[mike_d]

Waymo is amazing, knowing some of the stuff they do behind the scenes to ensure safety - I would feel safer riding in a Waymo than driving myself.

My biggest fear has always been that Cruise or Tesla would shit the bed so bad we don't get any self-driving, either because of regulatory constraints or ruining the public perception of them.

[ehsankia]

Absolutely. Glad California DMV is taking the more level headed approach here and evaluating them case by case based on merit.

[startupsfail]

In the case of that particular incident, it seems that the pedestrian was “thrown” into the lane with Cruise when it was hit by another (human) driver.

[TuringNYC]

Can any engineers in the field comment on whether the [very cool!] Waymo vs Cruise performance anecdotes being discussed here would have been the result of Google's millimeter-level scanning of SF? Or is it just better algos only? Or both?

[jprete]

I’d think better algorithms. A detailed scan of a city’s streets and layout is obsolete before it’s completed.

A driver (bot or not) needs very good reactions to bad situations observed in the field. And ideally a good sense for when a situation is one step from trouble. As a driver, if I can’t see through a bus then I’m going to assume a crazy pedestrian is waiting to jump in front of me.

(A lot of drivers are quite bad frankly and don’t do most of the stuff you’d want someone to do if they’re directing a one-ton blunt weapon around town at 35mph.)

[giantg2]

"A lot of drivers are quite bad frankly and don’t do most of the stuff you’d want"

I'd say 90% of driving is good/safe driving is mental, with at least half of that bening based on the decision making capability. Yet our tests are mostly focused on physical ability with a small amount of memorizing a subset of the rules.

[solardev]

I wish we put our drivers in driving sims the way pilots have to accrue flight sim hours.

[rrr_oh_man]

Pilots (at least to get the PPL) train at least around 40-60 hours on a live aircraft before flying solo. That's more rigorous, yet very close in principle to what we have in place to obtain a driver's permit, I'd argue.

[solardev]

I think it only took me a half dozen hours or so to get my license? But also that only tested me in best case scenarios with good visibility, other good drivers, etc. A better sim should test and train my response to unexpected events in less than ideal conditions.

[daveguy]

I certainly didn't have 40-60 hours in a car before I got my license. Much less learning permit.

[AustinDev]

Cruise vehicles have sufficient sensor payloads to produce a 'milimeter-level scanning of SF' with many fewer hours than they've been on the road there. I doubt that provides the biggest difference. It's all down the software at this point. Especially because they're not trying to do anything crazy like vision only.

[kbos87]

We keep hearing this is crazy, but is there evidence to back that up? At this point with hundreds of thousands of Tesla vehicles logging millions of miles a month, I feel like we’d be hearing non-stop reports of destruction on the highways if it really is as crazy as the average HN reader has been led to believe.

[Galaxeblaffer]

the only reason we don't hear non-stop about destruction is that you're required to keep both hands on the wheel/yolk at ALL times while FSD is activated. comparing the current cameras on a Tesla to the human eye is just silly really, sure they cover 360 which is superior, but in almost any other way, the human eye is superior having better dynamic range, better contrast, better resolution and best of all, we have eye lids. if you would just sit down and think about it for 10 mins you'll realize that gimping your self driving car by using cameras only is just a very very silly proposition when we have all this cool tech like lidar essentially giving the car super powers.

[kbos87]

If FSD drivers were having to constantly intervene because the car couldn’t accurately map obstacles, we’d be hearing a lot more about it. I drive with FSD all the time - I could give you a list of things it needs to improve on, but not a single one has anything to do with its accuracy of understanding its surroundings.

[bumby]

>not a single one has anything to do with its accuracy of understanding its surroundings.

This has been my gripe for a long time. I feel like many in tech have conflated two problems. With current software the problem of perception (ie "understanding its surroundings") is largely solved*, but this shouldn't be conflated with the much more difficult problem of self-driving.

*for sure, there have been issues with perception. A glaring example is the Uber fatality in AZ.

[space_fountain]

I do see stories about FSD seemingly trying to drive into obstacles fairly often. It’s true that it does see most obstacles, but most is not good enough for this.

[daveguy]

Accuracy of surroundings is absolutely something it could improve on. Adding a different modality (like lidar) would be like adding another sense. Seeing an 18 wheeler without under guards would be easier with an additional sense. It makes the intelligence part easier because the algorithm can be more sure about it's interpretation of the environment.

[electrondood]

No longer true. Now they use the camera to verify you're looking at the road.

Also, we can drive just fine with two cameras and a neural net, no LIDAR.

In engineering, you don't add something unless it's necessary, not because "well we may as well have it."

[goalonetwo]

you listened to Musk a bit too much.

And no, a neural net and two cameras are not "just fine". The day cameras will be as good as your eyes and your neural net will be on the level of human intelligence (AGI) then maybe it would be possible. But until then you will need to rely on extra hardware to get there.

Go check on youtube how FSD behaves in city with 1/10th the complexity of SF/Waymo. And remember the difficulty is with the long tail of unexpected events.

[solardev]

We don't drive just fine, we routinely kill each other, often because of poor visibility or not noticing motion. Backing into a busy street? Bam. Open your door without checking? Biker down. Passing a bus at a crosswalk? Pedestrian dead. Driving at night with heavy fog? Off the cliff.

Even your basic non-fancy non-AI car these days have a variety of sonar/radar assists to help out their cameras. Tesla is just being cheap (and getting people killed because of it).

[trogdor]

>with hundreds of thousands of Tesla vehicles logging millions of miles a month, I feel like we’d be hearing non-stop reports of destruction

Why do you believe that Tesla vehicles are driving millions of miles a month on autopilot?

I suspect that autopilot accounts for a vanishingly small percentage of Tesla miles-driven. Mostly because it sucks, but also because of how hard it is to not miss a nag on a long road trip, with the result that you are locked out of autopilot until your next stop.

(Yes, even when paying attention to the road, with your hands on the wheel. Hell, autopilot locks you out if you exceed 85 mph for more than a few seconds by pressing the accelerator, such as when passing. This is true even in places where the speed limit is 85 mph, and the slower end of the flow of traffic is 95 mph.)

I love the performance of my Model S Plaid. Autopilot, however, is a joke.

[sgustard]

Wow, I've had much more positive experiences with Autopilot on multi-hour highway trips. No shutoffs, no phantom braking in the last year or two. Autopilot on the highway is so much better than any other car I've driven with adaptive cruise. Anyway, the parent is comparing city and highway self-driving and they are completely different.

[kbos87]

Tesla reported that there were 400,000 cars in the US and Canada who had access to FSD Beta in January 2023, so I don’t think that’s at all hard to believe.

Edit: Looks like back in January they were adding 10M cumulative miles driven per month (roughly) - https://insideevs.com/news/633328/tesla-fsd-beta-now-active-...

[trogdor]

Assuming those data are accurate, it looks like you’re right.

Surprising, but it is what it is.

[_ea1k]

I probably use AP >80% of the time on the highway. From observing comments in the local Tesla groups, I'm sure there are plenty of people who behave similarly.

Millions of miles per month doesn't surprise me at all.

[nohuck13]

"This is true even in places where the speed limit is 85 mph, and the slower end of the flow of traffic is 95 mph."

Where do you drive where the speed limit is 85 mph?

[gundmc]

Highway 130 outside Austin, Texas has an 85 mph zone.

[gambiting]

Poland? 140km/h is the motorway speed limit. Admittedly there aren't that many Teslas here. But they are extremely popular in Germany and you see people drive them at silly speeds.

[bumby]

There are a number of highways in Texas with an 85 mph limit.

[F-W-M]

> This is true even in places where the speed limit is 85 mph, and the slower end of the flow of traffic is 95 mph.

Once traffic rules are enforced by autonomous cars that don't let you go over the speed limit, we can start raising the limit to more realistic levels. I guess currently it is priced in that everyone goes over by x km/h.

[salawat]

>Once traffic rules are enforced by autonomous cars that don't let you go over the speed limit, we can start raising the limit to more realistic levels. I guess currently it is priced in that everyone goes over by x km/h.

Bloody noghtmare fuel that, but it's been the inexhorable direction manufacturers and Governments have been pushing for for a while.

Stop making things that enable people more than it constrains them. The act of velvet glove societal manipulation is pretty damn obnoxious once you realize it's a thing.

[amelius]

It's the sensors and Lidar.

[qeternity]

> knowing some of the stuff they do behind the scenes to ensure safety

Any info on this? I would love to learn more about what they're doing.

[xnx]

https://waymo.com/safety/ is probably your best resource. You can follow the links from there to learn about all the things they're doing with simulation, etc.

[coding123]

I know most of it was PR but the stuff Waymo did about showing a test course they built and actually having "unit tests" related to driving showed me they're not dumping these things on the road to make money, they're trying to get it right.

The competition:

Uber kills someone and turns out they disabled "breaking due to objects in the way" because it makes it jerkier.

Tesla is often killing the drivers because they marketed it as "auto-pilot"

Cruise rolled over someone for 20 seconds.

I would have to agree. I honestly hate regulatory capture in most industries, but perhaps Waymo's "unit tests" should be turned over to the DMV and made as the gold standard driving test for AI drivers.

[littlestymaar]

> My biggest fear has always been that Cruise or Tesla would shit the bed so bad we don't get any self-driving, either because of regulatory constraints or ruining the public perception of them

Exactly this. When I saw Tesla releasing their half-backed crap I feared they will damage the public perception so badly that we wouldn't have self-driving cars for years.

[nonethewiser]

> Waymo is amazing, knowing some of the stuff they do behind the scenes to ensure safety

Such as?