Autonomous Car Progress

[diplomat33]

I was thinking about this yesterday. It rained a few days ago in SoCal and my car is filthy - gonna get a wash today. I was sitting at red lights, and the ultrasonics were popping up like the car is going into parking mode, with the line showing orange, as if I'm too close. However, I was more than a car length away from the car ahead of me. Then the USS line went away - then it came back, over and over until the light turned green. My thought is that the front sensors are really dirty, causing false responses. Oddly it was the only time that happened during the drive. I'm also wondering if it was interference from the lead car's own USS, like I was picking up their sonics as a return.

If an L3 car is driving around and the sensors are really dirty, how will that affect performance? In my case, it was transient, so the system may still engage but then suddenly think it's going to hit something.

With L3, the system would likely notify the driver to take over. L4 has to handle the situation on its own, like pulling over if performance is too degraded. I know Waymo cars have self-cleaning systems for their sensors so the car can keep the sensors clean while driving to deal with this type of situation.

 

[rbt123]

You could still have consumer ownership. But I think the owner would need to sign an agreement stating that certain actions would void the manufacturer's liability. So if the owner modifies the car in certain ways that could affect the self-driving, uses unauthorized repair shop or fails to do basic maintenance, then the owner would assume liability.

In most US states that works on a legal level; though the publicity of after-market modifications causing crashes would still be a burden.

On a global scale that doesn't work legally though; you cannot dictate restrictions or remove features after product sale in many places. You can void a warranty, but even then only for components covered by the warranty (an after-market door replacement cannot void a motor warranty).

Again, a lease or exclusive unlimited service is different because ownership is not transferred.

 

[diplomat33]

In most US states that works on a legal level; though the publicity of after-market modifications causing crashes would still be a burden.

On a global scale that doesn't work legally though; you cannot dictate restrictions or remove features after product sale in many places. You can void a warranty, but even then only for components covered by the warranty (an after-market door replacement cannot void a motor warranty)

Thanks. That is helpful. I just think the manufacturer needs an out if the owner does something that compromised the self-driving system. As long as the self-driving system is operating with the design specs set by the manufacturer, I think the manufacturer should be liable for accidents caused by the system. But I don't think you can reasonably expect the manufacturer to still accept liability for an accident if the owner screwed up the self-driving system. Certainly, I could see the manufacturer making a strong case in court that they should not be held liable because the owner damaged or changed the system and therefore the system was not operating as it was designed to.

 

[diplomat33]

In honor of Mobileye's IPO, they released this short PR video of their history. The video highlights key moments in Mobileye's history that I find pretty helpful to see their progress to-date:

 

[DanCar]

Cruise paid ride hail program in SF. 20 miles per day during june - Aug. 1,807 miles.

https://i.redd.it/odme9s60zfw91.jpg

https://www.reddit.com/r/SelfDrivingCars/comments/yf8izl

 

[EVNow]

Mobileye has promised much before and not delivered.

Don't you know - only Tesla's promises can't be taken seriously. Everyone else will always come through

 

[diplomat33]

https://twitter.com/x/status/1586058151773835264

 

[DanCar]

Aurora says they are hiring and have 160 open positions, focusing on trucks.

Aurora CEO Chris Urmson weighs in on Argo AI shutdown - The Robot Report

Ford and VW announced that the companies would be ending their investment in Argo AI and that the company would be shutting its doors.

www.therobotreport.com

 

[Goose66]

From a system design/development standpoint, there’s a BIG jump from L2 to L3, and another HUGE jump from L3 to L4. L4 is basically the same design/development-wise as L5 with a limited operational domain. “Know your limitations.”

Tesla’s current design and development efforts appear to me to be attempting to bridge the jump from L2 to L3.

 

[powertoold]

From a system design/development standpoint, there’s a BIG jump from L2 to L3, and another HUGE jump from L3 to L4. L4 is basically the same design/development-wise as L5 with a limited operational domain. “Know your limitations.”

No, there's not.

The levels are so stupid. They're a constant source of poor logic in AV discussions. People often use the levels to convey progress, difficulty, or performance, but they aren't meant for that.

Every time I see any mention of levels, I know it's going to be a stupid discussion that goes nowhere.

 

[SCTes1aMan]

I thought some may find this interesting. Lex Fridman talks with Andrej Karpathy, the former director of AI at Tesla.

https://twitter.com/x/status/1586422266354614278

 

[DanCar]

Sounds silly to me. All sensors are noise? Drop cameras then?

 

[SCTes1aMan]

Sounds silly to me. All sensors are noise? Drop cameras then?

That definitely caught my attention. It's going to be fun to look back on all of this in 4-5 years.

 

[DanCar]

1:39:30 into the video, Lex asks Andrej what is the timeline for Autonomous driving. Andrej says no idea.

1:45:00 into the video. Why leave Tesla. Andrej said he was stuck in an executive role and not what he enjoys. Prefers to work on technical stuff. Wants to learn and teach.
1:53:00 - 1:54:00 - Tesla Bot: Andrej: it currently thinks it is a car.

 

[powertoold]

That definitely caught my attention. It's going to be fun to look back on all of this in 4-5 years.

In that time, we'll look back and see how obvious it all was, and how much of a scam most FSD companies were lol

Waymo and Cruise are driven by marketing and pleasing their bosses / sources of $$$, not real progress towards scalable FSD

Waymo opened their public service in Chandler, and to do so, they required constant remote surveillance along with roadside assistance. You have to wonder: why start a service that requires so much $$$ to operate? It's all about the marketing.

 

[diplomat33]

No, there's not.

The levels are so stupid. They're a constant source of poor logic in AV discussions. People often use the levels to convey progress, difficulty, or performance, but they aren't meant for that.

Every time I see any mention of levels, I know it's going to be a stupid discussion that goes nowhere.

In that time, we'll look back and see how obvious it all was, and how much of a scam most FSD companies were lol

Waymo and Cruise are driven by marketing and pleasing their bosses / sources of $$$, not real progress towards scalable FSD

Waymo opened their public service in Chandler, and to do so, they required constant remote surveillance along with roadside assistance. You have to wonder: why start a service that requires so much $$$ to operate? It's all about the marketing.

Do not speak of things you know nothing about.

 

[DanCar]

...
Waymo and Cruise are driven by marketing ...

Agree. Example: how many autonomous miles do they drive customers? Recently learned that Cruise drove customers 20 miles per day in June through August.

 

[ThomasD]

Has Cruise removed the safety driver from the Walmart deliveries? Are they able to deliver to the entire Tempe and Chandler area? Or are they restricted to just certain roads and neighborhoods. This would be perfect for Thanksgiving and Christmas deliveries. The week before Christmas you could have your Christmas shopping done for you and delivered within a 2 hour window. You would not have to fight crowds.

 

[diplomat33]

Agree. Example: how many autonomous miles do they drive customers? Recently learned that Cruise drove customers 20 miles per day in June through August.

All companies are driven by marketing to some degree. Tesla sold FSD in 2016 based on nothing but marketing. But to suggest that companies like Waymo and Cruise are just marketing scams is completely wrong. They have real autonomous driving. They do serious ML and solving real autonomous driving problems. They are scaling their AVs gradually as their autonomous driving matures. Cruise only drove customers 20 miles per day in June through August because their ride-hailing is still very early, not because it is a "scam". Every company starts small at first. And this is especially true when rolling out new tech, especially safety critical tech like autonomous driving. The number of AV miles will increase over time.

 

[Goose66]

No, there's not.

The levels are so stupid. They're a constant source of poor logic in AV discussions. People often use the levels to convey progress, difficulty, or performance, but they aren't meant for that.

Every time I see any mention of levels, I know it's going to be a stupid discussion that goes nowhere.

I disagree. You can certainly design an L2 system with no thoughts of it evolving to L3. In fact, high-end cars have had L2 systems for a decade. They didn’t design those systems to ever operate without supervision. They lack integrated navigation, interaction with stop lights, stop signs, or other signaling, lane selection algorithms, intersection traversal, and the like. Developing a system intended to include all these capabilities in order to control the full driving function is a completely different undertaking than an L2 system that simply controls speed and steering.

Similarly, an L4 system is built to not require a driver at all and operate independently within a limited operating domain. It can’t just throw up a “Take Control Immediately” warning every time something comes back from a neural network outside of a threshold confidence level. It has to be able to handle all driving functions and fallback by itself. Accordingly, we see systems built from the ground up to be L4 having far more, redundant sensors, rely on
high resolution mapping and other externally supplied data, contain more computing power, and often include communication capabilities for remote monitoring and control.

On the other hand, L4 to L5 is just an expansion of the ODD in which the car is capable of handling the driving function. This may be an expanded geographic area — everybody wants to think of the limitations as geofencing — but could also be expanded weather conditions, road types, speed, traffic conditions, whatever limitations exist on the L4 system. But the design, programming, and hardware is basically all the same. A particular system moving from L4 to L5 over time seems far more like a natural evolution than L2 to L3 or L3 to L4.