Hacker News Reader

Comment by YeGoblynQueenne | Hacker News Reader

>> Keeping the safety drivers in the beginning, scaling up, and getting a year or more worth of actual real world use, before slowly scaling back the drivers is what I'd call a very responsible choice.

I'm sorry but where does this "one year" expectation come from? What says that Waymo needs a year (or two years, or ten, or three months, etc) to start "scaling back the drivers"?

From what I can tell, the only way to judge the competence of their driverless systems is to compare where they are today with were they were 10 years ago when they started out. However, this is next to impossible- that data is not public.

So, where does that "one year" optimism come from? Autonomous driving won't magickally become easier in a year from now.

View on HN

No, the only way to know how well it's doing is for it to get enough "interesting" miles, and then compare the success rate of those miles to similar human-driven miles. Right now Waymo is past 8 million miles driven. Miles driven is a better and normalized metric, with comparable real world statistics, unlike time.

Miles driven is a less valuable number if the miles are largely along "easy" paths. What interests me is challenging events, perhaps what you call "interesting miles" because the more of those that cars successfully navigate, the more real-world ready the autonomous vehicles will be. There are hundreds of different types of these situations that are common enough that these vehicles will have to react appropriately (construction, detours, officer using hand signals, missing or misleading road paint, inaccurate maps, etc.) I don't know where Google's cars have driven, so they may have encountered these things, but I don't know.

I did explicitly say "interesting" miles, but we shouldn't mix up Phoenix miles with New York miles or Canada miles. No one claims they are ready to run anywhere in the world, but they have enough miles in Phoenix that they're confident they've seen almost all interesting miles there.

Waymo is past 8 million miles total, how many miles on the latest version of the software? Certainly there are new features that are there in the last half than the first half? I think miles count is a bad metric and if it wasn't, Tesla would be the front runner.

All those miles are recorded, fuzzed and replayed in simulation, where they've driven over 5 billion miles. I assume they retest all the miles + fuzzed versions on every version release.

It’s past 10 million miles and it’s driving more than 500k mi per month. If the last version of the software was in August that it’s more than 2 million miles by now. The miles per month figure will only increase.

And how many of those miles are up and down 101 after midnight on weekdays?

From what I can tell, the only way to judge the competence of their driverless systems is to compare where they are today with were they were 10 years ago

That’s hardly the only way to judge the competence of their driverless systems.

> the only way to judge the competence of their driverless systems is to compare where they are today with were they were 10 years ago when they started out.

How does a self-relative measure capture the absolute goodness of the vehicle?

I'm much better at skiing than I was two years ago, but I'm not ready for the Olympics. My history is irrelevant to my ability.

The exact paragraph you quoted says “a year or more”.

That paragraph is not in the article or anything Waymo has said.

It's just my speculation, if after a year of a huge number of successfully completed rides without incident, they'll obviously be more confidence in the system.

My point is, the responsible thing for self driving companies to do is maintain safety drivers until confidence increases.

Reply on news.ycombinator.com