So when are you planning on moving to China?
I just don't see autonomous vehicles hitting state side anytime soon. It's perpetually 2-3 years away.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Autonomous Car Progress
- Thread starter stopcrazypp
- Start date
-
- Tags
- Autonomous Vehicles
So when are you planning on moving to China?
I just don't see autonomous vehicles hitting state side anytime soon. It's perpetually 2-3 years away.
diplomat33
Average guy who loves autonomous vehicles
Wall Street Journal has a good video on the current progress of autonomous driving:
My only criticism is that the video messes up the SAE levels. They define L4 as full autonomy that needs pre-loaded map data to work whereas L5 is full autonomy that does not need maps or pre-loaded. But the SAE says nothing about L4 needing maps or L5 not needing maps. Maps are not part of the definitions of the SAE levels. I think the WSJ is making assumptions because L4 currently uses HD maps.
My only criticism is that the video messes up the SAE levels. They define L4 as full autonomy that needs pre-loaded map data to work whereas L5 is full autonomy that does not need maps or pre-loaded. But the SAE says nothing about L4 needing maps or L5 not needing maps. Maps are not part of the definitions of the SAE levels. I think the WSJ is making assumptions because L4 currently uses HD maps.
Last edited:
If you have time, I think it would be a good idea to contact the paper or the reporter with the byline if there is contact info. WSJ should be pretty good about correcting the story. It's better to catch these things early otherwise this will get picked up and propagated all over the place.
diplomat33
Average guy who loves autonomous vehicles
I emailed the writers of the article and I already got a response. The person said that they have passed along my concern to the makers of the video.
diplomat33
Average guy who loves autonomous vehicles
I got the following response from the Senior Video Producer at the WSJ:
I disagree with the response that the video only gave examples of maps and did not make it a requirement. The video clearly said that L4 requires maps and data:
That screenshot is very misleading about the levels IMO. I get that the WSJ wants to make the complex topic of autonomous driving easier to understand for a general audience but I feel like there is a better way to do it without misrepresenting the levels.
powertoold
Active Member
There's widespread misunderstanding of the SAE levels. I blame the SAE for not making it clearer.
Although L4/5 don't explicitly require maps, maps are needed for navigation, so you need them to make the L4/L5 feature useful. Otherwise, you can't set a destination, so what's the point.
Although L4/5 don't explicitly require maps, maps are needed for navigation, so you need them to make the L4/L5 feature useful. Otherwise, you can't set a destination, so what's the point.
diplomat33
Average guy who loves autonomous vehicles
I don't blame the SAE. They put out documents that try to be very clear in the technical terms that they use. The define all the terms that they use. And the SAE J3016 even rejects terms like "self-driving" that are considered too imprecise. And the charts about the levels that they put out are clear. I blame "journalists" and layman who try to "dumb it down" for the public without even bothering to read the SAE document. The fact that the senior video producer for the WSJ apparently put out a video to explain the levels to the general public and does not care about the SAE, the authoritative body on the levels, makes my point.
All autonomous driving needs basic maps for navigation. So I don't think they are talking about navigation. I assume they are talking about HD maps for autonomy. And the WSJ says that L5 does not need any maps at all which does not fit your interpretation. How would L5 handle basic navigation without any maps at all? That does not make sense.
powertoold
Active Member
Their one-page graphic of the SAE levels could be clearer. It should say things like:
1) The levels don't define any performance or safety standards.
2) Levels apply to paved publicly-accessible roads only. (not sure if this is true, but it should make it clear)
3) Maps are required in order to complete trips.
4) Etc.
In fact, in J3016, it says:
A level 5 ADS-DV capable of operating on all mapped roads in the US that are navigable by a human driver.
So level 5 needs maps.
Edit: to correct myself, that statement above doesn't say that level 5 requires maps. It just says that L5 should work on all mapped roads.
Last edited:
Yes, it's almost common sense that any of these advanced assistance levels will typically work in conjunction with maps for navigation and/or recognition of traffic law context, even though maps are not explicitly required by SAE definition.
The only L4-specific use of maps that comes immediately to mind is if we say that L4 vs. L5 in a partucular comparison depends on geofenced operating-area restrictions. But I think this is not a general distinction. L4 can be L4 and not L5 because of weather conditions or fast-developing scenarios, not necessarily because of geofenced limits. And L5 can be considered L5 even if it's operating in a restricted area (I believe that you and others often describe Waymo in Chandler as L5 within its boundaries, correct?).
So, the distinction made in the WSJ video is incorrect, confusing and also an unnecessary (aside from wrong) detail, especially if the purpose was to simplify the Levels of Autonomy explanation.
Mr. Downs should have simply owned the issue in a professional way. It's perfectly understandable that it's late and expensive to update and properly fix the video at this point. He could have said that he understood the point and that he thinks the video is still useful but they will put an overlay errata caption at that scene (you see this quite often in YouTube videos), or at least into the accompanying story/description.
diplomat33
Average guy who loves autonomous vehicles
No. By definition, L5 cannot be geofenced. L5 is defined as having an unlimited ODD. L5 cannot have any design restrictions placed on it's ODD.
No, I've never said that Waymo is L5. Waymo is L4 because of the geofence.
OK thanks and I do agree based on the definition. I wasn't sure who's claimed what. I think that sometimes "people", in some of the more heated discussions, have claimed Waymo has L5. But maybe the main claim I'm thinking of is that Waymo is far closer to future L5 (than Tesla), and a common counter-claim is that the geofence restriction is actually a more serious hurdle than is Tesla's currently intervention-heavy L2, but which is usable and presumably improving throughout the country (-> Western world ->...->world).
I personally believe that L4, usable in one's usual geographic area of interest, is a major achievement. There will eventually be degrees or sub-categories of L4, while true L5, no backup support ever required, may remain elusive and/or unnecessary in practice.
diplomat33
Average guy who loves autonomous vehicles
Yeah, I do think Waymo is closer to L5 than Tesla is. Obviously, Waymo still has FSD problems to solve but a lot of the autonomous driving that Waymo already has would work in other areas. In fact, there are probably lots of areas in the US where if Waymo deployed today, they could do driverless.
IMO, geofencing is not the main obstacle to Waymo's expansion because creating HD maps is not a big deal. The main obstacles for Waymo are economical, logistical and safety. Waymo can't expand until they can be sure that the ride-hailing will be profitable. There are logistical hurdles with a large ride-hailing fleet. Managing a large fleet of robotaxis, with remote operators on standby if there is a problem, would be very difficult and expensive. Lastly, Waymo would need to make sure their FSD really is safe before doing a large expansion. Otherwise, there could be a large number of accidents that could force them to shut down. So Waymo's best approach IMO is to expand slowly in major metro areas. If Waymo can offer convenient, cheap, safe, driverless ride-hailing, even in just 2-3 major cities, that would be huge. And if the ride-hailing services are profitable, then Waymo can scale up.
Tesla already makes a profit just by selling the car to the consumer. Tesla does not need ride-hailing to make a profit. And if Tesla gets FSD, they can offer it direct to consumers to use in their personal cars. So Tesla is not really in the ride-hailing business. And Tesla already has a large fleet of cars on public roads. Tesla's main obstacle is technological. Tesla can offer L2 features and give value to their customers. But if they want to offer true autonomous driving (no driver needed), then they need to "solve FSD". And for Tesla, "solving FSD" is much harder than for Waymo. That's because it is much harder to solve all of FSD problems everywhere than it is to just solve FSD in a smaller geofenced area. Smaller areas have fewer edge cases.
Definitely. Geofenced L4 is still a major accomplishment. That's because achieving driverless even in a relatively small area, is still hugely difficult. There is a huge leap between autonomy with a driver that can take over if needed and autonomy where there is no driver in the car at all and the car is truly on its own.
Personally, I view L5 a bit like an asymptote. It is something we will get closer and closer to without ever touching. That's because L4 is a limited ODD. So as we remove limits and expand the ODD, the L4 will get closer to L5. But there will still be "holes" in the ODD that we can't do yet, keeping the L4 just shy of true L5. But eventually, I think we might get so close to L5 that we will consider it close enough. Maybe there is a hole in the ODD that the L4 can't do, but if the L4 can be used for like 99.999% of driving, then I think for all intents and purposes, most people would consider it to be L5.
Your post summarizes the state of things very well.
Tesla is trying to solve FSD in a much more general sense with more constraints - Vision only (or primarily at least), lower net cost for sensors so cars are still profitable to sell to consumers, and no geofencing. This is both admirable but also make it MUCH harder to achieve L4. But that's okay, because as you pointed out, Tesla is profitable today based on their car sales. While their sky high evaluation may be hurt if they can't live up to the robotaxi hype from Musk before Waymo and others are already established, Tesla will still continue to be relevant and very profitable.
Their FSD approach is definitely much higher risk and much more technically challenging. The flip side of this is that IF they do achieve L4 with their approach on a similar timescale as Waymo and others, then it is game-over for their competition and they will dominate the market because they will just have an insane edge over the competition (a huge fleet already paid for by consumers and ready to serve as robotaxis). But that is a huge "if" and as someone who works in the field of computer-vision/ML, I think a more robust sensor suite has an edge over a vision-only system at this point in time.
For Waymo, they are already clearly demonstrating that with more constraints (geofencing, HD Maps, more extensive and expensive sensor suite, remote operators on standby, etc.) they are much closer to achieving L4. This is admirable and they certainly have a lead over Tesla if the metric is can you achieve L4 within a smaller subset of areas. The challenge for them on the other hand is that scaling their solution up to be general enough in the Tesla sense of general FSD everywhere is still going to be incredibly challenging and also result in a lot of logistical challenges as you mentioned in your post.
I think Tesla's approach is the right approach for them. They should be focused on making exceptional EVs that are affordable and that provide best-in-class driver assists / FSD features while remaining profitable and growing their market share. Waymo's approach is also the right approach for them.
The problem really is all the hype Musk has been creating around their FSD for years now with not a whole lot to back it up. I wish he would just shut up about all things FSD and AI and let pragmatists like Andrej Karpathy who actually know what they are talking about and understand the state of things do all the talking.
powertoold
Active Member
Based on what we've seen in JJRick's Waymo videos, it's clear to me that Waymo is years away from level 4 that encompasses an entire area. The current level 4 service in Chandler has a very limited geofence. And even with in this limited geofence, which they constantly make even smaller by restricting routes and avoiding problematic roads and maneuvers, I'm estimating that the disengagement rate is once every ~1000 miles (JJRicks has had 4 in less than 1500 miles).
My guess is that a lot of the execs recently left because Waymo won't be able to engineer themselves out of this.
Currently, the Waymo geofence EXcludes (sorry if I'm wrong on some of these):
1) Certain streets / roads / routes
2) Certain intersections
3) Unprotected lefts through cross traffic
4) Any rain / wet roads
5) Highways
6) Certain pickup and drop-off spots
Saying Waymo is closer to level 5 than Tesla is premature, as Waymo is very very very far away from level 5.
I doubt Waymo will ever have an unrestricted geofence (include all roads, maneuvers, and weather conditions within the 50mile^2) in the Chandler area, let alone San Francisco.
I'm also sure that Waymo's goal is not level 5, at least not anymore.
My guess is that a lot of the execs recently left because Waymo won't be able to engineer themselves out of this.
Currently, the Waymo geofence EXcludes (sorry if I'm wrong on some of these):
1) Certain streets / roads / routes
2) Certain intersections
3) Unprotected lefts through cross traffic
4) Any rain / wet roads
5) Highways
6) Certain pickup and drop-off spots
Saying Waymo is closer to level 5 than Tesla is premature, as Waymo is very very very far away from level 5.
I doubt Waymo will ever have an unrestricted geofence (include all roads, maneuvers, and weather conditions within the 50mile^2) in the Chandler area, let alone San Francisco.
I'm also sure that Waymo's goal is not level 5, at least not anymore.
Last edited:
That kind of depends on what the cause of the disengagements is, whether there are common patterns that are solvable, etc. It could be one month or twenty years. Without a more detailed breakdown, it's anybody's guess.
The question, of course, becomes why those specific roads and intersections are excluded. Is it because of schools (safety concerns)? Construction? Because the previously obtained HD maps are no longer accurate enough? Because of bad road conditions (potholes, etc.)? Too many people pulling out of parking lots without looking? Too many delays because of pedestrians/slow traffic/badly timed traffic lights? Something else entirely?
I'm pretty sure Waymo does unprotected left turns. They were definitely doing them in demos in Phoenix well over two years ago. It may try to minimize them, though (as should any good driver, really).
That one, I suspect, is mostly the fault of LIDAR, whose performance degrades substantially in inclement weather. That and lawyers.
That's presumably because there are only a couple of highway exits within that tiny service area.
Okay, I just looked, and it's more like four or five within Chandler, all on the 202 loop. The point is that there's not much point in getting on the highway to go such a short distance.Also, if you've driven on the highways in Phoenix... I figure one engineer realized that the cars would get run over if they didn't go 20 over the limit, and that engineer's manager decided that there was too much legal liability if they got pulled over for going 20 over in an autonomous car, and that was the end of highway driving.
And that again becomes a question of why. Too many people? Too many drivers not paying attention? Something broken in the path planning? Inaccurate/incomplete HD map data? Incorrectly tagged HD map data? Because the car won't fit on the escalator in a shopping mall?
Waymo did 30,000 miles per disengagement in California last year. So if they're really disengaging 30x as often in Arizona, there must be some huge difference either in the roads, the quality of the map data, or the competence/behavior of other drivers/pedestrians. I'd be curious to know what that difference is.
Barklikeadog
Active Member
diplomat33
Average guy who loves autonomous vehicles
Beware the small sample fallacy. You can't judge all of Waymo's performance based on only one car. JJ Ricks has done less than 2000 miles. If you look at Waymo's CA DMV report, there is a car with VIN 2C4RC1K74HR805516 that had a disengagement rate of 1 per 1,688 miles similar to what you think JJ has experienced. If we just judged Waymo based on that one car, we might also assume that Waymo's disengagement rate is approximately 1 per 1,700 miles. But if we look at the entire CA DMV disengagement report, we see that Waymo also has cars with disengagement rates of 1 per 10,000 miles. And some cars did thousands of miles with zero disengagements. Over a bigger sample of 600,000 miles, Waymo's average disengagement rate is 1 per 30,000 miles.
You are making an assumption that Waymo can't solve these engineering problems. You don't know that. Waymo has some of the best FSD engineers on the planet. There is no reason to think that they can't solve these problems.
You are definitely wrong about #3. Waymo does unprotected left turns through cross traffic all the time in JJ's videos.
Waymo is very far from public commercial deployment as L5 since they have only deployed commercial ride-hailing in a small geofenced area so far. But Waymo's tech could be closer to L5 than you think. The driverless rides only represent where Waymo is confident the Driver has achieved 99.9999% reliability. But for all we know, the Waymo Driver could be 99.9% reliable outside the geofenced area in Chandler, not good enough for driverless, but still pretty good. We know Waymo is testing in other areas with safety drivers with good disengagement rate (See CA DMV report). So the Waymo Driver can handle more situations than just what we see in Chandler. I am not saying that Waymo's tech is almost L5. But it is not "very very far away" from L5 IMO.
I think it will be really telling when Waymo launches their ride-hailing with safety drivers in SF with the 5th Gen i-Pace. We will see how it performs in that new area that is different from Chandler. And it will be our first look of the new 5th gen that is better than the 4th gen. I think it will give us another important data point of how close or how far Waymo is to L5.
Sure they will. Why not? The Waymo Driver is constantly improving. I am sure eventually, they will be able to do driverless in an unrestricted geofence.
Last edited:
powertoold
Active Member
Have we found any marketing or documentation that Waymo is aiming for level 5?
In this recent video, the new Waymo CEO says:
"We just simply think level 4 is the goal for any domain." (At 2:15)
I agree with Waymo actually. A true level 5 system will be prone to abuse by the public.
It's possible that a company will achieve a level 5 feature in the future, but I doubt they'll deploy it for public use.
In this recent video, the new Waymo CEO says:
"We just simply think level 4 is the goal for any domain." (At 2:15)
I agree with Waymo actually. A true level 5 system will be prone to abuse by the public.
It's possible that a company will achieve a level 5 feature in the future, but I doubt they'll deploy it for public use.
diplomat33
Average guy who loves autonomous vehicles
I don't think Waymo is aiming for L5 deployment. Waymo is strictly aiming for L4 deployment because it matches their commercial goals. But Waymo is not interested in only solving FSD problems in Chandler. They are trying to solve as many FSD problems as possible. As Waymo solves more and more FSD problems, the Waymo Driver will get closer and closer to L5. So in that sense, Waymo is trying to "solve L5". But we need to make a distinction between deployment and capability. A system can be capable of L5 but be deployed as only L4. For example, I can geofence a L5 system and it becomes L4. So, Waymo could be trying to "solve L5" but only be interested in deploying as L4.
Hope that makes sense. Basically, I am suggesting that Waymo is "aiming for L5 capability" but only interested in "L4 deployment".
Last edited:
powertoold
Active Member
Well, based on that, anyone improving their level 4 is getting closer to level 5. If we use that logic, then Waymo is closer to level 5 than Tesla is.
But again, I don't think Waymo's goal is level 5, whereas Tesla is explicitly aiming for it.
It seems Waymo's goal is to restrict a level 4 service as much as possible such that it's still useful, while meeting their safety and reliability thresholds.
Similar threads
- Article
