Driving through an obviously flooded street thinking "I'll easily make it" and getting stuck in the middle? Yeah, these cars have achieved human level intelligence.
That being said... it's actually somewhat uncommon for humans to drive into flooded streets. To the degree that people think it's notable enough to take videos and post them to social media. I don't have the data, but would be interested to see how many times per passenger mile travelled human-directed and remotely-operated vehicles like Weymos drove into flooded streets.
I can appreciate the cameras and lidar on the Weymos don't give their remote operators a lot of good data about the depth of water on the road-way. As you point out, humans in cars often don't get this right. I think the humans that don't drive into deep water are the ones who a) give any amount of water on the roadway a big NOPE and b) people familiar with the local environment and use multiple visual clues to judge the true depth of the flooding.
This is why I personally feel like Tesla's approach is more likely to "win". The fundamental blocker to self-driving cars is not sensing / sensor fusion, it is intelligence. And the Tesla approach seems much more likely to achieve functional intelligence than Waymo's.
While I agree with basically all of this, and find the FSD on my Tesla to be quite useful, but a question pops into my mind.
Why can't Waymo ALSO develop the same smarts and just also solve the sensor fusion issue such that they can use the right set of sensors in the right environmental conditions, and then leapfrog Tesla's capabilities?
They could in theory. If they put at least as much emphasis on the AI side as Tesla does. Or if someone else cracked vehicle AI wide open and left it open for them to copy, and then they did exactly that, and found a way to bolt on their extra sensors in a useful fashion while at it.
As is, Waymo's playing it smarter than Cruise did, but they're not all in on AI yet. So I don't expect them to "leapfrog Tesla" in that dimension - and it's the key dimension to self-driving.
> such that they can use the right set of sensors in the right environmental conditions
Because this part is really hard, and that's why Tesla abandoned the fusion approach. You cannot possibly foresee all the conditions in which LIDAR or any active sensor will malfunction/return wrong data/return data that's only slightly off for that ONE specific time. And even if it doesn't, you need to trust it to not return noise. And when it does return noise, how do you classify it as noise?
Cameras are passive sensors - they get whatever light comes in and turn it into an image. Camera is capturing shapes that make sense to the neural nets: it's working. See all black/white/red/cannot see any shapes? Camera is not working, exclude it from the currently used set of sensors or weigh it less when applying decisions, because it's returning no signal (and yes, neural nets have their own set of problems).
EDIT: cameras also provide more continuous context: if 1 pixel is off, is clearly bright red in a mostly-green scene where no poles can be identified, the neural net will average it out and discard it as noise. If 1 pixel says "object" in LIDAR, do you trust it to be correct? Perhaps the ray just hit a bird or a fly, but you only see a point, it's a lossy summary of the information you need.
I thought about this and I think it boils to how the model is trained.
Tesla trains it models from actual drivers purely based on (input) Vision and (output) actuators - Brake, Steering, Accelerators.
Human output is based on what they and the camera sees. So, it's a 1:1 match.
If Waymo were to do that, it'll muddle the training set. The Lidar input may override camera input.
I always struggled when Musk mentioned Lidar will make it ambiguous. It didn't make any sense to me why having a secondary failback sensor messes things. But, if you put it in the training data context, it absolutely makes sense.
Because they don't have a fleet of millions of people labeling the data for them and paying for the privilege of doing so. Waymo has about 3700 vehicles. Tesla has millions. Waymo only operates in known environments and collects a very limited range of data. Tesla collects data everywhere that people drive their cars.
I got downvoted for saying this last time the topic came up but constraints focus a project. It’s best to start work with as few variables as possible, and only add new ones when absolutely necessary.
I'm working on a similar problem in computer vision and we're quickly approaching the point where our pure vision work is better than our Lidar supported track because we've had to deal with the constraints instead of having a crutch to lean on.
This is really my bear case against AI. I am not against it. I actually think it is really neat! But we have been working on driverless cars for how long and spent how much? And still things like a flooded roadway completely throw them.
Tesla failed to deliver driverless cars but now is pivoting to the much more complex fully autonomous robots. And we can’t get AI to stop hallucinating facts, but any day we are going to be at AGI in a few years? I get people want these things to happen, but I just don’t see it happening any time soon. The whole tech industry feels built on what maybe, someday, possibly, could happen but most likely won’t, but we are all going to act like is a sure thing and is just around the corner.
Are there no responsible adults left at these tech companies?
I was (I think the search bar will prove this out) a pretty committed skeptic of driverless cars, but I've come around on them in some use cases. I'm not optimistic about them on highways. But they solve some important problems in regional/local transit.
We're contemplating standing up an EV shuttle service in Oak Park. It will fail. As I understand it, we've piloted non-EV versions of a shuttle service; they failed. The problem is that in small local areas, the staffing for a useful transit service is too expensive; that's because "useful" imposes constraints about responsiveness, coverage, and most of all hours of service, which mean the service won't pencil out with the ridership it'll get.
An autonomous vehicle transit service in our muni would probably work fine; it's a strict grid system with very low speed limits (AVs will, in our area, be strictly better drivers than the median human drivers --- this isn't a statement about human fallibility so much as an observation about scofflawry in our area). And if the product existed, we could afford it, because we wouldn't be paying fully loaded headcount costs for 2+ shifts of drivers at epsilon levels of utilization.
For whatever it's worth, I don't really have "autonomous vehicles" and "LLMs" in the same bucket in my head. I'm bullish on both, but for very different reasons. It usually doesn't occur to me to think of Waymos as "AI", though, obviously, they are.
I'm bullish on AI as a replacement for Uber from airports well behaved climates I frequent but bearish on how long it'll take to actually make a damn for me needing my car in Ohio until the mid-late 2030s at this rate. It's just so close and so far away at the same time.
Maybe a dumb question, why do electric cars have issues with water?
My understanding was that ICE cars have trouble because water get's drawn into the engine. Water in the engine causes it to stall. And the engine must have air in flow and out flow.
An electric car doesn't need air in the same way (no oxygen to ignite with gasoline, no air to compress and expand).
Shouldn't electric cars to much better at driving through water?
They can drive through surprisingly deep water, but you'd still rather avoid it for a lot of reasons. Dangerous loss of traction and risk of getting swept away, soaked passengers will want a refund, and a sopping wet interior will take the vehicle out of service for a while.
Snark aside, there will probably always be conditions in which waymo is not the right answer. Are they going to do hurricane evacuation? I think removing the driver just necessitates this.
While this is going to be an overly optimistic scenario: Imagine how smooth a hurricane evacuation would go if _everyone_ used a self-driving car to do the evacuation - atleast there might be less gridlock than there is during any usual hurricane evacuations. And assuming the self driving cars don't do something stupid that causes every car behind it to essentially lock up and stop moving
That said, I know a scenario like that would never happen, probably for the best.
The problem is they're not designed for that. They aren't spending resources on some master control networking system because in 99% of use cases that won't be useful anyways as most of the traffic being dealt with isn't other waymo's willing to communicate.
There might be some level of adoption where they would, but honestly we're back to "but what about trains/trucks?".
Half the problem with evacuations is people don't want to leave behind their stuff to get destroyed. You'd basically be better off getting a fleet of semi's with some quick and dirty cube system thrown up than a bunch of automated sedans.
Sort of. There is no built in support for evacuation methods, but the WayMo absolutely does use a master control system for network the cars. This is how the database of streets is kept and is why WayMo vehicles occasionally swarm private non through way ally streets when there is some glitch in the database that indicates private ways are available roads or an ally that looks like a through way turns out to have a fence between properties.
With human drivers: traffic light turns green. The first car starts driving. The 2nd car waits 2 seconds and then starts driving. The third car waits another 2 seconds (4 seconds total) and then starts driving. The fourth car waits another 2 seconds (6 seconds total) and then starts driving. etc.
With computers driving: traffic light turns green. All cars simultaneously start driving. It'd be like a train but without the efficiency.
Similarly, with human drivers: some jackasses drive into the box and the light turns red. Now perpendicular traffic is either fully blocked or must proceeed slower to maneuver around the jackasses. With computer drivers, they shouldn't intentionally break the law and they should have plenty of sensors to figure out that they cannot make it through the box.
As a sorta informed outsider, conceptually this makes intuitive sense. But in practice, how does this work? It seems a lot of the intuition breaks down if we don't assume it's network (aka 1 vendor). Fundamentally it's a bunch of external actors where we cannot verify trust and in order to solve for the needs of the individual, suboptimal choices must be made. To put it another way, even if computers can drive cars, what _else_ needs to be in place for this vision?
Safety margins still will require some level of delay between cars that aren't mechanically linked. Even with perfect reaction times, the physics of driving (maximum acceleration rates, possible loss of traction) dictate this, it's a non-trivial control theory problem. Besides, it doesn't seem to be a goal of Waymo; I've seen lines of their vehicles before and they all behave the same way as in mixed traffic.
Traffic is usually caused by adding inefficiencies across a system with little slack - someone brakes too hard or too early, and if all the cars are stacked up, that one brake event can ripple through hundreds of following cars, getting worse and worse because each person brakes more. Self driving cars can perfectly sync up and move like a train. Theoretically there could be no traffic on highways if all cars are self-driving. Rarely is a highway so full that there couldn’t be more cars (eg. The entrance ramps are backed up) which implies the issues are related to the driving flow and not the capacity of the street itself.
Ideally, robot drivers will some day be better drivers than humans in all road conditions. They'll be able to coordinate fast lane merges and busy intersections by subtly adjusting speed without vehicles having to stop.
Imagine a busy intersection where all the cars fly past one another at 40 miles an hour without stopping but none of them crash. Humans can't do this, but machines could, if, and when the technology gets there. To be clear, there's still a way to go.
Once all cars are autonomous, that day is certainly coming. Even before then, it's very likely we'll see platooning in the future, even if there are still some human drivers.
Also, this already exists in some places. Look at a video of how to cross the street as a pedestrian in Vietnam: You literally just start walking across and people weave around you. Or look at driving in India and similar places.
In principle the driverless cars are more able to organize fleeting, operating in a way that's not actually practical if you don't share a single guiding directive.
I don't know that you'd ever see this in practice, but it's much more practical in theory for almost identical machines running the same software than for a bunch of humans in a variety of vehicles who've maybe only half understood how to do this.
Also, for this specific problem we know humans are idiots. They should all be driving an agreed route to the agreed evacuation point, but some real humans will decide they know a shortcut, they want to drop past Jim's place, or whatever. Just as there's a difference between what the protocol says happens when you have to abandon an aircraft on the tarmac versus the reality that people will decide they want to self-evacuate and they need their carry on bags and chaos ensues and maybe people die.
Same reason there's less gridlock when people obey traffic lights and other rules of the road and don't brake randomly. If every car on the road drove itself then there would never be traffic.
Well, probably not the current generation of driverless cars. Those would be a nightmare. Contrary to what some want to believe self driving cars do random shit all the time.
But in the future, if there is a coordination standard among driverless cars, that could allow much higher density at higher speed. Coordination standards + higher density of self driving should reduce the self driving cars doing random shit too.
It would be a failure. Turns out they do something stupid. People tested this in sf by calling a bunch of waymos at once for a prank, but I guess that is the best case example of what a panicked evacuation on the service might be like. It was like a ddos attack. They ended up gridlocking themselves and turned it into a real life version of one of those rush hour board games. No one got out of the little area they called the waymos in.
I doubt it's less actual throughput in most cases. In a place like Atlanta there's no place where it's bus after bus. The BRT line they built nearby is a bus every 10 minutes. Which being very generous to the bus usage is equivalent to like 5 cars a minute.
Evacuation is a use case in my mind. Having a fleet of shuttles on command to move people in preparation of a hurricane would be a benefit. They would obviously need to put weather limitations during actual storms because no one should be driving in a hurricane.
Evacuation you want to prioritized throughput - think of how little road space 100 people in a bus take up vs say 50 cars with 2 people each. Or even 25 cars with 4 people each.
If you have central control you might even be able to get away with changing the rules. i.e. most roads are now one-way leading out of the city. voilà we nearly doubled outbound throughput. Even just for commuting that would be awesome, not that it is happening anytime soon, but one can dream, especially while sitting in gridlock traffic.
Guessing the depth of a puddle is not an easy task. Many untrained horses will refuse to step into shallow puddles. Then we also have human drivers driving into flooded road.
I wonder how much of this is trouble perceiving water depth vs integrating that understanding into the larger driver model without creating regressions elsewhere.
I don't think there's a good solution right now. You can't just go based on surrounding traffic because humans are also stupid and flood their cars all the time.
You could maybe use short-wave infrared cameras combined with ground penetrating radar, but it'll get real expensive so probably not commercially viable.
I think the only "good" solution is to have the car be overly paranoid, and if it detects water on the roadway that's bigger than some arbitrary diameter (to rule out mud puddles), then the car has to assume its a flood, stop, and escalate to a human or change the route.
Alternatively, just don't run Waymo operations during flood/flash flood warnings. Maybe we as a society need to top forcing everything to still operate normally during natural disasters. It's OK to shut things down when safety calls for it, and that applies to human drivers too. If areas are flooding, stay home.
> Alternatively, just don't run Waymo operations during flood/flash flood warnings.
FTA
> the company said that it shipped an update to its fleet that placed “restrictions at times and in locations where there is an elevated risk of encountering a flooded, higher-speed roadway,”
> But even those precautions apparently were not enough to stop the Waymo robotaxi from entering the flooded intersection in Atlanta. Waymo told TechCrunch on Thursday that the storm in Atlanta produced so much rainfall that flooding was happening before the National Weather Service had issued a flash flood warning, watch, or advisory.
Their fleet is constantly scanning the area with lidar, which is assembled into maps. If those maps are in 3d rather than a 2d road grid you can calculate puddles very accurately with no extra sensors:
- Find the edge of the water using vision or lidar
- look up the ground height at that position in your map data. That is the water level
- run a flood fill of the local 3d map starting from that point, with that water level. That gives you an exact shape of the puddle
- for any point on your planned path, you can now check if the point is in the puddle (per the flood fill above) and how deep the water is (difference between puddle's water level and ground height)
- use that either as a go/no-go for a planned path, or even feed this into your pathfinding to find a path with acceptable water level
The main limitation is that it assumes that the ground hasn't changed. It won't help in a landslide, or on muddy ground where other cars have disturbed the ground. But for the classic case of the flooded underpass or flooded dip in the road it should be very accurate
The vehicles have enough information to make the determination. Ground data is available in the point cloud and usually labeled as such. Water sometimes shows up in point clouds, sometimes it doesn't depending on conditions and wavelength.
If the apparent road surface is higher than the mapped ground surface, probably a puddle. If your point cloud has a big hole, also probably a puddle.
This assumes you aren't doing ground plane removal, of course. But it's quite likely that Waymo is using a heavily ML approach these days, and I can imagine the poor thing getting very confused if it's not an explicit training goal.
I feel like re-reading this sentence a few times sends me right to the twilight zone of AI psychosis.
It’s 2026 and self-driving cars can’t tell the difference between a puddle and a flooded street, something a 3 year old can do.
Google literally just got off stage telling us that AGI is almost here. Wake me up when this doesn’t feel like an NFT ape fever dream.
And here we are talking about this like “oh gosh golly I wonder if this is some simple thing that could have been easily solved but they were trying to avoid regressions”
Get out of town, man.
I wish every dollar spent by investors on Waymo went into more frequent public bus service instead. A regular-ass bus with a human driver.
During the “winter”, sure, but it dumps rain during the same and there are flash floods occasionally. I agree with the parent comment that Miami is a great area to test - especially given that the bad weather is seasonal. They can run 24/7 during the good weather seasons.
Also, the drivers in Miami are a bit more unpredictable than the average driver around the country in my experience, so good challenge cases for self-driving development.
Unpredictable drivers aren’t a challenge compared to weather. They’re just 3D objects to avoid. That’s a solved problem.
The thing about weather is that with a fully automated fleet they can just stop and give up on driving instantly. Rain in Miami doesn’t tend to last very long except in specific storms like hurricanes. Waymo can just not operate during those times.
I’m very doubtful that a lot of these inherent problems with the technology are being rapidly solved. See: the article.
I think another way of framing it is "Waymo pauses Atlanta service due to weather conditions", which doesn't sound at all unreasonable to me. It's no different from "Chicago O'Hare pauses flight departures due to a winter storm" or whatever.
I think that self driving cars won't ever be able to handle every condition out there, and so there's probably a time when the system will be paused / shutdown when conditions aren't safe to drive in. Honestly, I wish we could do this with human drivers for that matter, too, but some will press on even when they shouldn't...
Well except that there were incidents of cars getting stuck in floods with passengers before they paused the service.
A closer analogy would be ""Chicago O'Hare pauses flight departures due to a winter storm after 3 planes slide off the runway due to ice"
Absolutely I think there will be a disconnect between when people think they should be able to drive somewhere (ie to work in a no-visibility blizzard) and when ideal self-driving cars would allow themselves to operate. Maybe society will adjust to be more flexible to natural conditions, or maybe people will get frustrated and drive themselves into the poor conditions as always.
Self driving will never handle all corner cases until they essentially have a frontal cortex. They probably need something like an LLM to help with very high level abstract situations, e.g. avoiding a hurricane like someone else mentioned in this thread.
A frontal cortex isn't enough; there are plenty of corner cases that humans fail at too. The real test is if self-driving performs on par, or better than, humans in the vast majority of cases. If it saves 50,000 lives a year to go with self-driving, it's a net-win even if there are a few people who die in situations where they would have survived with a human driver behind the wheel.
Self driving cars are not going to be accepted if they have only marginally better success rates than humans. Just look at the news. Every minor self driving incident is endlessly magnified by the media while millions of human-caused accidents are just a part of life. That's just how our brains work. All major decisions are made primarily based on emotion, not analytics.
Human accidents don't get treated as "just a part of life", serious human driving errors are often considered so egregious that the person making the error picks up a driving ban or even a custodial sentence.
So it's actually entirely rational that the bar for companies to be able to ship software that makes those fatal errors without consequence other than an insurance payout should be higher (especially since when fatal error rates can only be estimated accurately over the order of millions of miles, driverless systems are more prone to systematic error or regression bugs than the equivalent sized set of human drivers, and the cost and appeal of autonomy probably means more experienced drivers get replaced first and more journeys get taken)
Humans don't handle all corner cases. People can be slow to react to completely novel or surprising situations. There will be corner cases where humans generally do better than a machine, but the simple rule to slow down and come to a halt if things look too weird or confusing will almost always be the right answer.
Ideally, driverless cars will one day be better drivers than humans and this will save tens of thousands of traffic deaths per year. Holding up progress because cars will be confused in extremely rare or improbable situations will cost more lives than it saves.
Not only are people slow to react to unusual situations, but this is taken advantage of by city designers to force people to slow down.
Random planters in the middle of the road? Streets that narrow and then widen? Drivers start slowly creeping along, which means they are less likely to injury pedestrians.
I think self-driving cars will only become better once they can do all the learning in real time and on-board. Otherwise, they will only be as good as the data they trained on - which is ultimately real meat driver data and a derivations of said data.
They will add flooded streets to the training simulation and this problem will go away. Eventually, the corner cases not in the training simulation will be so corner they basically never happen. Waymo can be incredibly successful without dealing with "surprise clown parade" or whatever.
The driving ML model will take care of the next 10 seconds of driving, in a fast loop deciding what steering and throttle commands to give.
The LLM will apply the high level reasoning needed to deal with longer time horizons and complex decisions, like deciding that the best way to reach the car wash 100 yards away is by walking.
they should probably put some sort of metal strip into the roads that a vehicle can follow reliably, future iterations could make continuous contact to the strip to deliver power to these vehicles, and this would also allow them to become larger by reducing fuel weight or even allow cars to travel very close together for efficiency gains
Clearly they haven't actually had any serious problems getting stuck or anything because it'd be all over the news.
I don't think they're barreling into foot+ deep water.
I think they're driving into shallower "perfectly navigable but still deep" puddles at normal for the roads speed and this pizza delivery boy type behavior is making passengers clutch their pearls because they are expecting their robotaxi to drive like a high end chauffeur.
> One of Waymo’s robotaxis was spotted driving through a flooded street in Atlanta, Georgia on Wednesday before it ultimately got stuck for about an hour, according to local news reports. The vehicle was recovered and removed from the scene, Waymo told TechCrunch. Waymo says it paused service in the city, just like it has in San Antonio, Texas, while it figures out a solution.
That title sounds so much more dramatic than it seems it actually was. I imagine headlines like: “Billions of python 3.14.4 programs were recalled today when a bug was found in the core itself. No word yet on whether the successor product, Python 3.14.5, will avoid a similar fate. How long will we tolerate being used as test subjects in the developer’s risky games?”
I thought Weymo's were supposed to be "supervised" by humans in the Philippines. Maybe driving in circles in the suburbs and driving into flood waters happens only when the cars are out of mobile data range? Did Weymo pay their mobile phone bill? Does the (somewhat) autonomous system on the car decide when to flag a human for help? I would have expected a human to be watching all the time. Are they experiencing labor problems in the Philippines? Maybe Weymo doesn't want to pay their remote operators as much as the remote operators want to get paid?
Driving through an obviously flooded street thinking "I'll easily make it" and getting stuck in the middle? Yeah, these cars have achieved human level intelligence.
Just get a jeep snorkle
What happens when you you start floating?
I guess water propulsion... and a rudder?
That being said... it's actually somewhat uncommon for humans to drive into flooded streets. To the degree that people think it's notable enough to take videos and post them to social media. I don't have the data, but would be interested to see how many times per passenger mile travelled human-directed and remotely-operated vehicles like Weymos drove into flooded streets.
I can appreciate the cameras and lidar on the Weymos don't give their remote operators a lot of good data about the depth of water on the road-way. As you point out, humans in cars often don't get this right. I think the humans that don't drive into deep water are the ones who a) give any amount of water on the roadway a big NOPE and b) people familiar with the local environment and use multiple visual clues to judge the true depth of the flooding.
Let’s redirect the problem: it’s not the car, it’s the flooding! We should address that first
This is why I personally feel like Tesla's approach is more likely to "win". The fundamental blocker to self-driving cars is not sensing / sensor fusion, it is intelligence. And the Tesla approach seems much more likely to achieve functional intelligence than Waymo's.
While I agree with basically all of this, and find the FSD on my Tesla to be quite useful, but a question pops into my mind.
Why can't Waymo ALSO develop the same smarts and just also solve the sensor fusion issue such that they can use the right set of sensors in the right environmental conditions, and then leapfrog Tesla's capabilities?
They could in theory. If they put at least as much emphasis on the AI side as Tesla does. Or if someone else cracked vehicle AI wide open and left it open for them to copy, and then they did exactly that, and found a way to bolt on their extra sensors in a useful fashion while at it.
As is, Waymo's playing it smarter than Cruise did, but they're not all in on AI yet. So I don't expect them to "leapfrog Tesla" in that dimension - and it's the key dimension to self-driving.
> such that they can use the right set of sensors in the right environmental conditions
Because this part is really hard, and that's why Tesla abandoned the fusion approach. You cannot possibly foresee all the conditions in which LIDAR or any active sensor will malfunction/return wrong data/return data that's only slightly off for that ONE specific time. And even if it doesn't, you need to trust it to not return noise. And when it does return noise, how do you classify it as noise?
Cameras are passive sensors - they get whatever light comes in and turn it into an image. Camera is capturing shapes that make sense to the neural nets: it's working. See all black/white/red/cannot see any shapes? Camera is not working, exclude it from the currently used set of sensors or weigh it less when applying decisions, because it's returning no signal (and yes, neural nets have their own set of problems).
EDIT: cameras also provide more continuous context: if 1 pixel is off, is clearly bright red in a mostly-green scene where no poles can be identified, the neural net will average it out and discard it as noise. If 1 pixel says "object" in LIDAR, do you trust it to be correct? Perhaps the ray just hit a bird or a fly, but you only see a point, it's a lossy summary of the information you need.
There is noise on LIDAR returns too. No one considers a single LIDAR point to be a collision hazard.
I thought about this and I think it boils to how the model is trained.
Tesla trains it models from actual drivers purely based on (input) Vision and (output) actuators - Brake, Steering, Accelerators.
Human output is based on what they and the camera sees. So, it's a 1:1 match.
If Waymo were to do that, it'll muddle the training set. The Lidar input may override camera input.
I always struggled when Musk mentioned Lidar will make it ambiguous. It didn't make any sense to me why having a secondary failback sensor messes things. But, if you put it in the training data context, it absolutely makes sense.
Because they don't have a fleet of millions of people labeling the data for them and paying for the privilege of doing so. Waymo has about 3700 vehicles. Tesla has millions. Waymo only operates in known environments and collects a very limited range of data. Tesla collects data everywhere that people drive their cars.
I got downvoted for saying this last time the topic came up but constraints focus a project. It’s best to start work with as few variables as possible, and only add new ones when absolutely necessary.
I'm working on a similar problem in computer vision and we're quickly approaching the point where our pure vision work is better than our Lidar supported track because we've had to deal with the constraints instead of having a crutch to lean on.
You can have intelligence with lidar.
You can have even more intelligence with both.
They never advertised that they did. Its not even real true AI. They just struggle with new scenarios.
People drive into floods too. They just don't get sensational articles written about it, just posted on reddit.
Taxi drivers with passengers don’t tend to though. At least not at the same rate.
Whoosh...
This is really my bear case against AI. I am not against it. I actually think it is really neat! But we have been working on driverless cars for how long and spent how much? And still things like a flooded roadway completely throw them.
Tesla failed to deliver driverless cars but now is pivoting to the much more complex fully autonomous robots. And we can’t get AI to stop hallucinating facts, but any day we are going to be at AGI in a few years? I get people want these things to happen, but I just don’t see it happening any time soon. The whole tech industry feels built on what maybe, someday, possibly, could happen but most likely won’t, but we are all going to act like is a sure thing and is just around the corner.
Are there no responsible adults left at these tech companies?
I was (I think the search bar will prove this out) a pretty committed skeptic of driverless cars, but I've come around on them in some use cases. I'm not optimistic about them on highways. But they solve some important problems in regional/local transit.
We're contemplating standing up an EV shuttle service in Oak Park. It will fail. As I understand it, we've piloted non-EV versions of a shuttle service; they failed. The problem is that in small local areas, the staffing for a useful transit service is too expensive; that's because "useful" imposes constraints about responsiveness, coverage, and most of all hours of service, which mean the service won't pencil out with the ridership it'll get.
An autonomous vehicle transit service in our muni would probably work fine; it's a strict grid system with very low speed limits (AVs will, in our area, be strictly better drivers than the median human drivers --- this isn't a statement about human fallibility so much as an observation about scofflawry in our area). And if the product existed, we could afford it, because we wouldn't be paying fully loaded headcount costs for 2+ shifts of drivers at epsilon levels of utilization.
For whatever it's worth, I don't really have "autonomous vehicles" and "LLMs" in the same bucket in my head. I'm bullish on both, but for very different reasons. It usually doesn't occur to me to think of Waymos as "AI", though, obviously, they are.
I'm bullish on AI as a replacement for Uber from airports well behaved climates I frequent but bearish on how long it'll take to actually make a damn for me needing my car in Ohio until the mid-late 2030s at this rate. It's just so close and so far away at the same time.
Maybe a dumb question, why do electric cars have issues with water?
My understanding was that ICE cars have trouble because water get's drawn into the engine. Water in the engine causes it to stall. And the engine must have air in flow and out flow.
An electric car doesn't need air in the same way (no oxygen to ignite with gasoline, no air to compress and expand).
Shouldn't electric cars to much better at driving through water?
They can drive through surprisingly deep water, but you'd still rather avoid it for a lot of reasons. Dangerous loss of traction and risk of getting swept away, soaked passengers will want a refund, and a sopping wet interior will take the vehicle out of service for a while.
Deep water can still damage an EV by getting into connectors, sensors, wheel bearings, brakes, and cabin electronics.
They can also float just like a regular car.
Snark aside, there will probably always be conditions in which waymo is not the right answer. Are they going to do hurricane evacuation? I think removing the driver just necessitates this.
While this is going to be an overly optimistic scenario: Imagine how smooth a hurricane evacuation would go if _everyone_ used a self-driving car to do the evacuation - atleast there might be less gridlock than there is during any usual hurricane evacuations. And assuming the self driving cars don't do something stupid that causes every car behind it to essentially lock up and stop moving
That said, I know a scenario like that would never happen, probably for the best.
The problem is they're not designed for that. They aren't spending resources on some master control networking system because in 99% of use cases that won't be useful anyways as most of the traffic being dealt with isn't other waymo's willing to communicate.
There might be some level of adoption where they would, but honestly we're back to "but what about trains/trucks?".
Half the problem with evacuations is people don't want to leave behind their stuff to get destroyed. You'd basically be better off getting a fleet of semi's with some quick and dirty cube system thrown up than a bunch of automated sedans.
Sort of. There is no built in support for evacuation methods, but the WayMo absolutely does use a master control system for network the cars. This is how the database of streets is kept and is why WayMo vehicles occasionally swarm private non through way ally streets when there is some glitch in the database that indicates private ways are available roads or an ally that looks like a through way turns out to have a fence between properties.
Why would there be less gridlock if people were in a driverless car instead of a regular car?
With human drivers: traffic light turns green. The first car starts driving. The 2nd car waits 2 seconds and then starts driving. The third car waits another 2 seconds (4 seconds total) and then starts driving. The fourth car waits another 2 seconds (6 seconds total) and then starts driving. etc.
With computers driving: traffic light turns green. All cars simultaneously start driving. It'd be like a train but without the efficiency.
Similarly, with human drivers: some jackasses drive into the box and the light turns red. Now perpendicular traffic is either fully blocked or must proceeed slower to maneuver around the jackasses. With computer drivers, they shouldn't intentionally break the law and they should have plenty of sensors to figure out that they cannot make it through the box.
As a sorta informed outsider, conceptually this makes intuitive sense. But in practice, how does this work? It seems a lot of the intuition breaks down if we don't assume it's network (aka 1 vendor). Fundamentally it's a bunch of external actors where we cannot verify trust and in order to solve for the needs of the individual, suboptimal choices must be made. To put it another way, even if computers can drive cars, what _else_ needs to be in place for this vision?
Safety margins still will require some level of delay between cars that aren't mechanically linked. Even with perfect reaction times, the physics of driving (maximum acceleration rates, possible loss of traction) dictate this, it's a non-trivial control theory problem. Besides, it doesn't seem to be a goal of Waymo; I've seen lines of their vehicles before and they all behave the same way as in mixed traffic.
Traffic is usually caused by adding inefficiencies across a system with little slack - someone brakes too hard or too early, and if all the cars are stacked up, that one brake event can ripple through hundreds of following cars, getting worse and worse because each person brakes more. Self driving cars can perfectly sync up and move like a train. Theoretically there could be no traffic on highways if all cars are self-driving. Rarely is a highway so full that there couldn’t be more cars (eg. The entrance ramps are backed up) which implies the issues are related to the driving flow and not the capacity of the street itself.
> Rarely is a highway so full that there couldn’t be more cars
Yep, here in Chicago you might even go as many as 12 hours between such events
Ideally, robot drivers will some day be better drivers than humans in all road conditions. They'll be able to coordinate fast lane merges and busy intersections by subtly adjusting speed without vehicles having to stop.
Imagine a busy intersection where all the cars fly past one another at 40 miles an hour without stopping but none of them crash. Humans can't do this, but machines could, if, and when the technology gets there. To be clear, there's still a way to go.
Evidence suggests... no, that day is never coming.
Once all cars are autonomous, that day is certainly coming. Even before then, it's very likely we'll see platooning in the future, even if there are still some human drivers.
Also, this already exists in some places. Look at a video of how to cross the street as a pedestrian in Vietnam: You literally just start walking across and people weave around you. Or look at driving in India and similar places.
All I'm saying is never say never
busy intersections have more than just cars, my jay walking is going to cause a massive pile up
In principle the driverless cars are more able to organize fleeting, operating in a way that's not actually practical if you don't share a single guiding directive.
I don't know that you'd ever see this in practice, but it's much more practical in theory for almost identical machines running the same software than for a bunch of humans in a variety of vehicles who've maybe only half understood how to do this.
Also, for this specific problem we know humans are idiots. They should all be driving an agreed route to the agreed evacuation point, but some real humans will decide they know a shortcut, they want to drop past Jim's place, or whatever. Just as there's a difference between what the protocol says happens when you have to abandon an aircraft on the tarmac versus the reality that people will decide they want to self-evacuate and they need their carry on bags and chaos ensues and maybe people die.
Same reason there's less gridlock when people obey traffic lights and other rules of the road and don't brake randomly. If every car on the road drove itself then there would never be traffic.
This is literally not true, roads still have finite capacity, and sometimes demand exceeds capacity.
Well, probably not the current generation of driverless cars. Those would be a nightmare. Contrary to what some want to believe self driving cars do random shit all the time.
But in the future, if there is a coordination standard among driverless cars, that could allow much higher density at higher speed. Coordination standards + higher density of self driving should reduce the self driving cars doing random shit too.
It would be a failure. Turns out they do something stupid. People tested this in sf by calling a bunch of waymos at once for a prank, but I guess that is the best case example of what a panicked evacuation on the service might be like. It was like a ddos attack. They ended up gridlocking themselves and turned it into a real life version of one of those rush hour board games. No one got out of the little area they called the waymos in.
I mean the logical conclusion is a dedicated lane for automated cars..
At which point we've reinvented privatized buses with a last mile convenience vs greatly reduced throughput trade-off.
I doubt it's less actual throughput in most cases. In a place like Atlanta there's no place where it's bus after bus. The BRT line they built nearby is a bus every 10 minutes. Which being very generous to the bus usage is equivalent to like 5 cars a minute.
Just take away the sidewalk and bike lane :-/
Evacuation is a use case in my mind. Having a fleet of shuttles on command to move people in preparation of a hurricane would be a benefit. They would obviously need to put weather limitations during actual storms because no one should be driving in a hurricane.
Evacuation you want to prioritized throughput - think of how little road space 100 people in a bus take up vs say 50 cars with 2 people each. Or even 25 cars with 4 people each.
If you have central control you might even be able to get away with changing the rules. i.e. most roads are now one-way leading out of the city. voilà we nearly doubled outbound throughput. Even just for commuting that would be awesome, not that it is happening anytime soon, but one can dream, especially while sitting in gridlock traffic.
> No one should be driving in a hurricane.
I agree, but there are a number of people here in Florida who will do it or die trying (emphasis on the die trying)
Except the Waymo can do 150 mph bumper to bumper with other Waymos if you let them.
.. well until it hits the flood
Guessing the depth of a puddle is not an easy task. Many untrained horses will refuse to step into shallow puddles. Then we also have human drivers driving into flooded road.
I wonder how much of this is trouble perceiving water depth vs integrating that understanding into the larger driver model without creating regressions elsewhere.
I don't think there's a good solution right now. You can't just go based on surrounding traffic because humans are also stupid and flood their cars all the time.
You could maybe use short-wave infrared cameras combined with ground penetrating radar, but it'll get real expensive so probably not commercially viable.
I think the only "good" solution is to have the car be overly paranoid, and if it detects water on the roadway that's bigger than some arbitrary diameter (to rule out mud puddles), then the car has to assume its a flood, stop, and escalate to a human or change the route.
Alternatively, just don't run Waymo operations during flood/flash flood warnings. Maybe we as a society need to top forcing everything to still operate normally during natural disasters. It's OK to shut things down when safety calls for it, and that applies to human drivers too. If areas are flooding, stay home.
> Alternatively, just don't run Waymo operations during flood/flash flood warnings.
FTA
> the company said that it shipped an update to its fleet that placed “restrictions at times and in locations where there is an elevated risk of encountering a flooded, higher-speed roadway,”
> But even those precautions apparently were not enough to stop the Waymo robotaxi from entering the flooded intersection in Atlanta. Waymo told TechCrunch on Thursday that the storm in Atlanta produced so much rainfall that flooding was happening before the National Weather Service had issued a flash flood warning, watch, or advisory.
Their fleet is constantly scanning the area with lidar, which is assembled into maps. If those maps are in 3d rather than a 2d road grid you can calculate puddles very accurately with no extra sensors:
- Find the edge of the water using vision or lidar
- look up the ground height at that position in your map data. That is the water level
- run a flood fill of the local 3d map starting from that point, with that water level. That gives you an exact shape of the puddle
- for any point on your planned path, you can now check if the point is in the puddle (per the flood fill above) and how deep the water is (difference between puddle's water level and ground height)
- use that either as a go/no-go for a planned path, or even feed this into your pathfinding to find a path with acceptable water level
The main limitation is that it assumes that the ground hasn't changed. It won't help in a landslide, or on muddy ground where other cars have disturbed the ground. But for the classic case of the flooded underpass or flooded dip in the road it should be very accurate
The vehicles have enough information to make the determination. Ground data is available in the point cloud and usually labeled as such. Water sometimes shows up in point clouds, sometimes it doesn't depending on conditions and wavelength.
If the apparent road surface is higher than the mapped ground surface, probably a puddle. If your point cloud has a big hole, also probably a puddle.
This assumes you aren't doing ground plane removal, of course. But it's quite likely that Waymo is using a heavily ML approach these days, and I can imagine the poor thing getting very confused if it's not an explicit training goal.
Do you how often you get flash warnings in Atlanta? And local roads flood far more often than flash food warnings are issued.
If you can’t handle this issue, you really can’t operate in Atlanta.
Would be interesting if you can compare the surface roughness of pavement vs. the surface of water, wind would disturb it too
In many situations, the depth of the water doesn't matter as driving into it will likely result in death.
I feel like re-reading this sentence a few times sends me right to the twilight zone of AI psychosis.
It’s 2026 and self-driving cars can’t tell the difference between a puddle and a flooded street, something a 3 year old can do.
Google literally just got off stage telling us that AGI is almost here. Wake me up when this doesn’t feel like an NFT ape fever dream.
And here we are talking about this like “oh gosh golly I wonder if this is some simple thing that could have been easily solved but they were trying to avoid regressions”
Get out of town, man.
I wish every dollar spent by investors on Waymo went into more frequent public bus service instead. A regular-ass bus with a human driver.
I assumed they went to Miami to develop their foul weather capabilities. It's still pretty early.
Doubtful. They probably just pause service when it rains. Miami weather is ideal most of the time.
These self-driving companies have made very little progress on dealing with weather for how long they’ve spent on the problem.
During the “winter”, sure, but it dumps rain during the same and there are flash floods occasionally. I agree with the parent comment that Miami is a great area to test - especially given that the bad weather is seasonal. They can run 24/7 during the good weather seasons.
Also, the drivers in Miami are a bit more unpredictable than the average driver around the country in my experience, so good challenge cases for self-driving development.
Unpredictable drivers aren’t a challenge compared to weather. They’re just 3D objects to avoid. That’s a solved problem.
The thing about weather is that with a fully automated fleet they can just stop and give up on driving instantly. Rain in Miami doesn’t tend to last very long except in specific storms like hurricanes. Waymo can just not operate during those times.
I’m very doubtful that a lot of these inherent problems with the technology are being rapidly solved. See: the article.
I think another way of framing it is "Waymo pauses Atlanta service due to weather conditions", which doesn't sound at all unreasonable to me. It's no different from "Chicago O'Hare pauses flight departures due to a winter storm" or whatever.
I think that self driving cars won't ever be able to handle every condition out there, and so there's probably a time when the system will be paused / shutdown when conditions aren't safe to drive in. Honestly, I wish we could do this with human drivers for that matter, too, but some will press on even when they shouldn't...
Well except that there were incidents of cars getting stuck in floods with passengers before they paused the service.
A closer analogy would be ""Chicago O'Hare pauses flight departures due to a winter storm after 3 planes slide off the runway due to ice"
Absolutely I think there will be a disconnect between when people think they should be able to drive somewhere (ie to work in a no-visibility blizzard) and when ideal self-driving cars would allow themselves to operate. Maybe society will adjust to be more flexible to natural conditions, or maybe people will get frustrated and drive themselves into the poor conditions as always.
Is it so hard for LiDAR/Camera to detect flood water on road. Water on a road looks like a flat surface to sensors.
hard part is that cars should drive through shallow water... but how to know the depth?
given accurate mapping + realtime imaging, this should be possible albeit a Big Project(tm).
Biblical.
Coming to New Orleans soon...
Self driving will never handle all corner cases until they essentially have a frontal cortex. They probably need something like an LLM to help with very high level abstract situations, e.g. avoiding a hurricane like someone else mentioned in this thread.
A frontal cortex isn't enough; there are plenty of corner cases that humans fail at too. The real test is if self-driving performs on par, or better than, humans in the vast majority of cases. If it saves 50,000 lives a year to go with self-driving, it's a net-win even if there are a few people who die in situations where they would have survived with a human driver behind the wheel.
Self driving cars are not going to be accepted if they have only marginally better success rates than humans. Just look at the news. Every minor self driving incident is endlessly magnified by the media while millions of human-caused accidents are just a part of life. That's just how our brains work. All major decisions are made primarily based on emotion, not analytics.
Human accidents don't get treated as "just a part of life", serious human driving errors are often considered so egregious that the person making the error picks up a driving ban or even a custodial sentence.
So it's actually entirely rational that the bar for companies to be able to ship software that makes those fatal errors without consequence other than an insurance payout should be higher (especially since when fatal error rates can only be estimated accurately over the order of millions of miles, driverless systems are more prone to systematic error or regression bugs than the equivalent sized set of human drivers, and the cost and appeal of autonomy probably means more experienced drivers get replaced first and more journeys get taken)
Maybe. But insurance rates, and the government's enforcement of laws, are based on analytics, and overcome a lot of human emotional bias.
Humans don't handle all corner cases. People can be slow to react to completely novel or surprising situations. There will be corner cases where humans generally do better than a machine, but the simple rule to slow down and come to a halt if things look too weird or confusing will almost always be the right answer.
Ideally, driverless cars will one day be better drivers than humans and this will save tens of thousands of traffic deaths per year. Holding up progress because cars will be confused in extremely rare or improbable situations will cost more lives than it saves.
Not only are people slow to react to unusual situations, but this is taken advantage of by city designers to force people to slow down.
Random planters in the middle of the road? Streets that narrow and then widen? Drivers start slowly creeping along, which means they are less likely to injury pedestrians.
I think self-driving cars will only become better once they can do all the learning in real time and on-board. Otherwise, they will only be as good as the data they trained on - which is ultimately real meat driver data and a derivations of said data.
They will add flooded streets to the training simulation and this problem will go away. Eventually, the corner cases not in the training simulation will be so corner they basically never happen. Waymo can be incredibly successful without dealing with "surprise clown parade" or whatever.
this is absolutely already a thing under development, you can see Waymo is hiring for reasoning roles
how would a llm help
maybe a little biological brain engineered to think it is a car with api access to the car hardware via the llm?
imagine you get into the car and in the center console you just see a floating brain in vat like fallout
The driving ML model will take care of the next 10 seconds of driving, in a fast loop deciding what steering and throttle commands to give.
The LLM will apply the high level reasoning needed to deal with longer time horizons and complex decisions, like deciding that the best way to reach the car wash 100 yards away is by walking.
they should probably put some sort of metal strip into the roads that a vehicle can follow reliably, future iterations could make continuous contact to the strip to deliver power to these vehicles, and this would also allow them to become larger by reducing fuel weight or even allow cars to travel very close together for efficiency gains
Clearly they haven't actually had any serious problems getting stuck or anything because it'd be all over the news.
I don't think they're barreling into foot+ deep water.
I think they're driving into shallower "perfectly navigable but still deep" puddles at normal for the roads speed and this pizza delivery boy type behavior is making passengers clutch their pearls because they are expecting their robotaxi to drive like a high end chauffeur.
> One of Waymo’s robotaxis was spotted driving through a flooded street in Atlanta, Georgia on Wednesday before it ultimately got stuck for about an hour, according to local news reports. The vehicle was recovered and removed from the scene, Waymo told TechCrunch. Waymo says it paused service in the city, just like it has in San Antonio, Texas, while it figures out a solution.
Thousands of Waymos recalled after robotaxi swept into a creek https://www.bbc.co.uk/news/articles/cwy2011dl4xo
> It follows an incident on 20 April in San Antonio, Texas, where an empty Waymo vehicle entered a flooded road and was swept into a creek.
Nobody in it but sounds serious enough.
That title sounds so much more dramatic than it seems it actually was. I imagine headlines like: “Billions of python 3.14.4 programs were recalled today when a bug was found in the core itself. No word yet on whether the successor product, Python 3.14.5, will avoid a similar fate. How long will we tolerate being used as test subjects in the developer’s risky games?”
There was one in Atlanta that made the local news where it went too deep and stalled out, was stuck for over an hour.
does Waymo use Lidar or is it like Musk's "cost saving" cameras only
Waymo uses lidar. There's lots of information about it on the web.
Lidar is much less accurate in the rain.
The spinny things on the vehicle are LIDAR.
If they only would use lidar. Oh wait…
I thought Weymo's were supposed to be "supervised" by humans in the Philippines. Maybe driving in circles in the suburbs and driving into flood waters happens only when the cars are out of mobile data range? Did Weymo pay their mobile phone bill? Does the (somewhat) autonomous system on the car decide when to flag a human for help? I would have expected a human to be watching all the time. Are they experiencing labor problems in the Philippines? Maybe Weymo doesn't want to pay their remote operators as much as the remote operators want to get paid?
Your assumption that Waymos are "(somewhat) autonomous" is wrong, which is why your questions and conclusion don't make any sense
It's an interesting illustration of how widely and quickly misinformation spreads, though.
Maybe the solution is to put in more billions. Every fad creates jobs.