Hacker News Reader

Comment by rectang | Hacker News Reader

The ability to discern when the car needs to stop for a stationary object right in front of it needs to be an ironclad requirement for self-driving.

It seems that sophisticated LIDAR is currently the best way to achieve this, but LIDAR is expensive. So companies like Uber and Tesla skimp on the LIDAR, and build self-driving cars that can more or less follow lanes but plow right into obstacles? Whoa.

The feasibility of widespread "self-driving" tech is going to advance only as the cost of LIDAR falls.

View on HN

I wonder why they can't just use a less sophisticated lidar. Just have a few fixed beams that just look ahead. Do you really need a scanning system to detect an obstruction?

Lidar isn't reliable in conditions that are less than ideal (rain, fog, snow, etc). If you build a car that is dependent on real-time lidar observations, it is worthless much of the time in some climates and some of the time in all climates.

Most companies are pursuing a Lidar approach because dev_speed(Lidar approach) > dev_speed(camera approach). Tesla is pursuing a camera approach because max(camera approach) > max(lidar approach).

In such conditions, cameras are not unreliable, they're downright useless. Heck, even people have trouble driving during snowfall. I have a feeling that this is a massive set of scenarios that aren't even edge cases, yet SDVs are completely unprepared for them. ("No snow in coastal California - it doesn't exist at all!")

I think you may be imagining what a human sees in a foggy camera image, rather than what a neural network can see in it. In fact, atmospheric obstructions degrade Lidar's signal much faster than a camera's.

If you check out NVidia's DriveNet demo from a couple years back, you'll see that they already have a vision-based NN that outperforms humans in fog and snow[1]. We can debate whether people should drive in those conditions at all, but today's expectations will be the baseline that SDVs are up against, and cameras are much better suited than Lidar to achieve that baseline.

[1] Start at 7:21 https://www.youtube.com/watch?v=HJ58dbd5g8g

What human sees in a foggy camera image is something quite different from what we see directly - that's the Uber-scapegoat-camera-human-wouldn't-have-seen-anything-either argument again.

To the effect of "nobody should be driving like that" - I have driven at walking pace at times, considering faster speed unsafe. I do agree that would be a better task for a SDV, iff it can full fill the promise of that marketing video.

I also agree that various conditions are suited to different sensor types - thus for an autonomous vehicle, multiple sensor types are needed.

Reply on news.ycombinator.com