Here's my useless opinion.
Ok, so lidar (I refuse to capitalize it anymore, we don't capitalize radar) is a big umbrella for a whole lot of technologies, some are really quite good some are really quite bad. The better ones tend to be more expensive, involve moving parts, and be hard to keep well calibrated. So a lot of relative performance discussion depends on what class of lidar is being discussed, but...
LiDAR should not be used for vehicle operations as it:
- generally do no better than optical cameras through fog, snow, heavy rain
They do marginally better in dust and fog because the infrared wavelengths penetrate better. While these phenomena will scatter the laser beam and impact diffuse its return, for a camera these will scatter all ambient light back to the camera which is worse.
One advantage of lidar is that it can provide multiple returns from a single laser pulse-- so you might get a return from the raindrop is passed through, and then a second return from the object behind it. Cameras accumulate all the reflected light together with no real indication of what depth the light returned from.
Radar, with a much longer wavelength, does tend to penetrate weather better (with the tradeoff that it also provides much less lateral resolution).
- cannot differentiate colors
True, but color isn't always as useful as one might think.
- vulnerable to external light sources (especially LEDs and other LiDAR sources)
Less so than cameras are. LEDs aren't really special in this scenario. A good lidar scans, and only detects light in the direction that it just sent the light, so it would be unlikely for another lidar to send a pulse at the exact moment that you're looking right at it and if it did it would only be one bad point in the cloud. Good lidar also use very narrow bandwidth optical filters to only receive light of the wavelength that it transmits. Then it filters on all kinds of parameters. Yes, external light sources can interfere with lidar, but it's not nearly as susceptible as a camera is.
Radar has a similar vulnerability but it's just at a different wavelength.
- limited range compared to radar
- falls under the same weakness of diffuse scattering (x2 out and back ) as radar
Range is a limitation for all of these sensors. Cameras can essentially see 2D to the horizon, but only if there is light to detect and cameras lose the ability to resolve depth at much closer range than lidar does. Lidar range is limited mostly by eye safety regulations. Radar range is limited by FCC spectrum regulations.
LiDAR can offer superior vision is low or zero light scenarios; although this has proven unreliable in driving scenarios where there can be hundreds of external lights hitting the LiDAR sensors. It is also adversely affected by bright sunlight situations. It is possible that LiDAR would not have done any better vs the white box truck on road with bright background.
The white box truck is actually the kind of scenario a lidar would be excellent in resolving. Cameras will struggle to resolve the edges of a backlit object, and if everything is white and featureless cameras can't measure depth because they can't generate stereo correspondences. Lidar would paint it with an IR pulse and know there is something there and how far away it is.
The benefits of LiDAR do not outweigh the overall system complexity. I am definitely NOT saying LiDAR by itself is complicated, it really is not. I'm only saying (typing) that 'adding another sensor' is not the solution.
I disagree. Adding sensors that help compensate for each others' weakness is the right approach. That's the approach taken on any safety critical system-- diversity and redundancy. I also wouldn't advocate removing the cameras.
I'm pretty sure that Tesla's point of view on this isn't that lidar is bad-- I suspect their point of view is that lidar is expensive. I agree with them on that. People, even luxury vehicle customers, won't pay what those Waymo or Cruise vehicles cost-- if you want to take autonomy to the mass market you need to get the price down. The question is how.
Most of the industry is working to get the cost down by cost reducing the sensors. Tesla is hoping they can sidestep that effort and eliminate the sensors, and provide a good-enough solution without the extra hardware. That's a big gamble, and not one I'm convinced will pay out.
First, they need to develop the tech to work with less rich information which is a harder problem and will take more effort and likely time to complete. Then they need to provide a product that regulators are willing to put on the road first as a prototype but then as a consumer product. Automotive regulations are very conservative on safety and very often demand diversity and redundancy-- they also demand predictability in software execution which the Tesla black box AI model isn't going to provide. So Tesla is banking that regulators are going to bend either under the weight of evidence from field tests or from public or political pressure. Finally, they need to make a product that technically ignorant customers will trust-- and customers like that tend to focus on consensus views. So if everyone in the industry is saying you need more sensors and Tesla is the lone voice saying otherwise, it's going to take some effort to pull more than the Tesla fanatics over.
When the price of the sensors comes down, I'm sure we'll see multisensor Teslas in the future just because they'll need to compete on features. I guess I can't really be sure. I was sure Apple would eventually sell a multi-button mouse... but for the stubborness of one man. This has the same feeling of being a point of principle.
We are still a ways out on creating autonomous vehicles that operate equal the average population; but that is not enough, they need to be far superior.
I tend to agree. Musk is saying they're going to deploy unsupervised FSD and be producing cybercabs before 2027, but I think that's just going to support their field tests all of which are in Texas and California (and probably Arizona eventually). There's a lot of ground to cover between field testing in Texas and navigating Boston winters well enough to remove the steering wheel and get regulatory approval. I'd be shocked if anyone is selling a true autonomous vehicle to consumers before the 2030s. Maybe if they geofence it and rely on state level approvals the can sell something sooner, but that'll still be niche compared to the low cost EV Tesla apparently cancelled to focus on this.
I also agree that they need to be superior to humans to succeed. Really they should just need to match humans, but people are weird-- they expect perfection from machines in a way they don't expect it from themselves. People are afraid of flying even though its much, much, much safer than driving, because they feel they're not in control. I don't know what the current FSD accident rate is relative to humans for the same number of miles, but every single Tesla wreck gets reported in the news like it's a plane crash. The public has a high bar for this stuff.
