This used to be my job. They’re not controlling the cars. They’re basically completing real-time CAPTCHAs, telling the car whether the cameras see a stop sign, a bicycle, temporary barriers, etc. If the car can’t identify an object that could possibly cross its path, it pulls over and stops until an operator can do a sanity-check on whatever the car’s confused by. They only need to be able to identify objects on the road, not know the rules of the road.
I think the interventions here are more like: “that’s a trash can someone pushed onto the road - let me help you around it” rather than: “let me drive you all the way to your destination.”
It’s usually not the genuinely hard stuff that stumps AI drivers - it’s the really stupid, obvious things it simply never encountered in its training data before.
Saw this blog post recently about waymo’s sim setup for generating synthetic data and they really do seem to be generating pretty much everything in existence. The level of generalization of the model they seem to be using is either shockingly low or they abort immediately at the earliest sign of high perplexity.
I’m guessing it’s the latter, they need to keep accidents to a minimum if they’re ever going to get broad legislation to legalise them.
Every single accident is analysed to death by the media and onlookers alike, with a large group of people wanting it to fail.
This is a prime example, we’ve known about the human intervention for a while now but period people seem surprised that those people are in another country.
Broadly speaking, an AI driver getting stumped means it’s stuck in the middle of the road - while a human driver getting stumped means plowing into a semi truck.
I’d rather be inconvenienced than killed. And from what I’ve seen, even our current AI drivers are already statistically safer than the average human driver - and they’re only going to keep getting better.
Ai drivers have run over and crushed people slowly before too though because they didn’t see the person as an “obstacle” to be avoided, or because they were on the ground, it didn’t see them
And they always will. You need to look at the big picture here, not individual cases. If we replaced every single car on US roads with one driven by AI - proven to be 10 times better a driver than a human - that would still mean 4,000 people getting killed by them each year. That, however, doesn’t mean we should go back to human drivers and 40,000 people killed annually.
current AI drivers are already statistically safer than
As long as they use level 3 autonomous cars and then cheat with remote operators instead of using real level 5 cars, such statistics remain quite meaningless.
However, they tell about the people who use them as arguments.
As the OP stated, the low velocity cases are not causing deadly accidents. And you can’t drive by wire at high speed (too much latency). So I doubt it’s affecting the stats in any meaningful way.
Honestly I much prefer they have a human as a backup than not.
This is how it generally behaves, but they are capable of taking direct control in more difficult situations. It’s only very slow maneurvers though, it’s not like they would be driving it down the street. They could move it off the road onto the shoulder though if needed.
Edit: I am trying to find the source, but having problems. It was only ever mentioned in 1 official waymo document that I’ve seen that it was technically possible. My guess is they say their remote helpers can’t / don’t do it because they truly can’t, and it’s some highly restricted type of person who can, who isn’t classified like these other employees. The whole misleading but technical true kinda speak. I’ll keep looking though because I was really surprised to see them admit it when I saw it in an official document.
In very limited circumstances such as to facilitate movement of the AV out of
a freeway lane onto an adjacent shoulder, if possible, our Event Response agents are able to remotely move the Waymo AV under strict parameters, including at a very low speed over a very short distance.
Looks like I was right as well on terminology, it’s not the remote operators that can do it, it’s the “Event Response” team that can.
As far as I know this is the only official acknowledgement it’s possible. Everywhere else they say it isn’t, and this is a footnote in that document.
And these foreign crowd workers know the local traffic rules? Maybe they even have regular drivers licenses?
This used to be my job. They’re not controlling the cars. They’re basically completing real-time CAPTCHAs, telling the car whether the cameras see a stop sign, a bicycle, temporary barriers, etc. If the car can’t identify an object that could possibly cross its path, it pulls over and stops until an operator can do a sanity-check on whatever the car’s confused by. They only need to be able to identify objects on the road, not know the rules of the road.
I think the interventions here are more like: “that’s a trash can someone pushed onto the road - let me help you around it” rather than: “let me drive you all the way to your destination.”
It’s usually not the genuinely hard stuff that stumps AI drivers - it’s the really stupid, obvious things it simply never encountered in its training data before.
Saw this blog post recently about waymo’s sim setup for generating synthetic data and they really do seem to be generating pretty much everything in existence. The level of generalization of the model they seem to be using is either shockingly low or they abort immediately at the earliest sign of high perplexity.
I’m guessing it’s the latter, they need to keep accidents to a minimum if they’re ever going to get broad legislation to legalise them.
Every single accident is analysed to death by the media and onlookers alike, with a large group of people wanting it to fail.
This is a prime example, we’ve known about the human intervention for a while now but period people seem surprised that those people are in another country.
Feels like the robot hoovers when they encounter an unexpected poo.
Ancient texts show that robot hoovers did not have a means of intervention
Hm. Interesting. But that makes them look even mode incapable than I feared.
Broadly speaking, an AI driver getting stumped means it’s stuck in the middle of the road - while a human driver getting stumped means plowing into a semi truck.
I’d rather be inconvenienced than killed. And from what I’ve seen, even our current AI drivers are already statistically safer than the average human driver - and they’re only going to keep getting better.
They’ll never be flawless though. Nothing is.
Ai drivers have run over and crushed people slowly before too though because they didn’t see the person as an “obstacle” to be avoided, or because they were on the ground, it didn’t see them
And they always will. You need to look at the big picture here, not individual cases. If we replaced every single car on US roads with one driven by AI - proven to be 10 times better a driver than a human - that would still mean 4,000 people getting killed by them each year. That, however, doesn’t mean we should go back to human drivers and 40,000 people killed annually.
By that logic…
We should really be investing in trains and buses, not cars of any type.
As long as they use level 3 autonomous cars and then cheat with remote operators instead of using real level 5 cars, such statistics remain quite meaningless.
However, they tell about the people who use them as arguments.
As the OP stated, the low velocity cases are not causing deadly accidents. And you can’t drive by wire at high speed (too much latency). So I doubt it’s affecting the stats in any meaningful way.
Honestly I much prefer they have a human as a backup than not.
Here’s a short video of someone receiving help. They explain briefly that they provide instructions to the vehicle, they don’t do the actual driving
https://youtube.com/shorts/8ltYKcswq6A
This is how it generally behaves, but they are capable of taking direct control in more difficult situations. It’s only very slow maneurvers though, it’s not like they would be driving it down the street. They could move it off the road onto the shoulder though if needed.
Edit: I am trying to find the source, but having problems. It was only ever mentioned in 1 official waymo document that I’ve seen that it was technically possible. My guess is they say their remote helpers can’t / don’t do it because they truly can’t, and it’s some highly restricted type of person who can, who isn’t classified like these other employees. The whole misleading but technical true kinda speak. I’ll keep looking though because I was really surprised to see them admit it when I saw it in an official document.
https://www.cpuc.ca.gov/-/media/cpuc-website/divisions/consumer-protection-and-enforcement-division/documents/tlab/av-programs/tcp0038152a-waymo-al-0003_a1b.pdf
Looks like I was right as well on terminology, it’s not the remote operators that can do it, it’s the “Event Response” team that can.
As far as I know this is the only official acknowledgement it’s possible. Everywhere else they say it isn’t, and this is a footnote in that document.