-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Waymo One launches
- Thread starter strangecosmos
- Start date
What is useful depends on the person.
I think the visible progress is mostly irrelevant due to (what I see) as a big difference between Tesla and others in that for some, TaaS is the product, for some driver assist is a feature that is more or less locked in terms or functionality. Tesla is weird in that they sell cars and are also working on the self driving features that they will be able to roll out to the fleet when ready. HW 2.x will need updated, but 3.0 with be out soon and hopefully sufficient.
My point (if I have one)? Tesla can play the long game of a generalized or all encompassing solution because they don't need to hit X level of progress to generate revenue, they can keep selling cars (with or without EAP) and upgrade them once the hit their milestones (of the cars not hitting milestones).
Other companies must have a certain level of functionality to have a saleable product and so focus on certain driving aspects/ regions to get good enough to use regardless of whether that solution will work on the wider range of environments.
Just a peanut in the gallery rooting for anyone working to solve a problem.
CarlK
Active Member
CarlK
Active Member
Don't know what you're trying to prove. Unless you're superman or you have camera eyes you have just so much instantaneous dynamic range like everyone else. Read the first few paragraphs of the article I linked to understand why you think you can do better than it actually is. When you say to "look around" you're not taking a shot but you're taking a video. Cameras can do even better when combining several shots too (for example in HDR photography).
Last edited:
Google HDR and the camera parameters. Learn about iso, aperture, shutter speed and the digital sensor. Then you may understand my message.
Almost every new electric car company is going with this mantra so not really wierd anymore. Before? Yes.
Here's the thing. If Tesla are still struggling with highway autonomy today how will they acheive L5 when surface street autonomy is orders of magnitude more difficult.
Elon reminds me of trump in that he can say anything, double down on it and his fans will still cheer him on and defend him. Elon knows he wont have level 5 in 2019, he's not crazy. But it won't stop him from building up hype and making money off the idea.
We are 18 days away from 2019. For L5 to actually happen in 2019. You have to believe in a fairytale magical improvement that Trent believes in.
Actually it's the opposite. It's the car manufacturer who dont depend on features. They are content to doing the same thing they have done over and over again because they can still sell cars regardless. They have shown this with adas and EVs.
Well point is to have a set goal and a game plan you can actually execute and accomplish that can provide the best use and save lives. Level 3/4 highway is a doable goal. Level 5 in 2019 is a fairytale and impossible. Eventually everyone will be working on a all encompassing solution.
When HW3 gets here (by june 2019). Elon will only have 18 months till 2021. The real question is, can Tesla even have a good highway system by then?
Uber was creating a self driving product. Waymo's car is a self driving product. Mobileye is a self driving product (that goes into OEM cars).
Flip that around, if Tesla is trying to develop a solution that includes surface street autonomy, how good do you expect highway driving to be in the interim?
Also, we only get to see what the system downsized to HW 2.x can do. That must be a subset of the 3.0 intent capability.
'In 2019' is different from 'by 2019', so having 18 days left in 2018 is sort of meaningless. I'm also still unclear on what recent statements you base the L5 claim on.
Existing OEMs, sure. But look at what they buy, Supercruise is a product that achieve a usability level in a set of conditions that reached the point of being includable.
Again, why the level 5 love? We both know FSD shouldn't be taking muddy 2 tracks through the woods, even if people can...
To make the lives saved argument, you'll need the full development timeline, lives saved at each step, and take the integral.
Are you making a level playing field? In that view, isn't MobilEye in the wrong for not supplying Tesla with whatever capability they have to increase safety while Tesla works toward a higher numbered replacement? Along with ME supplying parts at cost to everyone to save lives?
Why is that the real question? Based in my senario, the goal is good everywhere, including highway.
However, to answer that question: 2021, highway, HW 3 (maybe higher due to fab improvements) I'd say yeah.
My personal feel is that right now they are trying to squeeze a 3.0 NN in a 2.x box .
Neither of those are automaker. The new automakers are using the same approach.
I like that you posted that question because it means their quality and performance ties in with the potential of their FSD. Meaning we can gage their overall FSD progress by looking at NAP. Which right now currently sucks.
This exact same excuse is used every time though and will be reused in 6 months as "they just installed HW3, give them time to collect data". We saw this excuse throughout the 2 years of HW2 development.
Its different in the sense of the psychological influence of 'this year' vs ' next year'
Not sure what you are getting at here.
A system that fails every ~10-100 miles isnt saving any lifes. A system that is substantially better at driving than a human driver is!
Nope the goal is substantially better than a human driver. Not a good ADAS system. Unless you take thay leap at eliminating the driver, you will never improve safety at substantial and impactful levels
CarlK
Active Member
Why don't you just tell me what I said that you disagree? I don't expect you could though. Did you even read what I said and that Cambridge aritcle about comparison of how our eyes and cameara work? BTW I've been shooting SLR and DSLR for decades. It's a lot more than aperture and shutter speed.
Last edited:
I don't see that. HW 2 was a complete (rushed due to ME breakup) reboot from AP1, HW 2.5 was a minor change to HW 2.0. This is the first time Tesla has had a fully developed HW set in parallel to the current one that has had significant development time/ resources allocated to it pre-launch. They already have HW3 to use and test (with an internal test fleet) and the hardware itself was designed around the NN architecture they have been developing on simulated HW.
That ME sales depend on having a product with seme useful feature set to the OEM who is buying it. No viable use cases, no sales.
Then the question is not of a good highway system (what you wrote that I responded to).
Broader scope, a timid FSD system that runs 100% of the time could still be safer than a human who is normally safe, but has lapses in attention.
I was thinking about highway driving on my non-highway commute this morning and realized that highway only system is a myth of sorts.
Q: Why limited access highway/ freeway vs general driving?
A: Because it is theoretically simpler.
Q: Why is it deemed simpler?
A: Less use cases.
But is that true?
A great highway driving system needs to handle the rare as well as the typical, that means having the same level of detection/ reaction as a general city system.
General -> pedestrians
Highway, broken down car -> pedestrians
General -> construction
Highway -> construction
General -> emergency vehicles
Highway -> emergency vehicles
General -> traffic lights
Highway -> lane usage traffic lights
General -> child running into street
Highway -> animal running into street
General -> broken down car blocking lane
Highway -> broken down car blocking lane
General -> bad drivers
Highway -> bad drivers
A lower rate of occurrence of a situation does not remove the requirement to handle the situation.
FloridaJohn
Member
This back and forth really does get tiresome to read. To get this thread back on track:
Ars Technica appears to have found an actual paying customer of Waymo One that is able to talk about the ride:
We finally talked to an actual Waymo passenger—here’s what he told us
Sounds intriguing, but still in the early stages. Short quote from the article:
Ars Technica appears to have found an actual paying customer of Waymo One that is able to talk about the ride:
We finally talked to an actual Waymo passenger—here’s what he told us
Sounds intriguing, but still in the early stages. Short quote from the article:
Then on Wednesday morning, a Phoenix-area technologist and entrepreneur named Michael Richardson told me that he'd been released from his NDA. He was willing to talk about his experience as a Waymo customer.
"I'm impressed by what the vehicle can do and how well it gets around," Richardson told me in a Wednesday phone interview. "It's very, very impressive."
Still, he said, the service had significant limitations—including a limited coverage map and too few drop-off locations. Due in part to those limitations, Richardson has only taken four Waymo rides—two round trips—in the three months he's been part of Waymo's program.
All of his rides had safety drivers, and he said he saw them take control of the vehicle at least once over the course of those four rides. He also said that Waymo vehicles sometimes seem to take less-direct routes to avoid tricky situations like left turns or freeway driving.
"I'm impressed by what the vehicle can do and how well it gets around," Richardson told me in a Wednesday phone interview. "It's very, very impressive."
Still, he said, the service had significant limitations—including a limited coverage map and too few drop-off locations. Due in part to those limitations, Richardson has only taken four Waymo rides—two round trips—in the three months he's been part of Waymo's program.
All of his rides had safety drivers, and he said he saw them take control of the vehicle at least once over the course of those four rides. He also said that Waymo vehicles sometimes seem to take less-direct routes to avoid tricky situations like left turns or freeway driving.
Tam
Well-Known Member
It's interesting that Waymo does not like freeways but Tesla welcomes freeways.
Why the difference?
just guesses here:
- maybe they need more city miles for training
- maybe the software is not fast enough for higher speeds (Tesla isn't fast enough as well otherwise it would not have missed stopped cars)
- or maybe the highway markings / road quality is not that great there. One company said, they hate 101 in the Bay Area because of the road quality. Markings are not clear and the car is bouncing a lot which needs a lot of computing (adjustment of coordinate system)
Last edited:
1) There is no such thing as a highway only taxi service.
2) Waymo is going straight to level 5. Death is less likely at slower speeds. You don't learn to drive a motorcycle on the expressway.
So what exactly (in details) according to you should we expect in 6 months when HW3 gets here?
what do you mean by run 100% of the time?
I'm so glad that you made this post. It's so thrilling! It actually backs up what I have been trying to explain all this time. So It's actually interesting that Trent liked it. Shows you that facts are sometimes rejected/accepted based on if its presented as a support to your position or not.
Yes a level 3/4 high-speed highway only system for example will require the same level of vision capability as a surface street FSD system. Which is what I have been saying about mobileye eyeq4 having solved vision for sdc and is capable of supporting level 3 and level 4 SDC.
Trent rejects this notion and claims that Tesla has always been ahead throughout their development even when they struggled to match AP1 performance, even when they rebooted their NN platform several times as admitted by you, even when we have seen windows into their NN vision progress and improvements over the years. He still insists that Tesla is far far ahead in vision and will not concede that fact.
But as you have just proven, you need just as capable and reliable vision.
You are 100% right "A lower rate of occurrence of a situation does not remove the requirement to handle the situation.".
Its not that these system won't be able to handle these situations, it would be that they will be locked out of these situations to guarantee a higher level of safety. For example a highway only system that supports construction and fails every 100k miles, but when restricted from working in construction zones only fail after every 1,000,000 miles. However the ability to work in constructions will enable it to achieve a minimal risk condition upon operational design domain (ODD) exit (aka it wonders into a construction zone).
The difference however between freeway and surface street is the axis of movements and density which increases exponentially the planning involved (which you call reaction).
In the highway, you have actors moving in one direction, no intersecting, with each in their respective lane, this includes pedestrians stopped on the side of the road. Anything that gets infront of your lane, you either slow down or move over slightly.
In the surface streets however, you have a vast density of actors all moving and coming from on all axis (forward, backwards, left and right) and all intersecting. And they are virtually unpredictable. This requires far greater level of planning and maneuvering than on the highway
Highway
Surface Streets
Last edited:
Beats the heck out of me. I'd expect a big improvement (although it may be offset from HW3 roll out to validate on larger pool)
That humans tend to screw up due to distraction, so a FSD system that doesn't get distracted could be better, even if its driving is sub optimal. ie: slow to accelerate, but never rear ends a car in bumper to bumper traffic.
I get what you are saying, but a highway is still not to the level of big sky/ airplane environment where you don't need mirrors. Consider the cases of wrong way drivers, cars reversing to a missed exit, illegal (or even authorized) use of median U turns, car broken down in middle lane, tow truck going to it, passengers getting out...
An ideal system would not rely on prediction for funcationality beyond the necessity to assume cars are not trying to hit them (otherwise a car would be so paranoid, it would never move, or get within a lane of anything...).
Density is also a must-have requirement, although, near cars tend to block far cars in terms of interaction.
diplomat33
Average guy who loves autonomous vehicles
I would think it is fairly obvious though? Waymo is Level 4/5 car/prototype with driver merely as a backup, Tesla is Level 2 with a driver in charge.
Freeway is much easier to solve on Level 2 because the traffic scenario is fairly simple so it is the place to start when working towards self-driving. That’s why Tesla and so many others started there of course. But it also means getting to the freeway and joining the freeway is done manually. And a human is in charge if things go wrong as they can go wrong quickly at speed.
Waymo on the other hand started with a car that can go from start to finish autonomously even without a driver. This means they have to solve the hard problems first anyway. That’s why they have started by solving city streets because they can not be avoided unlike freeways when going from start to finish...
Joining freeway autonomously is not easy and any accident on the freeway is much more dangerous. It makes sense to start at lower speeds when working on a car that drivers itself all the way from start to finish.
I am confident a Waymo is better on the freeway than a Tesla. That doesn’t mean joining the freeway and handling the whole process fully autonomously is still something that would be mature even for Waymo...
If it needs a backup, it is not level 5, and if the driver is needed at unexpected points (vs known geofenced areas), it is not level 4.
That was hardly my point though. As I said, a prototype of a Level 4/5 car. We know it can self-drive so certainly Waymo’s are Level 4 cars and I would say aim to be Level 5 hence I framed it as Level 4/5 car/prototype.
But my point was a Level 4/5 prototype going from start to finish can not avoid city streets but it can avoid freeways that are hard to join and where speed kills. Hence it makes sense to avoid them if needed.
A Level 2 car on the other hand likes freeways for their simplicity because it does not have to worry about self-driving to get there or about being in charge. So it makes sense a Level 2 car avoids city streets and ”likes” freeways.
Similar threads
