Comment by Animats | Hacker News Reader

Comment by Animats | Hacker News Reader

Parent Full thread

Waymo does seem to be ahead technically. They've been working on the hard cases for years, while the other guys are still struggling. They are trying hard to get good situational awareness. One of the few neutral measures we have are CA DMV's disconnect and accident reports, and in those, Waymo has been consistently ahead. Waymo is hooked up with FCA (Fiat-Chrysler Automobiles), of all things.

Uber is "fake it 'til you make it". Otto's truck demo only worked on a deserted highway surrounded by chase cars. Cruise also started out that way, but they seem to have improved once acquired by GM. Those guys seem to drive mostly on visual cues, without as much of a near-area model as Waymo.

Tesla has the self-crashing car. Four collisions with something partly blocking a lane, all on video. I've posted the links before. Musk made a big mistake trying to do it without LIDAR. Nobody else can do that reliably, and neither can Tesla. Tesla had to turn off most of the autonomy. Now they're shipping the Model 3 in small quantities with hardware that's only enough for lane keeping and auto-braking.

Volvo has a good, but limited 100 car demo with real drivers running. Freeway only. Their goal is "no crashes". They'll probably get there.

Ford is quietly doing something and not talking about it much.

Meanwhile, Continental and now, Delphi are quietly gearing up to make low-cost LIDARs by the millions. Despite all the talk about self-driving as a service, it may turn out to just be an auto part.

Tesla's approach may really have fundamental problems, but it's impossible to compare the results from a driver assist system used by untrained consumers with an autonomous driving system that's only being used with professional safety drivers in conditions for which is expected to function properly.

IMHO, the big difference between the Uber/Tesla approach and the Waymo approach is that Waymo seems to understand that this is an extremely hard problem to solve. Uber/Tesla seem to be happy with getting 90% of the way there and make headlines. I don't think either Uber or Tesla have really committed to the technology the way Waymo has.

As far as the legacy auto-makers, it's really unclear what they are up to. I think they appreciate the difficulty, but probably don't have the skills to move faster than Google.

It's quite possible to compare them. Google has never done anything like this.[1] This is a Tesla vehicle on "autopilot" crashing into a street sweeper without even slowing down. It's on an expressway, where Tesla's "autopilot" is supposed to work. The street sweeper is big, partly blocking the lane. The driver trusted Tesla's "autopilot" to do something reasonable. It didn't.

The driver was killed. That blood is on Elon Musk's hands.

[1] https://www.youtube.com/watch?v=1cfcumft5n8

This is incorrect, and you know it's incorrect. You're talking about it like it was a Level 3 system, but it's Level 2.

The details of that video are fuzzy at best. The crash happened 8 months before it was reported in western media, before Tesla made any announcements about fully self driving systems. It's not even entirely clear Auto Pilot was involved.

Still, there have been confirmed accidents involving Auto Pilot. That's not to say it hasn't been acting as designed. It is currently a system to help drivers when they make a mistake, not a way to prevent drivers from needing to drive. They've said it will do more eventually, but that's not what they're selling.

It's entirely possible that Waymo's system is completely incapable of dealing with street sweepers stopped the middle of a passing lane of a highway. If that's the case, there's no way the professional test drivers would be taking the car into that kind of situation. The public isn't given that degree of insight into their testing methodologies. This contrasts with shipping systems, where the public is implicitly given a users (potentially mis)using the system. It also means that Tesla has orders of magnitude more distance of real-world test data than non-shipping systems.

Excuses, excuses. Waymo's system regularly demonstrates its ability to navigate around unexpected obstacles. Tesla just has "lane changing", as if roads were railroad tracks.

One question I can't seem to find a good answer for is this:

Why doesn't the self-driving tech hand off control to a remote driver if the internet connection is good enough and there is enough warning?

Even with a couple seconds notice a primed human could take the wheel and do the right thing when the sensors aren't making sense.

There's no good reason; it's likely what they'll do.

Never will happen. Latency is a law of nature. 60mph === 8.8 ft / 100ms. Near misses and close calls with unexpected objects can't afford the additional delay of a radio round trip that may or may not be working to make safety critical decisions. All situational awareness has to be handled on board for immediate response when needed.

I've worked on a videogame where vehicles are controlled from the client but have their physics done on the server. Server sends car position, and client sends inputs. So I can give some insight into how driveable it might actually be:

- <40ms (round trip) is not really noticeable

- 40-~80ms is driveable

- 80-150ms is getting difficult

- 150-250ms you're overcorrecting all the time

- 250ms+ is completely undriveable

That's not trying to avoid sudden obstacles, that's just making turns etc that you already intended to make. Avoiding anything sudden would be impossible much earlier. I agree remote control over Internet will never happen.

If you set up a control center in each metro area, you could potentially have a sub 10ms RTT.

And I imagine that in a large percentage of cases the car's sensors might indicate potential danger with enough advance warning for the remote driver to take over.

Obviously the operators won't control directly the vehicule :-)

The operators will have a map on which they click to set the vehicule's destination, and the console will show what will be the exact path of the car, and ask for confirmation.

Thought experiment: If you had such a system and it was good enough that people would not notice it’s use, would you advertise that? I would look carefully at the wording of these companies. Such a system would let you paper over all sorts of issues and claim zero “driver interventions”...

"Your upcoming crash is very important to us. Please stay on the line and a representative will be with you shortly."

There is no "primed" human on the line or you might as well just have a driver. And leaving aside issues of controlling remotely, you're probably talking on the order of minutes for someone to establish situational awareness, figure out what's going on, and take control. Which may be fine for something like Onstar but not for time-critical vehicle control.

There is obviously an xkcd for that https://xkcd.com/1897/.

This is what they do to circumvent legal reporting requirements. You don't have to report a hand-off to a remote driver, but you have to report a hand-off to the safety driver in the car.

The latencies are too high. It’s not an aerial drone.

I think FCA had not yet started to build an automatic driving team in house, so pairing up with Waymo would make sense and Waymo could push for a deal it thought worth making. Maybe Waymo got a non-exclusive deal, where other companies might have wanted some exclusive period with Waymo tech in their cars. Similar to what happens with cell phones.

Here is the data to support your post.

https://www.wired.com/2017/02/california-dmv-autonomous-car-... The Numbers Don't Lie: Self-Driving Cars Are Getting Good - Wired

For those looking, Animats' past comment containing videos of Tesla crashes is here: https://news.ycombinator.com/item?id=12942316

Reply on news.ycombinator.com