Hacker News Reader

Comment by amelius | Hacker News Reader

LIDAR makes implementing FSD so much simpler. And you can still harvest data so you can use it as a stepping stone for video-only based FSD.

View on HN

Yeah, I think if we want driverless taxis in major cities ASAP, then Waymo is probably the way forward. It should be relatively easy to continuously map a city with a fleet of vehicles and keep them geofenced. It's will be a good revenue generator that can sustain the company while they work on a more general video based solution.

The counter argument is lidar isn’t strictly necessary (case in point humans use vision, and roads are designed for visual drivers). Furthermore, there is potential added complexity in fusing lidar and video data. When lidar says the path is blocked and video says it is not, you essentially have to either trust lidar and stop unnecessarily or just ignore lidar and go anyway (and then what was the point of the lidar?). Having to make a decision like this is a form of complexity and addressing that complexity potentially takes resources away from iterating on the vision part of the stack.

> When lidar says the path is blocked and video says it is not, you essentially have to either trust lidar and stop unnecessarily or just ignore lidar and go anyway (and then what was the point of the lidar?).

If it's actually blocked then decision becomes stop and prevent harm or go forward and hurt someone.

I don't think it's a very good counter argument.

Edit: without meaningful data as to which is more meaningful I'd lean towards LIDAR being safer because there's less info (depth) to infer, and to me color seems less important than depth, barring stuff like signs and lights. Which shouldn't be hard to correlate between the two.

But without more meaningful data this comes off as a "you should x people people did x anciently and it seems like that worked"

That's a very important difference. If it turns out that LIDAR (and maybe other sensors) are needed in addition to machine vision for level 5 automation, all the data Tesla cameras have been gathering may be worth much less than the 3D models Google is building.

Waymo is doing level 4 automation. Level 5 is still very far away, and will require more than just LIDAR.

https://www.synopsys.com/automotive/autonomous-driving-level...

> Level 5 is still very far away, and will require more than just LIDAR.

I don't think that's true. In my view, by scaling these AI models we will arrive there. But even if I am wrong, we don't need level 5 for this technology to be revolutionary. I am mostly interested in using Waymo as my new Uber, so if it can decently navigate cities costing half the price, then sign me in.

The problem is the definition of level 5 is amorphous.

That is, "everywhere in all conditions" exceeds what any human will do. If I won't drive in a blizzard or specific cities, does it make a car level 4 if it punts under these conditions?

I think we need a better definition that encompasses the different capabilities of things calling themselves level 4, and avoids an impossible standard for level 5.

I will be happy to consider a car to be "level 5" if it can drive to Lake Tahoe through a snowstorm and then park itself in a specific spot of a ski resort parking lot by following the vague and confusing hand signals given by the lot attendant.

LIDAR _and_ high resolution/detail 3D scans known ahead of time.

Reply on news.ycombinator.com