Comment by nlitened
4 hours ago
As I understand, lidars don't work well in rain/snow/fog. So in the real world, where you have limited resources (research and production investment, people talent, AI training time and dataset breadth, power consumption) that you could redistribute between two systems (vision and lidar), but one of the systems would contradict the other in dangerous driving conditions — it's smarter to just max out vision and ignore lidar altogether.
> lidars don't work well in rain/snow/fog.
Neither do cameras, or eyeballs.
When it's not safe to drive, it's not safe to drive.
I've been in zero-road-speed whiteout conditions several times. The only move to make is to the side of the road without getting stuck, and turning on your flashers.
Low-light cameras would not have worked. Sonar would not have worked. Infrared would not have worked.
I think the weather where cameras/sensors start having problems is much better than zero-vis whiteout.
If we could make sensors that lets an autonomous vehicle drive reliably in any snow/rain where a human could drive (although carefully) then we're good. But we are a long way from that. Especially since a lot of sensor tech like cameras tend to fail in 2 ways, both through their performance being worse in adverse condition but also simply failing to function at all if they are covered in ice/snow/water.
Radar might still have worked
1 reply →
If you have multi-return lidar, you can see through certain occlusions. If the fog/rain isn't that bad, you can filter for the last return and get the hard surface behind the occlusion. The bigger problem with rain is that you get specular reflection and your laser light just flies off into space instead of coming back to you. Lidar not work good on shiney.
Why does this matter? You have to slow down in rain/snow/fog anyway, so only having cameras available doesn't hurt you all that much. But then in clear weather lidar can only help.
If your vision is good enough to drive in rain/snow/fog, you don't need lidar in clear conditions. If you planned to spend $10B on vision and $10B on lidar — you would be better off spending $20B on better vision.
We have actual proof this isn’t true. Waymo is light years ahead of Tesla despite spending less.
Tesla is spending upwards of $6B/year to Waymo’s $1.5B. Only one of these companies makes an autonomous robotaxi that’s actually autonomous.
1 reply →
> If yo[u can] drive in rain/snow/fog, you don't need lidar in clear conditions
Of course you do, you're driving at much higher speeds and so is the surrounding traffic. You can't just guess what you might be looking at, you have to make clear decisions promptly. Lidar is excellent in that case.
Nothing works perfectly in all conditions and scenarios. Sensor fusion has been the most logical approach now, and into the foreseeable future.
Computer vision does not work exactly like human vision, closely equating the two has tended to work out poorly in extreme circumstances.
High performance fully automated driving that relies solely on vision is a losing bet.
No, it isn't "smarter." Camera-only driving is the product of a stubborn dogmatic boss who can't admit a fundamental error. "Just make it work" is a terrible approach to engineering.
Can hatred of Musk not derail this entire thread please? I have a camera-only ADAS that I think works quite well, but having both would be better.
Criticism of Musk isn't hate of Musk. The point is completely valid and the results of this management style infuses all of his businesses albeit with differing results.
It's significant that a truly hard problem like autonomous driving doesn't respond to a "brute force" management style. Rockets aren't in this category because the required knowledge and theory is fairly complete, whereas real autonomous driving is completely novel.
Shoe, meet foot.
1 reply →
Why does that strategy absolutely require the lidar to be absent from the car? When was less technology the solution to a software problem?
Limited resources? Billions per year are being thrown at the base technology. We have the capital deployed to exhaust every path ten times over.
Even if so, it doesn't mean that capital deployment efficiency and expected payoff make equal sense in all directions.
Then again, it's good that we have self-driving companies with lidar and without — we will find out which approach wins.
The Swiss cheese model would like to disagree.
When you have sensor ambiguity sounds like the perfect time to fail safely and slow to a halt unless the human takes over.
Do cameras work well in those conditions? Nope. Also cameras don't work well with certain answer of glare, so as a consumer I'd rather have something over-engineered for my safety to cover all edge cases...
Evidence clearly shows otherwise.
Also, military sensor use shows the best answer is to have as many different types of sensors as possible and then do sensor fusion. So machine vision, lidar, radar, etc.
That way you pick up things that are missed by one or more sensor types, catches problems and errors from any of them, and end up with the most accurate ‘view’ of the world - even better than a normal human would.
It’s what Waymo is doing, and they also unsurprisingly, have the best self driving right now.
This is silly. Cameras are cheap. Have both. Sensors that sense differently in different conditions is not an exotic new problem. The kalman filter has existed for about a billion years and machine learning filters do an even better job.
Cameras are cheap, but, as I understand:
1) it's not cheap to produce lidars at a stable predictable quality in millions;
2) car driving training data sets for lidars are much scarcer (and will always be much scarcer due to cameras' higher prevalence) and at a much lower quality;
3) combined camera+lidar data sets are even scarcer.
1. Automotive LiDAR is down to $350 in China already. BYD is starting to put LiDAR in even entry level cars. (It's been in their mid and high end cars for a while).
2+3. BYD collects extensive training data from customers, much like Tesla does. They will have no trouble with training.
Doesn’t that make it a sensible long term play to equip your car with $200 LIDAR and start gathering that data as a competitive advantage?
1 reply →