Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Waymo CEO dismisses Tesla self-driving plan: “This is not how it works”

arstechnica.com

Post reply on HN
Print view View original

521–530 of 581 posts

Re: Waymo CEO dismisses Tesla self-driving plan: “This is not how it works”

#521

Post by sudosysgen »

strangeattractr wrote: post #185
TheDong wrote: post #161

Earlier quoted context omitted.

Your impression of Delhi and Agadir doesn't seem right to me. By the numbers, Delhi has the highest crime-rate in india, but it has a 5 times lower rate of crime than San Francisco does. And also, self-driving cars will all have enough cameras and be constantly connected to a server that theft will be far less likely for them than regular cars... and regular calls already seem to get by just fine without people strip…

Unreported crimes aren't factored into the stats you quote.

That works on both sides of the equation, as someone who lived both in NA and Morocco. There's similar levels of unreported crime.

Re: Waymo CEO dismisses Tesla self-driving plan: “This is not how it works”

#522

Post by sudosysgen »

pas wrote: post #383
ashtonkem wrote: post #238

Earlier quoted context omitted.

Sure. But that’s a moot point if it turns out you need to basically replicate large parts of the human brain in order to drive with cameras alone.

But those are needed for L5 anyway, Lidar or noLidar. The big issue is the world model, what's the prior when uncertainty rises, how to spot really out of ordinary stuff, and how to spot really ordinary-looking but dangerous stuff.

I'm fairly certain L5 automation won't need anywhere near the processing capacity of the visual cortex.

Re: Waymo CEO dismisses Tesla self-driving plan: “This is not how it works”

#523

Post by throwaway316943 »

na85 wrote: post #340
meekrohprocess wrote: post #151

Earlier quoted context omitted.

I have one question for Tesla customers who trust the company to deliver full FSD. How do you reconcile that belief with the fact that Tesla's embedded developers did not understand the extremely simple concept of write endurance? The NHTSA opened an investigation into premature HUD failures because they prevented the backup cameras from working. But the fact of the matter is, the company used a small partition of in…

>You're also entrusting my life, and those of my family, to them. But we'll gloss over that, because it's expensive not to. This is my biggest concern. I'm going to get killed by some jackass' tesla while he's texting, because Elon Musk is a megalomaniac.

Don’t worry, it’s much more likely that you’ll be killed by someone who is texting and driving some other make of car.

Re: Waymo CEO dismisses Tesla self-driving plan: “This is not how it works”

#524

Post by pas »

sudosysgen wrote: post #522
pas wrote: post #383

Earlier quoted context omitted.

But those are needed for L5 anyway, Lidar or noLidar. The big issue is the world model, what's the prior when uncertainty rises, how to spot really out of ordinary stuff, and how to spot really ordinary-looking but dangerous stuff.

I'm fairly certain L5 automation won't need anywhere near the processing capacity of the visual cortex.

The visual cortex part is what current deep learning stuff handles pretty well. What's not so obvious is how to generate a model that can manage the "usual problems" (unclear road signs, debris on road, pedestrians/cyclists, harm minimization in accidents) without "freaking out" (phantom braking).

For example if the car in front did not even slow down while going through some obstacle (the proverbial plastic bag), then it's very likely that we don't have to either. But if something looks like a drunk guy trying to cross the highway, then it's caution time.

That said, without the actual data/studies/access on what Tesla does and how, it's hard to say how good they are at this. (Sure, in theory everything needed to drive a car well can be inferred from regular human vision visuals. But that means their model has to be able to do counterfactual reasoning - I see this thing as a wall, but other cars go through it, so it's not a wall, so maybe it's a marking. Otherwise it'll very hard to just figure out things frame-by-frame and simply by raw object detection.)

Re: Waymo CEO dismisses Tesla self-driving plan: “This is not how it works”

#525

Post by dontreact »

paulryanrogers wrote: post #56
petters wrote: post #49

Earlier quoted context omitted.

Are you aware how low the deaths per mile is for humans? I don't think we know if Waymo has a lower rate than a human yet. Tesla I don't know, but wouldn't be surprised if it is higher.

IIRC for humans it's like 1 death per million miles driven. I believe Waymo is closing in on that. Not sure if Waymo handles the variety of conditions though. Musk and Tesla seem very flippant in their marketing.

It’s not. It’s 1 per 100 million miles

Re: Waymo CEO dismisses Tesla self-driving plan: “This is not how it works”

#526

Post by dillondoyle »

jiofih wrote: post #437
dillondoyle wrote: post #306

Earlier quoted context omitted.

that is terrifying. the next 60 seconds after that is crazy bad. 'that is not ideal'... We need to legislate this before more people die

Nothing scary happens in the next 60 seconds. The car would obviously not go into the opposite lane if there was traffic, and given a few more seconds it would return to the right shoulder. The driver is jumpy and taking over before it has time to correct anything.

what?!

first I dont think that's so obvious that the car would obviously not go into the opposite lane, it doesnt even see all the cars as OP pointed out and Teslas have rammed solid objects stationary at full speed.

Not being able to 'see' standard lane markings is such a basic requirement!

The car shouldn't have to correct.

if this was a student driver during a test I'm guessing they would fail - been a while I don't know the points but there are multiple really bad mistakes.

Re: Waymo CEO dismisses Tesla self-driving plan: “This is not how it works”

#527

Post by Geee »

theshrike79 wrote: post #455
burmanm wrote: post #277

Earlier quoted context omitted.

Is snow an edge case? My cameras don't function when it's snowing or there's snow in the ground. Or there's snow banks around the road. My dash is lit with the "some cameras are blocked" 24/7 for the whole winter time. And proximity readings are blinking those yellow / red lights all the time from the right side as it thinks the snow bank next to the road is too close. Or the system thinks the camera is blocked becau…

Snow is a difficult problem even for humans. In worst cases the road turns from a 2 lane road to a 1.5-1.8 lane road when there's enough snow on the banks. Multi-lane motorways turn from 4 lanes to 2-8-3.5 lanes. Humans have trouble navigating that, dunno how any FSD system would manage it in our lifetime.

Also sometimes the "lanes" that form on the road can cross the center divider to the oncoming lane, which is really dangerous if you're not paying attention. I almost hit an oncoming semi-truck just one day because of this. It's an example of a rare edge case.

Re: Waymo CEO dismisses Tesla self-driving plan: “This is not how it works”

#528

Post by snovv_crash »

screye wrote: post #510
snovv_crash wrote: post #409

Earlier quoted context omitted.

There's still mechanical failure to deal with.

I am uniquely equipped to answer that question. Did an undergrad in mech E and worked as an automobile engineer for while before pivoting to computer vision and ML. The safety standard that mechanical parts need to abide by are orders of magnitude higher than any Vision/Software product. Additionally, mechanical failures are rarely catastrophic. The part will most likely alert you hours/days in advance that it has be…

I have a similar background, elec-mech undergrad, masters in robotics doing SLAM, spent time doing electrical design for industrial usage, then software doing computer vision, now working on drone autopilots.

I 100% agree on needing to design and regulate the design of safety critical CV (and particularly ML based CV) algorithms so that, if nothing else, the failure mechanisms are quantifiable and limited in their impact.

Re: Waymo CEO dismisses Tesla self-driving plan: “This is not how it works”

#529

Post by faisalhackshah »

La1n wrote: post #425
JonathanFly wrote: post #410

Earlier quoted context omitted.

>However, in the unlikely scenario that 2D videos are sufficient for L5 self-driving, Tesla wins BIG. No one will be even close, especially due to the scale of 2D data that they have gathered. Potentially Tesla's 2D data lead means they will still be the leader even if the compute resources in the current cars isn't up to the task, and maybe even if they have to add additional sensors like lidar. They can probably re…

>costs a fraction of the cost of lidar What makes lidar for use in self-driving that expensive? I have a robot-vacuum with a lidar that wasn't all that expensive.

Robot vacuums typically don't use Lidar. They user a laser pointer and an offset camera. Distance is estimated by the pixel offset between the laser dot, and the center of the image.

Object very far away: Laser dot in center of image. Object very close: Laser dot at maximal offset from center of image.

Re: Waymo CEO dismisses Tesla self-driving plan: “This is not how it works”

#530

Post by Marsymars »

necovek wrote: post #493
bartvk wrote: post #452

Earlier quoted context omitted.

Or: look-but-failed-to-see-errors, which are an "interesting" cause of accidents. When I took my motorcycle driver's test, my driving instructor sometimes warned me that I needed to make movements in a particular way. He claimed that even though I would make eye contact with a car driver, they may look-but-not-see-me. His reasoning was, as a motorcycle rider, I'm horizontal/upright when a car driver may be looking fo…

Riding a motorcycle is a tough one for car drivers, and not just because of the issue you mention: bikes can accelerate and brake much more rapidly due to their lower mass, and inattentive drivers can easily be caught by that. Them appearing where it shouldn't be possible for a car to show up also amplifies the issue (you don't need to look over your shoulders in a single lane street, but bikes easily show up there).…

> bikes can accelerate and brake much more rapidly due to their lower mass

Cars are typically able to brake faster than motorcycles. One of the reasons why tailgating on a bike is extremely dangerous.

Post reply on HN