Agreed. At least not with the sensors that come equipped on Tesla vehicles. Companies like Waymo are much closer, however.
Though I would also argue the presence of adversarial attacks doesn't necessarily mean self-driving cars are not close to being solved. There will always be some actions that bad actors can take to cause harm to others - that doesn't mean something is broken for regular use. You might as well put a bunch of nails in the road on the highway, for example. Luckily, most people aren't bad actors, and we have laws to prevent these things.
If you listened to the hype, yes we are not close; but that'd be your fault for giving in to the early hype, written by tech/science journalists without domain expertise or insider information.
If you look at the reality, certain companies have advanced a TREMENDOUS amount in the last few years. It was less than 2 decades ago that DARPA introduced its self-driving vehicle challenge and none of them could even make it a couple of miles.
If you asked me 20 years ago, would there be a driverless taxi service in Phoenix in 2020, I would have said absolutely no way, the y2k bug scare, tech so hard, we'll never get there.
No shit. What about when you erase all the lines on the road? What about when you remove the road? What's next, "Luxury Tesla defeated by a cheap piece of duct tape over the main sensors"? "We removed a wheel from a Tesla and it crashed into a wall!". These tests are so contrived, honestly.
You've never seen a car that's accidentally driven into wet concrete/pavement? If a construction worker places the pylons a little too far apart, it looks like you might be able to use the lane!
A case in point about human perception: while you read jbob2000’s comment that way, I read that same comment as nothing more than a counter to the idea that humans were somehow invulnerable to errors.
This seems a bit strange, given the main argument in favour of replacing human drivers with AI is that it should be possible to make one that works as well as the best human driver, only without getting tired, high, or distracted, and with reaction times measured in microseconds instead of hundreds of milliseconds.
We’re not there yet. I don’t know how far we are from that future.
And there are things you could do to roads that would get human drivers killed but not affect machine drivers.
The world doesn't descend into chaos just because chaos is possible. It descends into chaos when the people who cause chaos aren't caught and brought to justice.
Three white spots that look maybe a few inches square each are hardly 'contrived'. Even outside of adversarial attacks, the same situation could happen from children playing with chalk or someone crossing the road with a leaky paint can.
> What about when you erase all the lines on the road? What about when you remove the road?
Both of those conditions can be handled by human drivers...a "self-driving" car had better at least handle them safely (e.g., alert the driver and then disengage autopilot), even if it can't actually navigate under the circumstances.
> What's next, "Luxury Tesla defeated by a cheap piece of duct tape over the main sensors"?
Someone will do that. I'd hope that the car would alert the driver as soon as it started that there was no usable sensor data.
> These tests are so contrived, honestly.
Of course, so we should just assume that these things work perfectly and that bad actors don't exist. Sounds safe to me.
> What about when you erase all the lines on the road? What about when you remove the road?
Both of those conditions can be handled by human drivers...a "self-driving" car had better at least handle them safely (e.g., alert the driver and then disengage autopilot), even if it can't actually navigate under the circumstances.
Not always, at least not well -- I had a car drift into the side of my car in the rain when the lane markers were nearly invisible and the seams in the road made it seem like his lane was moving into my lane.
The article mentions the 'blurred line' test - but the more interesting one was the smaller interference dots - which is a more practical attack. You can see them in the video here:
While I'm hugely skeptical of the current state of self-driving cars, you could probably get human drivers to make the same mistake if you were to repaint the lines. However, humans will also notice the oncoming cars (if there are any) and avoid getting in a head-on collision.
The thing missing from this test is that critical practical piece: if there was an oncoming car, will the Tesla do something to avoid the collision? I would assume that not getting in a head-on crash is higher priority than staying in the lane markings.
Without oncoming traffic, all this is testing is what the Tesla considers valid line markings. I'm sure there's room for improvement here (such as checking where the other lane is, raising the requirement for how well-defined the lines have to be, etc), but those are also going to involve trade-offs where there are legitimate situations that will stop working.
I think you could just as easily title this video "Tesla auto-pilot follows road markings even if they're really bad".
Edit: The best shot I could get from the video [1] makes me even more upset at this test: these look like the temporary markings often used during construction, just before they come and paint the normal lines using the big line-painting truck. There's not even regular lane lines after this. I wouldn't even be surprised if this is something Tesla specifically trained the software to handle.
> The thing missing from this test is that critical practical piece: if there was an oncoming car, will the Tesla do something to avoid the collision? I would assume that not getting in a head-on crash is higher priority than staying in the lane markings.
Considering that a Tesla is happy to ram you into the back of a stopped vehicle, at highway speeds, I wouldn't have that hope.
Autopilot does two things - follows the vehicle in front of you, and stays between the lane lines.
It does more than that. You don't even have to look to Tesla to find cruise control tech that is capable of adjusting speed based on traffic conditions. e.g., Nissan uses tech that can adjust the speed of your vehicle relative to the distance between you and the car in front of you. You can even choose between three different follow distances for the car in front of you. It can go from 75mph to a full stop without intervention, and then back up to 75mph.
I get that there's skepticism related to self-driving vehicles, but let's be genuine here and recognize that this technology is young and is still in a state of development, and not misrepresent what it can and cannot do.
23 comments
[ 5.0 ms ] story [ 47.8 ms ] threadThough I would also argue the presence of adversarial attacks doesn't necessarily mean self-driving cars are not close to being solved. There will always be some actions that bad actors can take to cause harm to others - that doesn't mean something is broken for regular use. You might as well put a bunch of nails in the road on the highway, for example. Luckily, most people aren't bad actors, and we have laws to prevent these things.
If you look at the reality, certain companies have advanced a TREMENDOUS amount in the last few years. It was less than 2 decades ago that DARPA introduced its self-driving vehicle challenge and none of them could even make it a couple of miles.
If you asked me 20 years ago, would there be a driverless taxi service in Phoenix in 2020, I would have said absolutely no way, the y2k bug scare, tech so hard, we'll never get there.
No human would be tricked into driving into oncoming traffic by the three small stickers in the road used in this attack.
A favorite of gun enthusiasts and gun lobbyists.
This seems a bit strange, given the main argument in favour of replacing human drivers with AI is that it should be possible to make one that works as well as the best human driver, only without getting tired, high, or distracted, and with reaction times measured in microseconds instead of hundreds of milliseconds.
We’re not there yet. I don’t know how far we are from that future.
The world doesn't descend into chaos just because chaos is possible. It descends into chaos when the people who cause chaos aren't caught and brought to justice.
Both of those conditions can be handled by human drivers...a "self-driving" car had better at least handle them safely (e.g., alert the driver and then disengage autopilot), even if it can't actually navigate under the circumstances.
> What's next, "Luxury Tesla defeated by a cheap piece of duct tape over the main sensors"?
Someone will do that. I'd hope that the car would alert the driver as soon as it started that there was no usable sensor data.
> These tests are so contrived, honestly.
Of course, so we should just assume that these things work perfectly and that bad actors don't exist. Sounds safe to me.
Both of those conditions can be handled by human drivers...a "self-driving" car had better at least handle them safely (e.g., alert the driver and then disengage autopilot), even if it can't actually navigate under the circumstances.
Not always, at least not well -- I had a car drift into the side of my car in the rain when the lane markers were nearly invisible and the seams in the road made it seem like his lane was moving into my lane.
https://www.youtube.com/watch?v=6QSsKy0I9LE&feature=youtu.be
The thing missing from this test is that critical practical piece: if there was an oncoming car, will the Tesla do something to avoid the collision? I would assume that not getting in a head-on crash is higher priority than staying in the lane markings.
Without oncoming traffic, all this is testing is what the Tesla considers valid line markings. I'm sure there's room for improvement here (such as checking where the other lane is, raising the requirement for how well-defined the lines have to be, etc), but those are also going to involve trade-offs where there are legitimate situations that will stop working.
I think you could just as easily title this video "Tesla auto-pilot follows road markings even if they're really bad".
Edit: The best shot I could get from the video [1] makes me even more upset at this test: these look like the temporary markings often used during construction, just before they come and paint the normal lines using the big line-painting truck. There's not even regular lane lines after this. I wouldn't even be surprised if this is something Tesla specifically trained the software to handle.
[1] https://i.imgur.com/aLbhnzQ.jpg
Considering that a Tesla is happy to ram you into the back of a stopped vehicle, at highway speeds, I wouldn't have that hope.
Autopilot does two things - follows the vehicle in front of you, and stays between the lane lines.
I get that there's skepticism related to self-driving vehicles, but let's be genuine here and recognize that this technology is young and is still in a state of development, and not misrepresent what it can and cannot do.
I don't believe that they're likely to pick up something driving straight at them, as it's not something the neural net has been trained for.
My Model X often alerts me to parked cars, thinking I might drive into them :P