Comment by Fricken | Hacker News Reader

Comment by Fricken | Hacker News Reader

Parent Full thread

At the end of 2016 all companies testing autonomous vehicles in California filed their disengagement reports. Waymo reported 1 unplanned disengagement every 5000 miles, and GM reported 1 every 180. Nobody else was even in the running. These metrics are crude, but it's currently the only benchmarking system we've got, and we'll see in December where Cruise is really at when they report their 2017 numbers.

Cruise does distinguish itself by driving mostly on busy downtown SF streets, a challenging environment, and they've released impressive videos showing off their capabilities.

A few weeks ago Kyle Vogt claimed Cruise would surpass Waymo 'in a few months', whether they do or not remains to be seen, but they are making sustained progress and they have the full force of General Motors behind them.

Unlike Waymo, Tesla's Autopilot program, and Uber's autonomous driving program, Cruise's operation has been drama free, without any lawsuits, leadership blowouts, or departures of key people. This implies they've got a solid team under good leadership, and I think it matters a lot with complex software projects like this.

Waymo (Google) is not that far from level IV. Average U.S. driver has one accident roughly every 165,000 miles and disengagements is not 1:1 with crashing. I suspect drunk drivers would probably be safer handing over control which could make a huge difference.

Most interesting is the Disengagements by location. (page 9) https://www.dmv.ca.gov/portal/wcm/connect/946b3502-c959-4e3b...

  Interstate 0
  Freeway 0
  Highway 12
  Street 112

Which might be good or bad depending on the relative amount of driving in each situation.

PS: Disengagements per 1,000 miles also dropped from 0.8 (2015) to 0.2 (2016) which suggests very good things.

PPS: Been getting a lot of up and down votes I wander what people agree / disagree with?

I don't know why you're getting downvotes.

Does anyone know off-hand, how much testing have autonomous vehicles had at night (as in, full dark/need to use headlights).

Cruise has been testing cars 24 hours a day, 7 days a week for quite some time. Here's a few hours of fully autonomous night driving: https://www.youtube.com/watch?v=6tA_VvHP0-s

Sounds like you're assuming level 4 merely means parity or better with human error. Lots of problems with that. For instance if it's Google's negligent bug that kills the family of 4 instead of a driver, Google gets sued. Google has much deeper pockets than the driver who would typically declare bankruptcy, effectively capping the damages. No real cap for Google, every time. So that explodes the cost of autonomous vehicles. And then there's the matter of potential criminal liability in the case of gross negligence. So there's going to be a much higher bar for transitioning to level 4.

Also, while there may be fewer of some types of accidents, there will be lawsuits for a potentially very large set of other types that never used to get litigated. For example today when a driver causes their own accident, nobody gets sued.

Google however does have the ability to hand over the cost of insuring such a system to the consumer, so even at current accident rates the cost is around 100 dollars a month per car using such a system (driving 12000 miles in a year). This equates to something like a cent per mile. So let's say they want incredibly high quality insurance, charging an extra 5 cents per mile won't make a difference in the scheme of things should they be able to get the overall cost of commuting 15 miles (the average American commute in one direction), to something under 5 dollars.

This aspect of insurance is widely misunderstood. First of all an extra 4 cents/mile just for insurance is really expensive. But wrongful death and personal injury lawsuits can result in very large judgements, especially when we get to punitive damages for a large corporation. Individual offenders regularly go bankrupt. It may be wishful thinking that this can be easily solved with "insurance". There is also the criminal negligence aspect to consider.

4 cents/mile is roughly in line with rates for traditional auto insurance today. And as these rapid improvements are likely to continue, these costs will quickly come down.

Sure you can imagine some class action / criminal legal disaster but the auto industry has survived repeated scandals. Corporate criminal negligence is extremely rare.

In five or ten years this technology could be saving 100,000 lives every year. I hope we are not so risk adverse that these benefits are needlessly delayed.

http://www.insurance.com/auto-insurance/auto-insurance-basic...

Medical errors kill on the order of 700 people per day in the US (high error bars on that number though). And yes, insurance is a major cost in the medical field, but even large scale negligent deaths are regularly dealt with by the US legal system.

Yes, anyone putting out such a system is going to get sued, but so does every car company that exists today. It's just the cost of doing business and ultimately just get's passed on to consumers.

You're correct. The question is what is the cost passed on to consumers. If it is something on the order of the cost of medical malpractice insurance and average settlements, well, wow. That would make AV prohibitively expensive.

More likely the bar for level 4 will just be significantly higher than parity with human error, if only for liability reasons.

"In 2010, there were 1.1 fatal crashes per 100 million truck miles" https://www.truckdrivingjobs.com/faq/truck-driving-accidents... Truckers make ~27c/mile.

So, at similar rate to human drivers that's ~24* million dollars per fatal crash in saved wages. Plus presumably higher truck utilization and lower management costs etc. Don't forget the driver is often not at fault and automated trucks should have plenty of footage of accidents to demonstrate who was at fault. So it's a fraction of both fatal and non fatal accidents, with non fatal accidents generally having lower associated costs. Also, most importantly it's only the truck involved in many accidents associated with bad weather and more generally the driver's medical bills etc do not need to be covered if there is no driver.

Further, technology is unlikely to degrade over time so they just need a profitable starting point. Also, if the self drive button cost ~25c / mile there are plenty of times I would hit it.

*Granted, there will be labor costs linked with automated trucks, but a 95% savings seems likely.

> if the self drive button cost ~25c / mile there are plenty of times I would hit it

25c/mile just for the insurance surcharge, you mean? Sounds like we're in agreement that mere parity with human error would be expensive from a liability perspective.

I don't think the average fatal accident is going to average even 5 million, I am just saying if it cost 20 million that's still not a deal killer. Put in other terms Valet Parking is 5+$, but 'go park' is ~5 cents. In ~15MPG city / stop and go traffic that's 3.75$/hour to not pay attention.

Hell go to sleep and wake up in another city 500-1000 miles away without going to the airport or needing a rental is a major benefit even if your not saying money.

Based on their own reporting, hasn't Waymo been safer than the average driver for a few years now?

The real question is, what routes are being followed and in what conditions? Presumably Waymo et al stick to exquisitely-mapped routes with excellent road markings between well-known autonomous-friendly endpoints during good daytime weather for the vast majority of their testing. Performance in less-than-ideal conditions is a more important question. I'm curious if there's any publicly available data on the routes followed, times, conditions, etc.

Anecdotally I see them driving like snails on Valencia Street and elsewhere in the Mission day and night. It's very well mapped, but one of the more challenging areas to navigate in America.

I've never seen them pulled over in the bike lane on Valencia like very other Uber/Lyft, which is a good thing, but picking up and dropping off is far from autonomous friendly in that neighborhood.

Of course sometime it's obvious a person is in control and presumably they're generating training data.

There was actually a dispute about a potential Cruise cofounder getting compensation for equity at the acquisition. They recently settled the lawsuit out of court [1]. I agree they have good leadership, but they're not without any lawsuits.

[1] http://www.businessinsider.com/car-startup-cruise-settles-le...

>> Cruise does distinguish itself by driving mostly on busy downtown SF streets, a challenging environment, and they've released impressive videos showing off their capabilities.

I don't doubt that it's a challenging environment, but there are many, many different kinds of challenges that are not even present in that city.

True. But city driving is all they need to offer service in the most potentially profitable, geo fenced areas.

Driving is SF is actually quite challenging. It is quite congested with cars as well as pedestrians and cyclists. Due to lax enforcement many traffic laws are routinely ignored particularly by cyclists, pedestrians and transit vehicles. On top of this there is lots of construction, poorly maintained roads and frequent traffic flow changes.

Not sure what you are imagining but SF is free of winter driving conditions. Other than that it is as challenging as any city in the US, Europe or Japan.

Is disengagement per X miles really a good metric. It doesn't reflect the difference between driving in different environments, i.e dense urban areas, suburban streets, and open highways.

Yes, it's a crude metric. The other thing is disengagements are self reported, and an unplanned disengagement doesn't necessarily mean a crash was narrowly avoided, it may be an overly cautious test driver, or the result of some non-safety critical intervention.

After the core software problems have been solved, the key challenge shifts to discovering edge cases, which become progressively more marginal the further advanced the software becomes, and the only way to find these edge cases is through 1000s of miles of driving waiting for unexpected situations to show up.

Reply on news.ycombinator.com