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

The Camera Is the Lidar

medium.com

Post reply on HN
Print view View original

11–20 of 77 posts

Re: The Camera Is the Lidar

#11

Post by barbegal »

amelius wrote: post #4

The problem is that deep learning should not be allowed in safety critical systems, because (1) the accuracy is always less than 100% even in known test situations, and (2) we don't know how it works and under what conditions it breaks down.

I think it should be allowed but the tests it should pass must be far more strenuous than for traditional software. I'm happy with failure rates of around 1 catastrophic failure every million hours. Even humans sometimes fail catastrophically and black out at the wheel for no detectable reason.

That level of testing is well beyond what today's software and hardware is capable of. Waymo has to override their cars (disengagement) approximately every 6000 miles [1] which equates to about 200 hours of driving. To reach a confidence level of 1 in a million hours you would need a test fleet of a thousand vehicles operating for a whole year without any incident occurring that requires human intervention. The costs for such testing would run into the hundreds of millions of dollars which makes me feel like only the largest corporations in the world could develop this technology.

[1] https://www.dmv.ca.gov/portal/dmv/detail/vr/autonomous/disen...

Re: The Camera Is the Lidar

#12

Post by amelius »

barbegal wrote: post #11
amelius wrote: post #4

The problem is that deep learning should not be allowed in safety critical systems, because (1) the accuracy is always less than 100% even in known test situations, and (2) we don't know how it works and under what conditions it breaks down.

I think it should be allowed but the tests it should pass must be far more strenuous than for traditional software. I'm happy with failure rates of around 1 catastrophic failure every million hours. Even humans sometimes fail catastrophically and black out at the wheel for no detectable reason. That level of testing is well beyond what today's software and hardware is capable of. Waymo has to override their cars (dis…

And after certification, what if the company wants to push a quick update to all of its cars every now and then, through a remote update? Would that be allowed? How would we even know that it happens?

Re: The Camera Is the Lidar

#13

Post by maxerickson »

amelius wrote: post #12
barbegal wrote: post #11

Earlier quoted context omitted.

I think it should be allowed but the tests it should pass must be far more strenuous than for traditional software. I'm happy with failure rates of around 1 catastrophic failure every million hours. Even humans sometimes fail catastrophically and black out at the wheel for no detectable reason. That level of testing is well beyond what today's software and hardware is capable of. Waymo has to override their cars (dis…

And after certification, what if the company wants to push a quick update to all of its cars every now and then, through a remote update? Would that be allowed? How would we even know that it happens?

Just make update transparency part of the initial certification.

Of course they might cheat, but who really knows if their car has an airbag where it is supposed to.

Re: The Camera Is the Lidar

#14

Post by barbegal »

I'm sure this works well in bright light but I'm sceptical that this can perform at all well on overcast days or at night. The OS-1 device spins at 10Hz and the LIDAR samples 2048 points over one 360 degree revolution. This means each column of pixels is sampled at 1/20480 of a second. To sample that fast requires a lot of light which is fine on a sunny day but on a cloudy day you can have 100 time less light. And at night you would have no ambient near infrared light at all.

Re: The Camera Is the Lidar

#15

Post by haney »

So I'm not up on the Lidar industry but $12k for a sensor seems really expensive, but then again from my casual observations Lidar is really expensive. Is there a physical / first principles reason this is true or is it just really new technology?

Re: The Camera Is the Lidar

#16

Post by deepnotderp »

amelius wrote: post #4

The problem is that deep learning should not be allowed in safety critical systems, because (1) the accuracy is always less than 100% even in known test situations, and (2) we don't know how it works and under what conditions it breaks down.

You do realize that traditional computer vision doesn't work 100% of the time, right?

What would you rather use to interpret sensor data?

Re: The Camera Is the Lidar

#17

Post by wpietri »

21 wrote: post #5
thanatropism wrote: post #2

Is it just me, or are we seeing yet another impressive leap in Computer Vision that's soon going to be hyped as an incremental step into Skynet?

Unless you think Skynet is impossible, all impressive leaps are an incremental step towards Skynet.

Not necessarily. My expectation is that Skynet is highly unlikely, a side branch we probably won't take.

Think of the 1920s-1950s version of robots, for example. They were machines shaped like people and that acted like people. In retrospect, they seem not scary but silly. The human shape isn't particularly useful or easy to build; our most common robots are vacuums shaped like hockey pucks.

Skynet is another "what if machines acted like people" fairy tale. It makes sense if you imagine yourself as a computer that wakes up; we wake up all the time, so it seems normal to us. But self-awareness and self-preservation are biological systems that evolved over very long time scales. Those are intricate systems, again not really useful or easy to build. And also not likely to randomly occur.

It could be that we'll build those kinds of systems, of course. But I think it will take a long time to get them right, and then it's not really the skynet story, it's the mad scientist with the robot army story.

Re: The Camera Is the Lidar

#18

Post by Animats »

The hardware is described, vaguely, at [1]. It's a rotating drum scanner with 16 or 64 lasers. $12,000 for the 64-laser model. $24,000 for the upcoming model with 200m range. Still a long way from a useful auto part.

There are about a dozen companies in this space now. Nobody has the price down yet. Continental, the European auto parts company, is the most likely winner. Quanergy made a lot of noise but didn't ship much.[2] There's a conference on automotive LIDAR this month in Detroit.[3] Many of the exhibitors are major semiconductor packaging companies, with various approaches to putting lots of little LIDAR units in a convenient package at a reasonable price.

[1] https://www.ouster.io/faq/ [2] https://news.ycombinator.com/item?id=17755183 [3] http://www.automotivelidar.com/

Re: The Camera Is the Lidar

#19

Post by wpietri »

haney wrote: post #15

So I'm not up on the Lidar industry but $12k for a sensor seems really expensive, but then again from my casual observations Lidar is really expensive. Is there a physical / first principles reason this is true or is it just really new technology?

I'm not either, but I think there's no reason for them not to charge a lot of money. I expect goal is to advance their tech as rapidly as possible while claiming mindshare. The people buying this stuff are generally swimming in investor money, perceived value is often related to price, and it's much easier to lower prices than raise them. So as long as they're selling enough units to get useful real-world feedback that supports their development, in their shoes I'd basically gouge people.
Post reply on HN