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.
I wonder how much of this is trouble perceiving water depth vs integrating that understanding into the larger driver model without creating regressions elsewhere.
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.
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.
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.
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
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?
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...
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?
Robots vs cars? Robots are much much safer. One is a killing machine going 100km/h, other one is a slow moving thing. You don’t need fast reaction for a robot. Tesla should’ve started with robots first and then self driving.
> And still things like a flooded roadway completely throw them.
I guess that's what you get when you test your cars far 20 years in a state that's almost perpetually in a state of drought.
On the other hand, as someone who grew up in New England, laughing about news stories of highways in warmer states getting backed up because of an inch or two or snow wasn't an uncommon occurrence, so maybe having trouble driving during unfamiliar weather is just a sign that they're learning to drive like humans too well
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.
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?
To me this doesn't seem like a disaster but just the kind of thing that happens as you role out a service and expose it to new challenges.
Presumably they haven't had the chance to do a lot of flood training but now they have that chance.
The huge advantage they have over people in general is that ideally if they figure this out then it will stay figured out. Then they can slowly role out and watch for the next hitches from new situations.
> The huge advantage they have over people in general is that ideally if they figure this out then it will stay figured out. Then they can slowly role out and watch for the next hitches from new situations.
Sure, because human drivers famously have to be taught with each new generation that driving into six feet of water is a bad plan.
To me it looks like it's a problem with the "default attitude" (can't think of a better name) of the Waymo driving software. When a human sees that the road surface ahead is in some unknown condition (flooded, covered in lava, whatever) they usually default to caution - better stop and check first. While Waymo apparently defaults to blithely driving ahead, after all its maps tell it that there's a road ahead and it didn't detect any known obstacle, so what could possibly go wrong?
> The huge advantage they have over people in general is that ideally if they figure this out then it will stay figured out. Then they can slowly role out and watch for the next hitches from new situations.
That is not a given when dealing with "machine learning".
They will need to have metrics for all these scenarious and ensure when they solve the 20th problem down the line this one does not regress, but instead it becomes more and more generalized.
I can already see the horrified passengers in a robo-taxi going full "military-survival" mode, driving at rally speed over fast flooding back-roads, evaluating moral dilemmas like ("If i stop and pick up one more, i become a lorry on a rail at the next flood intersection").
Surprisingly good at things that get you otherwise killed.
Like - it auto-backs up once it detects ground rumbles of the ground moving during a mud avalanche.
An alternate viewpoint is that it looks like after 20 years they still haven't even started solving weather issues that you encounter anywhere outside a California climate.
We already have a huge number of safety regulations for cars, that take into account all these various things. There's also insurance that covers flood damage and cars. These are the things that red flag something you need to test, if you want to take over driving the car.
Sounds like they need to employ more "neurodivergents" to make these robots work correctly, before they are all Silent Greened, and it is only the CEOs left bashing each other's heads in with rocks.
And to me it seems like you're justifying a lack of oversight and dangers of this technology for what purpose exactly? Why are you defending a corporation?
This is just part of the slog that autonomous driving was always going to be.
Many many years ago I happened to be in a conversation with one of the guys on a team that participated in the 2005 DARPA Grand Challenge. It was only the second such race after the 2004 one, but arguably the one which set off the autonomous driving race we see today. (Sebastian Thrun's team came in 2nd.)
I went into the conversation thinking it was going to be an extremely challenging but tractable sensors + control-systems problem. But by the end of the conversation I was like, OMG this is going to be a long-haul slog of solving an unending stream of problems, some potentially even AI-complete (i.e. requiring human-level judgment.)
We mostly discussed why his and most other teams failed and the failures were so myriad and so technically intractable that I could not see a path to full self-driving for at least two decades. And all of this was offroad, so it didn't even approach the challenges of sharing human-occupied streets. I cannot remember any details unfortunately, but I remember that one car got stuck in a loop due to a problem that would have been trivial for a human to bypass... but that required human-level judgment. As an analogy it was something like a soft obstacle that could safely be driven over. But for the car to know that it would require a database and an "understanding" of all possible obstacles. An LLM could have helped, but back then they were still firmly in the realm of SciFi.
So the only feasible solution was to painstakingly identify all the edge-cases and work through them slowly, carefully, one-by-one. Which is what Waymo has been doing. This is also why when Elon made his "full self-"driving announcements I knew he had absolutely NO idea what he was talking about, and he was likely going to move fast and break people.
Flooded streets is just another "bump on the road" to full self-driving, but it seems we're actually getting there now. In retrospect, my 2-decade estimate was surprisingly accurate, I have no idea how I landed on that particular number!
What are the chances that google just shuts down waymo once they get whatever they need from it. Weren't there other ambitious projects under google that had a similar fate?
72 comments
[ 2.2 ms ] story [ 72.3 ms ] threadI 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.
given accurate mapping + realtime imaging, this should be possible albeit a Big Project(tm).
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...
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 guess that's what you get when you test your cars far 20 years in a state that's almost perpetually in a state of drought.
On the other hand, as someone who grew up in New England, laughing about news stories of highways in warmer states getting backed up because of an inch or two or snow wasn't an uncommon occurrence, so maybe having trouble driving during unfamiliar weather is just a sign that they're learning to drive like humans too well
- Only ship products when they are perfect against every possible edge condition?
or
- If Waymo fails in a few scenarios today after 20 years of effort, they can never succeed?
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?
Presumably they haven't had the chance to do a lot of flood training but now they have that chance.
The huge advantage they have over people in general is that ideally if they figure this out then it will stay figured out. Then they can slowly role out and watch for the next hitches from new situations.
Sure, because human drivers famously have to be taught with each new generation that driving into six feet of water is a bad plan.
That is not a given when dealing with "machine learning".
They will need to have metrics for all these scenarious and ensure when they solve the 20th problem down the line this one does not regress, but instead it becomes more and more generalized.
Surprisingly good at things that get you otherwise killed. Like - it auto-backs up once it detects ground rumbles of the ground moving during a mud avalanche.
This isn't a new challenge - it is a known one!
Many many years ago I happened to be in a conversation with one of the guys on a team that participated in the 2005 DARPA Grand Challenge. It was only the second such race after the 2004 one, but arguably the one which set off the autonomous driving race we see today. (Sebastian Thrun's team came in 2nd.)
I went into the conversation thinking it was going to be an extremely challenging but tractable sensors + control-systems problem. But by the end of the conversation I was like, OMG this is going to be a long-haul slog of solving an unending stream of problems, some potentially even AI-complete (i.e. requiring human-level judgment.)
We mostly discussed why his and most other teams failed and the failures were so myriad and so technically intractable that I could not see a path to full self-driving for at least two decades. And all of this was offroad, so it didn't even approach the challenges of sharing human-occupied streets. I cannot remember any details unfortunately, but I remember that one car got stuck in a loop due to a problem that would have been trivial for a human to bypass... but that required human-level judgment. As an analogy it was something like a soft obstacle that could safely be driven over. But for the car to know that it would require a database and an "understanding" of all possible obstacles. An LLM could have helped, but back then they were still firmly in the realm of SciFi.
So the only feasible solution was to painstakingly identify all the edge-cases and work through them slowly, carefully, one-by-one. Which is what Waymo has been doing. This is also why when Elon made his "full self-"driving announcements I knew he had absolutely NO idea what he was talking about, and he was likely going to move fast and break people.
Flooded streets is just another "bump on the road" to full self-driving, but it seems we're actually getting there now. In retrospect, my 2-decade estimate was surprisingly accurate, I have no idea how I landed on that particular number!