is "autonomous car fails to read human intentions/behaviour". This has been a question in the past, with people saying autonomous cars drive unnaturally through intersections because they cannot figure out the intentions of other human drivers. The issue in this particular case is that the car can react faster once the pedestrian is detected as an obstacle, but it will take longer than a human driver to figure out that the pedestrian is going to become an obstacle. So overall reaction time might not be quicker.
Can a NN learn enough human behaviour to predict a "pedestrians about to walk in front of car" situation without somehow possessing a very high level of reasoning? I think yes. But there will always be some edge cases. One other interesting aspect is how different human cultures will probably mean different training is needed for each particular culture.
THIS My friend Gautam Sinha was driving to lunch through an uncontrolled intersection and neatly judged two cars hesitating at crossroads. I complimented him - "Well done!". He said "That was nothing. I've done two cars, a motorcycle and a dog. Two dogs!" In India, you hesitate, you're going nowhere. How will autonomous cars ever survive there?
23 comments
[ 4.3 ms ] story [ 57.9 ms ] threadYou're being very tendentious rather than objective, rhetorical rather than dialectical.
A human driver, in the scenario as reported, can be expected to take a certain amount of time to react to an unexpected event, such as a pedestrian walking out in front of it. An autonomous driver, IMO, doesn't have that luxury. That vehicle should be aware of every other object it's sharing a length of road with (sidewalks and medians included). It should be tracking every other object at all times, anticipating a change of course. It should know the proximity of other vehicles, and be able to calculate whether it has space to swerve if necessary, and whether the following vehicle has enough stopping distance in an emergency. It should know EXACTLY what it will do in response to any of the potential hazards it can detect. It should absolutely not be exceeding the speed limit under any circumstances.
I suspect this incident will highlight one of the areas where investigators are not currently up to speed with technology. You can't look at an incident like this as if it was a human driver using multiple dashcams, because "It was dark and she came from nowhere" doesn't cut it. For an autonomous vehicle accident, the video isn't enough - investigators need to see and understand all the sensor readings taken by the vehicle in the lead up to the accident. Did the car detect the pedestrian? If not, why not? Why didn't it notice the change of course? How long before impact did it identify that a collision was imminent? Why didn't it brake?
Unlike a human driver, there's no human factor to consider. Every decision that vehicle made was based on rules. So "The Uber was not at fault" isn't enough - every day, thousands of people are involved in accidents where they're not at fault. That doesn't mean they couldn't have avoided it, and part of being a good driver is allowing for other people doing dumb things. The investigation needs to highlight exactly what decisions that vehicle made which resulted in this tragedy, and what can be changed to make sure it isn't repeated.
I have a hard time imagining how the normal police men/women on the street will be able to handle the information overload the black box of an autonomous car represents for them.
I'd assume law enforcement would need to create specialized units for this.
I can remember Continental testing emergency braking assistant in 2011. There was braking and accelerating noise near their office every day for weeks. They tested with giant rectangle and dummy pedestrian and apparently it worked well. Did Uber turned this feature off?
Edit: typos
Seem like your making a pretty big assumption that "Human driver would try at least an emergency braking. Even if it’s too late."
This failure mode (pedestrian jumping in front of the car) is quite often. I drive 10k miles early and encounter this at least once yearly. Last time it was an elder gentleman who felt off his bicycle on the street from pavement. Time before it was a boy with scooter. It happens often and the scenario is easy to test in closed area.
I'll forgive a human driver for not noticing a pedestrian and reacting in a quarter of a second. But all the marketing videos have trained us to hold an autonomous car to a higher standard. Certainly a lack of lighting isn't an excuse. The car should have known she was there, identified her as a potential hazard, and been prepared to take appropriate action should her course change. Speeding, along with the fact that the car made no attempt to stop before the collision, is a "back to the drawing board" level of failure.
This appears to me to be a suicide.
A good driver will look back at what they did after an accident to see if they could have done anything to prevent it, even if they weren't at fault. Uber need to do exactly the same thing. Fortunately, unlike a human driver, they have significantly more data than an adrenaline-fuelled mind and some grainy dashcam footage to work with.
Can a NN learn enough human behaviour to predict a "pedestrians about to walk in front of car" situation without somehow possessing a very high level of reasoning? I think yes. But there will always be some edge cases. One other interesting aspect is how different human cultures will probably mean different training is needed for each particular culture.