Hacker News Reader

Comment by petilon | Hacker News Reader

Meanwhile Cruise just obtained permission to charge passengers for driverless rides in San Francisco [1]. Cruise has been cleared to charge for rides in vehicles that will have no other people in them besides the passengers. There will be no back-up human driver present to take control if something goes wrong.

Why is Tesla so far behind? Tesla is generating these autopilot mishap news on such a regular basis that it is giving the entire industry a bad name.

[1] https://www.npr.org/2022/06/03/1102922330/driverless-self-dr...

View on HN

Presumably, these regulators are the same, that allowed this Tesla very alpha software to be tested on public roads. Also from your own reference it does not seem like Cruise capabilities are much above an amusement park level toy ride:

"...The regulators issued the permit despite safety concerns arising from Cruise's inability to pick up and drop off passengers at the curb in its autonomous taxis, requiring the vehicles to double park in traffic lanes..."

"...vehicles confined to transporting passengers in less congested parts of San Francisco from 10 p.m. to 6 a.m. Those restrictions are designed to minimize chances of the robotic taxis causing property damage, injuries or death if something goes awry..."

"... Cruise's driverless service won't be allowed to operate in heavy rain or fog either..."

Plus it seems the system is another alpha/beta release:

"A Fleet of Confused Self-Driving Cars Stopped in the Middle of a Street for Hours": https://gizmodo.com/cruise-confuse-self-driving-cars-san-fra...

Yeah but how often do you see negative stories about Cruise and Waymo? Where lives are endangered? The story about cars stopped in the middle of a street doesn't count since lives were not in danger.

Does Cruise on Cruise crash counts?

"A Cruise-on-Cruise Crash Reveals the Hardest Thing About Self-Driving Tech": https://www.wired.com/story/cruise-on-cruise-crash-hardest-t...

In this case a Cruise car was rear ended by a Cruise car driven by a human, because their unsophisticated software makes for the car to break hard. As explained in the article:

"...That said, the rear-endings demonstrate that the technology is far from perfect. Cruise cars follow road laws to a T, coming to full stops at stop signs and braking for yellow lights. But human drivers don’t—and Cruise cars will be self-driving among humans for decades to come..."

"...The fact that a driver Cruise trained to work with these vehicles still managed to rear-end one emphasizes exactly how flawed they are..."

This story is from 2018...the lidar-focused portion of the AV industry has advanced lightyears since then.

You may as well post an article about Eliza to make a pt about DallE.

And the limited domain you're describing for Cruise is the domain in which they've reached the statistical safety performance required to fully remove a safety driver. They and Waymo are driving all over the city during the day, with very few disengagements: it's been a couple years since they've been able to post credible evidence of long and complex disengagement-free drives.

The last epsilon% of safety performance is the hardest, and I'm not convinced they or Waymo can crack it soon. But your comments are horribly misinformed.

They are increasing the risk for others as:

And after all those light years the LIDAR focused industry can't still see in fog, dust, rain or snow. An Eliza, just with more compute power and some badly understood ML algos.

Yes, I have seen the demo videos of self driving cars on snow.

> Presumably, these regulators are the same, that allowed this Tesla very alpha software to be tested on public roads.

I don't think Tesla has permission from CPUC to run its public testing program. Last I read (articles from a year ago), Telsa is telling CPUC that Full Self Driving is only Level 2; CPUC only regulates levels 3 and up.

If we’re going to try to have autonomous vehicles, at some point these things have to be tested on real roads. Would you rather have a real human behind the wheel to take over in an instant or just have the computer plow through whatever if there’s a bug? When you are driving tesla “autopilot” you are still liable for everything the car does, you still need to be attentive just like you normally would. If the car crashes despite that, then you weren’t paying enough attention and that is your fault. In my experience any time it did something stupid, there was always plenty of time for me to take over and easily avoid an accident. In other words, it’s not going to rapidly swerve into oncoming traffic at 70mph. If you don’t want to accept the legal risk of what your car does on beta autopilot software or don’t want to continue paying attention to the roads while on autopilot then you shouldn’t be using it.

You get way more distracted if you are doing nothing rather than if you're actually doing something. So I'd expect the reaction times for people to stop an autopilot from killing them are actually much higher than the reaction times of a driver.

Then they should not be using autopilot. When I have autopilot engaged I am way more alert. Also you aren’t “doing nothing” when you have autopilot engaged. The car annoys you if it doesn’t sense any force on the wheel within 10 seconds and even if you are looking at your phone (or anywhere other than the road) the car annoys you too (yes there’s a camera in the cabin looking at you when you’re on autopilot).

So what you mean is that Tesla should be using trained safety drivers like real AV companies do? Also, to be clear, you are either lying or misinformed about the behavior if autopilot as far as detecting engaged driving goes.

> When I have autopilot engaged I am way more alert

Hm… so what's the point?

The point is acknowledging that over a million people die per year on the roads due to human error and trying to help beta test software that would help bring that number down over 90% so that our kids can have a better future.

Before testing on real roads, these things should be tested in realistic closed environments. Tesla has been too cheap and lazy to take that basic safety measure.

Is Tesla so far behind or are we just not seeing all the dangerous behavior of Cruise-piloted vehicles because they haven't made their software available to the general public?

IMHO none of this software should be getting tested on public roads.

Tesla is behind because they had a fucking big pile of C/C++ + machine learned models and wanted to brute force their way to an acceptable quality.

In a recentish Lex Fridman podcast Elon finally talks about how they need to solve the hard part which turns out to be ... drum roll .. world modeling! (Kids in a school zone or behind a school bus or something like that was the example I think.)

Who would have thought that without a model that can provide a consistent "abstract model" of the world it's hard to fit incoming noisy sensor data?

Cruise and Waymo have lidar. Lidar can give you much more certainty about detecting objects and distances. Teslas only use cameras and have to do complex image analysis to detect objects.

On the other hand lidar is more expensive. But it is getting cheaper and it does seem like it might be necessary for actual self driving.

The train appears on the car's HUD and the distance appears to be calculated properly before it tries to turn in front of it. I don't think object detection is what went wrong in this case (which is scarier imo).

Having LiDAR may make it easier to implement, but there are plenty of humans who can drive well without LiDAR, so having LiDAR isn’t necessary.

The human brain is far faster at image processing than any computer you may be able to put in a car. So shortcuts are necessary. Perhaps one shortcut is to get actual reliable object shape and position data from Lidar than trying to infer it from flat images.

We don't want robots to make human-level mistakes. Robotic drivers should work much better than humans, otherwise they should not exist.

Tesla is far ahead. Cruise was literally in the news like 2 days ago because their entire fleet just stopped driving in the middle of an intersection and dozens of their staff had to drive out to the cars and manually take them back to the parking lot. Cruise works (except for when it doesn't) in extremely specific geofenced areas in very low traffic. Tesla works anywhere and in a lot more conditions.

Tesla is only far ahead in their willingness to take on risk and make at least partially untrue claims. There are multiple companies running actual robo-taxis in other cities but not a single city Tesla is running actual driverless vehicles. If Tesla was actually capable of running robo-taxis in one of the cities with the regulatory atmosphere they would, but they can't.

I opted out of FSD because it doesn’t work. I don’t get how people can keep saying it works. It’s incredibly dangerous to say that.

If your standard for "works" is 100% flawless driving then you should probably opt out of human driving too since that also doesn't work

All about where it works. Geofenced solutions are much easier.

I wouldn’t say they’re behind, just solving a different problem. Should they have focused on inside city and highway driving? Probably. Would’ve had a good product sooner. Even if they don’t win this whole thing, they pushed the entire industry in a good direction.

Why do you believe that Tesla is behind?

Tesla is selling a large number of vehicles with FSD, used across a large operational design domain.

Cruise isn't selling cars. Small fleet, that can operate at limited hours at night, at limited locations in San Francisco. Small scale and limited operational design domain compared to Tesla's.

> Why do you believe that Tesla is behind?

Because Cruise does not need a back-up human driver present. Tesla seems far behind because as this story indicates, a human needs to be paying close attention when Tesla is in charge.

Overall, Cruise/Waymo appear to have the better approach. I would rather start with a system that works and is reliable and expand its capabilities, than start with a system that barely works and hack it till it works. Especially when human lives are endangered by the tech.

Roller-coaster rides don't need a backup human driver. Are they ahead of Tesla in tech?

You can't buy Cruise / Waymo car or use it like you can with Tesla.

Roller coasters don't crash into emergency vehicles... so yes.

Because as much as Tesla marketing wants us to believe, FSD is anything but. Tesla is so ridiculously far behind real AV companies that I must assume that you're a paid shill or an absolute idiot.

The answer is that Tesla is creating a general solution, whereas Waymo and Cruise and targeting specific geographical areas which they can map precisely.

Waymo and Cruise are also using more expensive, bulky and unsightly hardware, including lidar. They are starting at the top, with a solution that works, then working their way down. Their approach is confidence-inspiring. Tesla is starting at the bottom. The frequent mishap stories coming from Tesla does not inspire confidence.

You can also start at the bottom like most European manufacturers, and slowly build up to autonomous driving.

That is, start with collision warning, lane assist etc. Boom, you have a Volvo that drives itself on a highway, and requires driver assistance elsewhere.

Doesn’t most taxis revenue come from city centers only ( 80-20 rule) seems as if Cruise’s solution would be prudent

Because they don't have Lidar. They're just attempting to solve a completely different problem, which is much harder. IMO it's a horribly irresponsible approach

Reply on news.ycombinator.com