Self Driving and Waymo

[brobro777]

Bro Waymo is everywhere is Los Angeles, and increasing all the time it seems

 

I heard it's popular with women who don't like skeevy Uber drivers

 

[Loss Horizon]

On 8/29/2024 at 10:45 PM, Saluki said:

While Waymo is still by far the leader in total miles driven, it looks like Zook, by Amazon, is leading in "miles per disengagement."  That's the number of miles that you drive before a human has to touch the wheel. 

 

https://thelastdriverlicenseholder.com/2024/02/03/2023-disengagement-reports-from-california/

 

Is anyone familiar with Zook?  Are they testing on the highways (easier than a city like San Francisco) or are they racking up miles in a place like Arizona?  It seems odd that one company would be able to perform so much better than the others if it was working with a lot fewer cars, and thus fewer data points.  

 

It feels like "miles driven" and "miles per disengagement" are metrics which shouldn't be used with autonomous driving.

 

I can push a cart downhill and it will drive autonomously for a few seconds until it crashed into the ditch. If I'm obsessed enough I can do that continuously for long time and clock many "miles driven autonomously" and "hours driven autonomously". It will be good numbers, but they won't tell anything about quality of the product.

 

"Miles per disengagement" is scary as hell. If I need to take over an automatic driving system every 2 minutes, it would be easy, because I stay engaged myself. If a system makes a mistake once a year, I will be happily sleeping by the moment it drives me off a cliff.

 

Those metrics assume too much of good faith from the companies who develop the systems. You would think they are all highly skilled, motivated and honest people, so those indirect metrics indicate real progress, and if something goes seriously wrong then somebody would blow the whistle. I hope so. They could also be run by deceitful and greedy entrepreneurs or managers who make riches by pretending.

 

In airline industry, there is a respectable metric of flight hours per fatal accident. But that metric is never a goal by itself and it was not improved by focusing solely on it. It was improved by design of engineering systems to achieve reliability and safety, by specific standards and practices of operating flights, by developing the Crew Resource Management system to reduce human errors in the cockpit.

 

I would love to see specific factors and reasons of why the autonomous driving systems are reliable. So far I only see application of the modern pattern matching AI, which doesn't exactly demonstrate capability of systematic protection from errors and failures.

Edited June 22, 2025 by Loss Horizon

[Spekulatius]

Both miles driven and miles per disengagement are useful metrics. Time per fatal accident is also useful but hopefully the incident rate is so low that the stats will be very noisy around that one.

On the topic of sensors it is quite funny that Elon rejects  Lidar even though it is now used in robovacs (albeit less powerful ones with less range). I think he will change his mind on that one but by that time it will be clear that none of the vehicles sold so far with FSD without those will ever be truly self driving.

Edited June 22, 2025 by Spekulatius

[changegonnacome]

 

There are two parallel paths to widespread adoption - statistical safety outperformance & political/regulatory acceptance.......the car on the left above may one day get to statistical safety outperformance.....the question is will it outperform the car on the right?.....and therefore what becomes the widespread political and regulatory standard for sanctioned widespread autonomy.

 

Ask yourself.....as a regulator, politician tasked with coming up with the greenlight rules/legal framework when lives and public safety are at stake.....do you sign off on the car on the left or the car on the right as the minimum safety standard.....its an easy choice IMO and an obvious one.

 

 

Edited June 22, 2025 by changegonnacome