That's a tough problem - distinguishing wet pavement from deep water.
Humans make that mistake frequently.
Autonomous vehicles should probably be equipped with a water sensor. (We did that in our DARPA Grand Challenge vehicle back in 2005). Then they can enter water very cautiously and see if it's too deep. This may make them too cautious about shallow puddles on roads, though.
It doesn't take much of a rainstorm to see localized flooding. Some debris over the storm drain is enough to flood a street. Hard to anticipate that happening.
It’s a particularly hard problem in Texas. We get torrential rains and the landscape is relatively flat. Couple that with shallow soil over lots of limestone and it means flooding is really common. We also have roads that have a “low water crossing,” where a road crosses a creekbed that is normally dry but which will flood. There are often water depth signs there (basically a vertical ruler with feet marks so you can see where the water is up to). We lose people to this scenario (driving into flood waters) every year. It’s particularly problematic when it’s dark and you miss a warning sign. Before you know it, you’re in deep water and the flow can sweep the whole car downstream until it gets pinned against a tree, possibly with water forcing its way into the car.
Such detailed database of fine grained road geometry gets stale very quickly, due to road maintenance and road construction. In US highway lanes are shifted sideways frequently.
Waymo does have a database of every today they drive, but for this they don't need one.
If the car comes to a road covered with water, and that road is in the database, and the water level appears low compared to the historical level of the road in the DB, then the car could cross. if the road is not in the DB, then a different decision might be made.
This is similar to humans: you might make different decisions depending on whether you know the road well or not.
Isn’t that the Waymo data model, though? They extensively pre-drive every new market, building dense volumetric maps of the entire service area before they begin service, so they essentially do have that database of every road (that they drive on).
Granted, I am not sure exactly how Waymo operates, but I thought that the extensive testing was mostly for legal reasons+just handling edge cases.
I am saying this, because I noticed that they typically start with a low-tier restrictive permit to operate (with a rather small number of cars in the fleet). Then they run it for a year or two, iron out edge cases particular to a given city (e.g., climate particularities, crazy spaghetti junctions in ATL, etc.), and log a lot of data. Then they take that data, go to the city/state, say "we have all this data that demonstrates we were very above the board while running the test pilot program, we are safe, and now we want to expand out of a very limited test pilot program."
And then it either goes well (Bay Area, LA, etc.) or goes off the rails for other reasons (often failing earlier for entirely unexpected reasons, like the pushback against it from taxi driver unions in NYC).
My point being, I could be entirely wrong, but I don't think that they literally map every single inch of the road before being allowed to operate. I just don't see it as being possible in any large populated city, given how often things change there. Just in 3 years living at one apartment in Seattle, I had a road directly adjacent to me changed from 2-way to 1-way, and then had 3-4 lanes that were basically highway entrances/exits (a block away from me) created and the whole area being rerouted entirely.
They actually do significant mapping. Where it operates currently it is not unusual to see this. It will be a waymo with a human driver operating someplace not currently in the waymo zone and clearly not en route to any maintenance facility either. Stuff like windy canyon roads with no thru access anywhere that are currently gated away, you might see a waymo with a human today.
Waymo is not the only company making lidar maps right now either. I've seen UPS deliver trucks with retrofitted lidar scanners on the roof now. I've even seen this on a police car already, looked like a black rooftop industrial ventilator on a 2ft mast installed directly on the crown victoria roof.
Tesla is less "HD", they have standard maps like we all think of, and a lane level "see-ahead" system where they basically just grab a satellite image tile, and align it with what the car sees for "FSD".
By a water sensor do you mean a sensor to detect the water level relative to the chassis? It seems like a very inexpensive downward-facing ultrasound sensor could work.
Is ultrasound less expensive than a moisture sensor?
The problem with both is they effectively require the vehicle to be in the water already. They need something that can tell depth before the vehicle has to slow down.
I've never made that mistake; I'm not aware of anyone I know doing it. I very rarely see it myself, except on news footage. Of course it happens some time somewhere but that says nothing about frequency.
> That's a tough problem
Not really. Don't drive where you don't know it's safe. Definitely don't drive into moving water - puddles only, and only if not too deep: I can usually figure it out based on the rest of the road - unless it's a sinkhole, the geometry is somewhat consistent - and especially by looking at objects in the water such as other cars driving through it. Sorry your friend isn't competent to figure it out.
People here are always quick to defend the autonomous cars, like a close friend. How often will we fall in love with a technology or company? It always distorts the truth.
Article's current (possibly original), less ambiguous title: "Waymo recalls 3,800 robotaxis after glitch allowed some vehicles to ‘drive into standing water’"
IOW 3,800 Waymo vehicles aren't currently sat spinning their wheels in water.
It's an interesting case of whether it's possible to infer the condition of wading and avoid having to install a sensor specific to a one in a million trips circumstance.
The inference would come from standing water slowing down the vehicle and likely require steering correction, in combination with some machine vision for identifying standing water.
Then there's the advantage of being Google and having hundreds of thousands of people in the same area using Google maps and navigation. Accelerometers in phones can detect crashes pretty reliably. There's a good chance they can reliably detect deceleration from standing water and report the location of the hazard.
How about a Mastodon, Lamb of God take with Floods of Triton:
Heap data upon this modern age
All human drivers now phased away
A lidar's glow, the soft wheel's echo
Autonomous force of code remains
We are last of the before rides
Now hear the robot cars rise
Hum into eternity
Remember this, all roadways lead to the fleet
Since recall on cars no longer means doing anything to the car's physical location I think the regulator NHSTA should update this term
It just creates alarmist headlines for what's really an over the air update, although "recall" is still currently a regulatory accurate term in the vehicle space
Cars, especially EVs, have many similarities to being phones. Imagine if a routine software update from Apple was called a "recall", that functionally describes what's happening here
NHTSA should at least distinguish between "omg we have to get these cars off the road and bring them to the shop immediately!" versus "over the air software update"
Kind of unrelated. That issue was due to a misguided effort to be cautious by having vehicles requesting human-review when they didn't really need it. Waymo fixed the issue by allowing the vehicles to operate in their normal, independent, mode.
This is one of the reasons why I'm suspicious of camera-only systems, here in Finland. Half the year there's a lot of snow and ice around. Which I imagine means most of the view is "white" and "shiny". Coupled with the dark winters it's gotta be a nightmare to deal with.
Could you use a different spectrum of EM radiation to detect water? There are parts of the microwave band that attenuate the signal by absorption and I wonder if you could use that. The only clue a human driver has in that situation is in the visible spectrum. The lines of the road disappear from view, which can be challenging to see at night.
They are rolling these out in New Orleans soon. Standing water is everywhere, and sometimes you have big hidden potholes. You just need to know the roads. Should be fun.
Maybe you drive into flood waters, but I don't. That's not a difficult skill to pull off.
We're still in the early days of self driving cars, and as much simulation and miles as they have, they're still constantly getting exposed to real world conditions that are new to them. The world is dynamic, so this will always remain true.
It remains to be seen where we'll converge on capability, incident rate, and acceptance.
That you need world models to sensibly deploy "thinking" machines in the real world. Else they do stupid shit like drive straight into water. You can bruteforce some semblance of thinking by training on literally all knowledge that can be digitized but even that is proving to not be quite enough.
They suspended service areas they deem high risk until the software update can be applied. So while, yes, it's just a software update, it's a recall in the sense that they've temporarily pulled all the cars off the road in certain areas
Yes, this is a common terminology issue. "Recall" is legally defined in terms of the kind of problems that require one, not the solution to those problems, because the relevant regulations were written when there was no way to fix consumer products other than physically delivering them to the manufacturer or an authorized repair person.
Just this morning I was almost killed twice on my bike ride to work by two separate drivers, one of whom looked to be 80 and could barely see over the dashboard, and one who was on their phone. I didn’t even bother trying to remember the plate numbers, knowing that the odds of any kind of consequences are absolute zero.
No, we can’t go back to driving our own vehicles. Waymo everywhere and human driving outlawed, ASAP.
Agree. Multiple people I know have bought Teslas because they don’t trust themselves or their spouses to drive safely, and want them to use FSD. There should be incentives to get people onto self driving.
Different states have different rules about what sort of things insurers are allowed to charge different rates for. In the states that allow it, Tesla does offer insurance discounts for FSD usage.[1] Lemonade also offers discounts for FSD usage.[2]
That's a tough problem - distinguishing wet pavement from deep water. Humans make that mistake frequently. Autonomous vehicles should probably be equipped with a water sensor. (We did that in our DARPA Grand Challenge vehicle back in 2005). Then they can enter water very cautiously and see if it's too deep. This may make them too cautious about shallow puddles on roads, though.
Pretty sure the right answer mainly involves the car knowing about the weather and other emergency events.
It doesn't take much of a rainstorm to see localized flooding. Some debris over the storm drain is enough to flood a street. Hard to anticipate that happening.
It’s a particularly hard problem in Texas. We get torrential rains and the landscape is relatively flat. Couple that with shallow soil over lots of limestone and it means flooding is really common. We also have roads that have a “low water crossing,” where a road crosses a creekbed that is normally dry but which will flood. There are often water depth signs there (basically a vertical ruler with feet marks so you can see where the water is up to). We lose people to this scenario (driving into flood waters) every year. It’s particularly problematic when it’s dark and you miss a warning sign. Before you know it, you’re in deep water and the flow can sweep the whole car downstream until it gets pinned against a tree, possibly with water forcing its way into the car.
If they have a laser measurement of the road from before, couldn't they see that the level of water vs the expected road surface?
Such detailed database of fine grained road geometry gets stale very quickly, due to road maintenance and road construction. In US highway lanes are shifted sideways frequently.
But are they not continuously updating the road database with their fleet?
For common routes, yes. For getting to John's house, where the path there sometimes floods, no.
That seems a very risky assumption for any car (self driving or human driver) during flash floods. "Turn around don't drown":
You think you know how deep it is under because you've taken that road many times before (or in your case you have historical laser measurement)
But you don't know:
- Maybe the road under fully collapsed
- Maybe the flow of water is extremely strong, so you need to accurately estimate that too.
If they have a pre-existing database of every road, sure. And if it's kept up-to-date at all times in all vehicles.
Waymo does have a database of every today they drive, but for this they don't need one.
If the car comes to a road covered with water, and that road is in the database, and the water level appears low compared to the historical level of the road in the DB, then the car could cross. if the road is not in the DB, then a different decision might be made.
This is similar to humans: you might make different decisions depending on whether you know the road well or not.
Isn’t that the Waymo data model, though? They extensively pre-drive every new market, building dense volumetric maps of the entire service area before they begin service, so they essentially do have that database of every road (that they drive on).
Granted, I am not sure exactly how Waymo operates, but I thought that the extensive testing was mostly for legal reasons+just handling edge cases.
I am saying this, because I noticed that they typically start with a low-tier restrictive permit to operate (with a rather small number of cars in the fleet). Then they run it for a year or two, iron out edge cases particular to a given city (e.g., climate particularities, crazy spaghetti junctions in ATL, etc.), and log a lot of data. Then they take that data, go to the city/state, say "we have all this data that demonstrates we were very above the board while running the test pilot program, we are safe, and now we want to expand out of a very limited test pilot program."
And then it either goes well (Bay Area, LA, etc.) or goes off the rails for other reasons (often failing earlier for entirely unexpected reasons, like the pushback against it from taxi driver unions in NYC).
My point being, I could be entirely wrong, but I don't think that they literally map every single inch of the road before being allowed to operate. I just don't see it as being possible in any large populated city, given how often things change there. Just in 3 years living at one apartment in Seattle, I had a road directly adjacent to me changed from 2-way to 1-way, and then had 3-4 lanes that were basically highway entrances/exits (a block away from me) created and the whole area being rerouted entirely.
They actually do significant mapping. Where it operates currently it is not unusual to see this. It will be a waymo with a human driver operating someplace not currently in the waymo zone and clearly not en route to any maintenance facility either. Stuff like windy canyon roads with no thru access anywhere that are currently gated away, you might see a waymo with a human today.
Waymo is not the only company making lidar maps right now either. I've seen UPS deliver trucks with retrofitted lidar scanners on the roof now. I've even seen this on a police car already, looked like a black rooftop industrial ventilator on a 2ft mast installed directly on the crown victoria roof.
Waymo explicitly lidar scans and "HD maps" the area:
https://waymo.com/blog/2020/09/the-waymo-driver-handbook-map...
Tesla is less "HD", they have standard maps like we all think of, and a lane level "see-ahead" system where they basically just grab a satellite image tile, and align it with what the car sees for "FSD".
By a water sensor do you mean a sensor to detect the water level relative to the chassis? It seems like a very inexpensive downward-facing ultrasound sensor could work.
Is ultrasound less expensive than a moisture sensor?
The problem with both is they effectively require the vehicle to be in the water already. They need something that can tell depth before the vehicle has to slow down.
Doesn't Land Rover historically have like a wading sensor?
> frequently
I've never made that mistake; I'm not aware of anyone I know doing it. I very rarely see it myself, except on news footage. Of course it happens some time somewhere but that says nothing about frequency.
> That's a tough problem
Not really. Don't drive where you don't know it's safe. Definitely don't drive into moving water - puddles only, and only if not too deep: I can usually figure it out based on the rest of the road - unless it's a sinkhole, the geometry is somewhat consistent - and especially by looking at objects in the water such as other cars driving through it. Sorry your friend isn't competent to figure it out.
People here are always quick to defend the autonomous cars, like a close friend. How often will we fall in love with a technology or company? It always distorts the truth.
Article's current (possibly original), less ambiguous title: "Waymo recalls 3,800 robotaxis after glitch allowed some vehicles to ‘drive into standing water’"
IOW 3,800 Waymo vehicles aren't currently sat spinning their wheels in water.
"recall" = applies software update
The difference between that and usual software updates I'm guessing is the cars are pulled from service until the update takes place.
Also I think it's wrong to call something a recall if it's not owned by customers. Waymo is a service.
Gah, thanks for this. Thought I was used to that slight-of-hand but this one got me
Recall makes for better headlines.
aw, I was having fun imagining 3,800 Johnny cabs just immediately changing route to go to headquarters.
It's an interesting case of whether it's possible to infer the condition of wading and avoid having to install a sensor specific to a one in a million trips circumstance.
The inference would come from standing water slowing down the vehicle and likely require steering correction, in combination with some machine vision for identifying standing water.
Then there's the advantage of being Google and having hundreds of thousands of people in the same area using Google maps and navigation. Accelerometers in phones can detect crashes pretty reliably. There's a good chance they can reliably detect deceleration from standing water and report the location of the hazard.
Waymo: *locks doors, chorus to Floods by Pantera starts playing, guns it into the water*
“Wash away maaaaan, take him with the floooood”
How about a Mastodon, Lamb of God take with Floods of Triton:
Since recall on cars no longer means doing anything to the car's physical location I think the regulator NHSTA should update this term
It just creates alarmist headlines for what's really an over the air update, although "recall" is still currently a regulatory accurate term in the vehicle space
Cars, especially EVs, have many similarities to being phones. Imagine if a routine software update from Apple was called a "recall", that functionally describes what's happening here
NHTSA should at least distinguish between "omg we have to get these cars off the road and bring them to the shop immediately!" versus "over the air software update"
Does anyone with a better understanding about LIDAR vs camera approach to autonomous drivng explain how would Tesla handle such situation ?
Waymo has LIDAR and cameras, so it is better equipped for every situation.
Unless the power is out
https://abc7news.com/post/san-francisco-leaders-press-waymo-...
Kind of unrelated. That issue was due to a misguided effort to be cautious by having vehicles requesting human-review when they didn't really need it. Waymo fixed the issue by allowing the vehicles to operate in their normal, independent, mode.
LIDAR isn't helpful for water. Standing water behaves like a mirror on LIDAR.
This is one of the reasons why I'm suspicious of camera-only systems, here in Finland. Half the year there's a lot of snow and ice around. Which I imagine means most of the view is "white" and "shiny". Coupled with the dark winters it's gotta be a nightmare to deal with.
Could you use a different spectrum of EM radiation to detect water? There are parts of the microwave band that attenuate the signal by absorption and I wonder if you could use that. The only clue a human driver has in that situation is in the visible spectrum. The lines of the road disappear from view, which can be challenging to see at night.
If the LIDAR can sense the road close enough to the front of the car, then it could estimate how far underwater the car is.
They are rolling these out in New Orleans soon. Standing water is everywhere, and sometimes you have big hidden potholes. You just need to know the roads. Should be fun.
This is ok though because humans drive into flood waters too.
Look, you can't make progress without getting your feet wet and then diving straight into the deep end.
Maybe you drive into flood waters, but I don't. That's not a difficult skill to pull off.
We're still in the early days of self driving cars, and as much simulation and miles as they have, they're still constantly getting exposed to real world conditions that are new to them. The world is dynamic, so this will always remain true.
It remains to be seen where we'll converge on capability, incident rate, and acceptance.
Go fish
LeCun is right.
About what
That you need world models to sensibly deploy "thinking" machines in the real world. Else they do stupid shit like drive straight into water. You can bruteforce some semblance of thinking by training on literally all knowledge that can be digitized but even that is proving to not be quite enough.
[delayed]
What is a recall in this case? Is them getting a software update a recall now?
They suspended service areas they deem high risk until the software update can be applied. So while, yes, it's just a software update, it's a recall in the sense that they've temporarily pulled all the cars off the road in certain areas
I think so. For some kind of legalese reasons that's generally what a Tesla "recall" amounts to these days.
Yes, this is a common terminology issue. "Recall" is legally defined in terms of the kind of problems that require one, not the solution to those problems, because the relevant regulations were written when there was no way to fix consumer products other than physically delivering them to the manufacturer or an authorized repair person.
FFS, can we just go back to talking to each other in person and driving our own vehicles? Where'd the 90s go?
If the car drives itself we will have more time to talk to each other in person.
Or invest in public transport instead
Just this morning I was almost killed twice on my bike ride to work by two separate drivers, one of whom looked to be 80 and could barely see over the dashboard, and one who was on their phone. I didn’t even bother trying to remember the plate numbers, knowing that the odds of any kind of consequences are absolute zero. No, we can’t go back to driving our own vehicles. Waymo everywhere and human driving outlawed, ASAP.
Agree. Multiple people I know have bought Teslas because they don’t trust themselves or their spouses to drive safely, and want them to use FSD. There should be incentives to get people onto self driving.
Tesla cars are not capable of driving autonomously according to the company and regulators.
My dad doesn't care what the regulatory definition is, he just presses "Start Self-Driving" and off he goes.
If self-driving is better, then presumably cheaper insurance costs would be an incentive.
Different states have different rules about what sort of things insurers are allowed to charge different rates for. In the states that allow it, Tesla does offer insurance discounts for FSD usage.[1] Lemonade also offers discounts for FSD usage.[2]
1. https://www.tesla.com/support/insurance/fsd-discount
2. https://www.lemonade.com/fsd
> can we just go back to talking to each other in person
He posts on an internet message board