Autonomous Car Progress

[diplomat33]

Seems like Cruise is handling Chandler a lot better, based on JJ Rick's second impression:

https://twitter.com/x/status/1685504624219103232

 

[spacecoin]

https://twitter.com/x/status/1684582445340041217

Interesting stuff, and this survey is first of its kind afaik. Too bad they mixed the NA and EU responses.

 

[dloomis]

https://twitter.com/x/status/1684582445340041217

I’m in the middle of renting FSD Beta 11.3.6 for a month, and yes of course it’s “useful”. But even if version 12 is as mind blowing as Elon says it is there’s just no way it’s worth $15k for the average driver who otherwise doesn’t have a driving disability. The economics just don’t make sense.

 

[willow_hiller]

I’m in the middle of renting FSD Beta 11.3.6 for a month, and yes of course it’s “useful”. But even if version 12 is as mind blowing as Elon says it is there’s just no way it’s worth $15k for the average driver who otherwise doesn’t have a driving disability. The economics just don’t make sense.

You need to compare FSD Beta with offerings from other car makers to put things into perspective. For e.g. the Hyundai Kona doesn't get smart cruise control until you pay $8,000 for the package upgrade from SE to Limited. The package also includes a few things like leather seats and upgraded wheels, but if you don't care about those things and just want the ADAS, there's no way to opt for it without paying the full $8,000.

Personally, I think FSD Beta is easily worth 2x a limited trim package.

 

[diplomat33]

"Volkswagen Commercial Vehicles has begun testing a self-driving version of the ID. Buzz electric vehicle with Mobileye Drive technology on two continents. In Munich, Germany, and Austin, Texas, Volkswagen will validate the Mobileye Drive-equipped ID. Buzz vehicles with safety drivers on public roads, towards a goal of series production.

Mobileye Drive incorporates advanced EyeQ Systems-on-Chip (SoCs), as well as Mobileye’s sensing, mapping, and driving policy technologies, to create a unique, full-stack autonomous driving system that can adapt to new locations.

You can read more about Volkswagen's testing in Austin here and Munich here."

https://twitter.com/x/status/1686069463173513226

 

[diplomat33]

JJ Ricks 2nd Cruise ride in Chandler:

00:00 Hailing the car
00:45 Drive starts
01:10 Four way stop
02:02 Four way stop
02:50 Four way stop
03:30 Stop light
05:00 Four way stop
05:52 Much better routing, no more weird twists… ish
06:05 Four way stop left turn
06:53 Right turn on a speed bump
07:30 First bit of weird braking
08:40 Stop light
08:50 Teeny bit of a bumpy stop
10:33 I don’t quite understand the routing
11:05 Teeeeeny bit of shake on the right turn
11:20 Nice left there
11:40 Stop sign right
12:37 Odd stop sign encounter and twitching
13:15 Nice left turn
13:50 Left turn
14:15 Little bit of wheel twitch on speed bump

 

[ItsNotAboutTheMoney]

You need to compare FSD Beta with offerings from other car makers to put things into perspective. For e.g. the Hyundai Kona doesn't get smart cruise control until you pay $8,000 for the package upgrade from SE to Limited. The package also includes a few things like leather seats and upgraded wheels, but if you don't care about those things and just want the ADAS, there's no way to opt for it without paying the full $8,000.

Personally, I think FSD Beta is easily worth 2x a limited trim package.

+ SEL (+$4k)
10.25" touchscrreen (v 8.0")
Wired also Android Auto/CarPlay (v wireless only)
Wireless phone charger
Leather steering wheel
Digital key
Sunroof
+ Limited (+$4k)
LED headlights
LED taillights
Heated steering wheel
Leather trimmed seating surfaces
Ventilated front seats
Homelink auto-dimming mirror
Rain-sensing wipers
Rear backup parking warning
Highway driving assist (Level 2)
Auto high-beam

Nothing really fancy. Seems like you'd want SE or Limited. Limited is $41,550.
Previous Kona (I have it) had SEL, Limited and then a higher version with the ADAS stuff. As that tech has become normalized they moved down a package and the SEL has some of what the Limited used to have.

 

[hooty]

Take a quick step back from the polarized "us versus them" mentality:

On a scale from 0 to 100, where 0 represents a car driven entirely manually by a human, and 100 represents a car that can drive without a human in any place in any condition, do you honestly think FSD Beta represents a 0?

Because if so, then I don't think we have a shared basis of reality, which makes having a conversation almost impossible.

Elmo promised robotaxis. There won't be any robotaxis, no matter how much hopium one chugs

 

[willow_hiller]

Elmo promised robotaxis. There won't be any robotaxis, no matter how much hopium one chugs

All you had to say in response was: "I believe Tesla has made zero progress toward autonomy." But thanks for confirming that you're incapable of having a coherent discussion about this. You don't care about facts, you just hate Elon Musk.

 

[spacecoin]

All you had to say in response was: "I believe Tesla has made zero progress toward autonomy." But thanks for confirming that you're incapable of having a coherent discussion about this. You don't care about facts, you just hate Elon Musk.

If I may chime in.

An advanced ADAS lie FSDb is literately zero progress towards autonomy regardless of how impressive it is every once in a while. It's still an eyes-on, and as of now, a hands on system. You may get a drive without interventions here and there or "often". But quite frankly, it's not even at a point where 95% of the trips are without an issue, even in optimal conditions. You need a disengagement every 1000-5000 miles at least to call something "progress towards autonomy". To get there, you will need to limit the ODD and it still will not be anywhere near "eyes-off" safety levels.

An autonomous system has completely design criteria than "let's try to make it drive anywhere in any weather". "let's make it cheap", "let's remove radar and USS - we can make it good enough without it. oh wait let's add radar again, but only for S/X", "redundant power steering isn't important since it's covid right now" et.c.

An autonomous system designer would think "Hmm what failure modes do we need to handle", "Let's add sensor cleaning so we can drive more safely in some conditions", "Let's add a sensor that sees a boar, bison or a moose even if the cameras partially blinded ny oncoming cars". "Let's add directional mics so we know where the emergency vehicle if coming from.", and most importantly "how will we validate the system in it's operational design domain".

It's incredible to me that anyone still believes that HW3/HW4 will be autonomous. It's a Level 2 product on Level 2 hardware. Probably a very nice L2 in a few years time even - definitely one of the more capable, but nonetheless likely forever L2. No retrofits. "All cars that has all the necessary hardware needed for of Full Self Driving". Let that sink in. Turns out "Full Self Driving"(tm) is an ADAS and all other statements regarding "safer than a human" and "robotaxis" and Level 5 autonomy in 6-12 months for the better part of a decade are "forward looking". Didn't you read the fine print?

It doesn't help hope many times Elon says "Robotaxi" (he's really not doing that anymore) or "safer than a human". It's time to wake up from the dream, turn on some critical thinking and realise that he screwed us.

 

[willow_hiller]

If I may chime in.

An advanced ADAS lie FSDb is literately zero progress towards autonomy regardless of how impressive it is every once in a while. It's still an eyes-on, and as of now, a hands on system. You may get a drive without interventions here and there or "often". But quite frankly, it's not even at a point where 95% of the trips are without an issue, even in optimal conditions. You need a disengagement every 1000-5000 miles at least to call something "progress towards autonomy". To get there, you will need to limit the ODD and it still will not be anywhere near "eyes-off" safety levels.

An autonomous system has completely design criteria than "let's try to make it drive anywhere in any weather". "let's make it cheap", "let's remove radar and USS - we can make it good enough without it. oh wait let's add radar again, but only for S/X", "redundant power steering isn't important since it's covid right now" et.c.

An autonomous system designer would think "Hmm what failure modes do we need to handle", "Let's add sensor cleaning so we can drive more safely in some conditions", "Let's add a sensor that sees a boar, bison or a moose even if the cameras partially blinded ny oncoming cars". "Let's add directional mics so we know where the emergency vehicle if coming from.", and most importantly "how will we validate the system in it's operational design domain".

It's incredible to me that anyone still believes that HW3/HW4 will be autonomous. It's a Level 2 product on Level 2 hardware. Likely forever. Let that sink in.

It doesn't help hope many times Elon says "Robotaxi" (see he's really doing that anymore) or "safer than a human". It's time to wake up from the dream, turn on some critical thinking and realise that he screwed us.

It's very odd to me that you believe autonomous vehicles just pop into existence immediately achieving 1000-5000 miles between disengagements. Every AV company today started with safety drivers behind the wheel. And if they didn't focus on limited geo-fenced areas years ago, they would still all have safety drivers.

Level 3 is well within reach for HW3. We already know that Tesla is actively estimating the probability of disengagement during any given maneuver. As long as they're able to accurately estimate that figure, HW3 could drive autonomously, and the notify the driver and hand back control during low-confidence. And then the jump to Level 4 is just a matter of identifying the ODD under which the system has consistent high confidence.

And you don't need a "directional microphone" for level 3. If you want to exclude handing emergency vehicles from the ODD, you just have to identify the sight or sound of them in the first place.

 

[spacecoin]

It's very odd to me that you believe autonomous vehicles just pop into existence immediately achieving 1000-5000 miles between disengagements. Every AV company today started with safety drivers behind the wheel. And if they didn't focus on limited geo-fenced areas years ago, they would still all have safety drivers.

This argument is ridiculous. Literately everyone started at a test track, with the sole purpose of performing in a defined ODD. Take your example of MB Drive Pilot for example. Take Zoox as another:

Level 3 is well within reach for HW3. We already know that Tesla is actively estimating the probability of disengagement during any given maneuver. As long as they're able to accurately estimate that figure, HW3 could drive autonomously, and the notify the driver and hand back control during low-confidence. And then the jump to Level 4 is just a matter of identifying the ODD under which the system has consistent high confidence.

Is what ODD are you thinking HW3 would be autonomous?

And you don't need a "directional microphone" for level 3. Mercedes Benz Drive Pilot certainly doesn't have one:

They have at least one external microphone.

Mercedes-Benz's Drive Pilot L3 Is Super Smooth (but Limited) Hands-Free Driving

Mercedes' Drive Pilot tech is already offered on the S-Class and EQS sedans for road-legal use in Germany, but won’t make its U.S. debut until 2023.

www.thedrive.com

If you want to exclude handing emergency vehicles from the ODD, you just have to identify the sight or sound of them in the first place.

If you are going to provide time for a handover, you need a camera that sees it from behind even if there are cars right behind you, and you do need at least one external mic.

 

[willow_hiller]

If you are going to provide time for a handover, you need a camera that sees it from behind even if there are cars right behind you, and you do need at least one external mic.

You're really stretching right now. HW3 has 3 cameras that face rearward; the backup camera and both repeaters. That's more than sufficient to see emergency vehicles behind you. And although you've already changed the criteria from "directional microphone" to "external microphone," are you incapable of hearing a siren from inside your car?

And either way, this is a tangent from my original point and the topic of this thread. You and others are arguing that Tesla has learned nothing about autonomy from FSD Beta, and has made zero progress toward even Level 3. I frankly think that's preposterous.

 

[spacecoin]

You're really stretching right now. HW3 has 3 cameras that face rearward; the backup camera and both repeaters. That's more than sufficient to see emergency vehicles behind you.

Perhaps it will suffice in all situations. Perhaps not. Depends on the ODD… Hardly in city driving unless hw3 can bend light.

And although you've already changed the criteria from "directional microphone" to "external microphone," are you incapable of hearing a siren from inside your car?

Robotaxi have 360 sensors of all modalities, hence "directional".

It doesn't matter if the fallback drives hears the sirens. The ADS doesn't, if it has no mic input. You're stuck in an ADAS mindset.

And either way, this is a tangent from my original point and the topic of this thread. You and others are arguing that Tesla has learned nothing about autonomy from FSD Beta, and has made zero progress toward even Level 3. I frankly think that's preposterous.

I think you fail to understand what Level 3 is. There seems to be zero L3 design intent right now. Again, in what ODD do you imagine this happening?

 

[willow_hiller]

I think you fail to understand what Level 3 is. There seems to be zero L3 design intent right now. Again, in what ODD do you imagine this happening?

I think you're adding constraints into the SAE definitions that don't exist. Here's a concise definition for you:

"The sustained and ODD-specific performance by an ADS of the entire DDT, with the expectation that the human driver will be ready to respond to a request to intervene when issued by the ADS."

L3 will have, by definition, a limited and conditional ODD. But no where do I see that an automaker needs to have a fixed temporal or spatial ODD, or even that the ODD needs to be based on factors that would be evident to the human driver; just that it can inform the driver to intervene in a reasonable amount of time.

This is the internal measure of FSD Beta's probability of disengagement that I'm referring to:

https://twitter.com/x/status/1677764917494181888

Tesla could use those metrics to pick specific locations, times, weather conditions, etc. But they could also let the system learn and identify whatever contextual clues indicate the future inability for the system to drive safely, and hand control back over to the driver.

 

[spacecoin]

Again, in what ODD is “Level 3 well within reach” in your opinion? It was your statement so you must have an idea? Highway queue chauffeur in a MB-like ODD? More? Faster speeds? Other roads? Rain?

Also define what your idea of ”reasonable time” is please.

At present the system have no grasp over what it can and cannot do.

 

[willow_hiller]

At present the system have no grasp over what it can and cannot do.

This is false. I just showed you that Tesla is actively predicting moments where it expects the human driver to intervene.

I don't have this data, so I cannot define the ODD for them.

 

[spacecoin]

This is false. I just showed you that Tesla is actively predicting moments where it expects the human driver to intervene.

I don't have this data, so I cannot define the ODD for them.

But you can claim HW3 L3 is well within reach in some ODD? By what KPI or observations?

Please understand that an eyes off system need to let the driver start performing the OEDR before taking over the DDT, so at least 5-6 seconds I’d argue.

 

[willow_hiller]

But you can claim HW3 L3 is well within reach in some ODD?

Yes, the ODD defined by the data which shows the scenarios FSD Beta is incapable of driving in. These are private companies we're talking about; they're not going to share this data.

If you can find one crash metric or KPI for how Mercedes Benz defined their Level 3 ODD, I'll say you're right and that Tesla is years away from Level 3.

 

[spacecoin]

Yes, the ODD defined by the data which shows the scenarios FSD Beta is incapable of driving in. These are private companies we're talking about; they're not going to share this data.

If you can find one crash metric or KPI for how Mercedes Benz defined their Level 3 ODD, I'll say you're right and that Tesla is years away from Level 3.

So the approach is basically: Let’s develop a system that tries to do everything and then remove everything that we suck at and hope there is something that we do complete 99.999999% of the time? Defiine that as the ODD. Then take some duct tape and add hand over procedures and a bullet proof ux?

That sounds equally stupid and unlikely to me.

MB likely said, in order to not crash when the camera is blinded or if cv fails, add lidar. handle r157 emergency corridor? add mic. and so on. it was designed specifically to do what is does.