Musk has said repeatedly that LiDAR is too expensive, and also anything that it can do, vision should also be able to do. The war on radar is a newer thing, not sure where that stems from. [I just looked, and he has tweeted that vision should also be able to outperform radar]
LIDAR has certain advantages and vision has different ones. But there’s no denying that using them together is better than only one. Musk’s resistance is almost entirely due to cost (and bulkiness to a lesser extent).
This article is 2 years old but is still a good discussion of why he doesn’t like LIDAR.
This recent article from The New York Times (paywalled) talks about how “metamaterials” would allow for solid-state LIDAR (meaning no moving parts and cheaper to manufacture).
BTW, just how expensive is present-day LIDAR that Musk thinks it’s too expensive?
Tesla has never been fans of LIDAR, supposedly because it is expensive and bulky. However, I have a vacuum cleaner that uses LIDAR, and it works great!
The most recent thing is about radar, not LIDAR. Musk has tweeted several things to the effect of the “coming soon” version of self driving will be a pure vision based system, and new cars will not even have a radar unit. This refers to the forward looking radar that is in the front bumper, and detects objects as far as 160 meters in front of the car.
This page describes the current autopilot sensors on the Tesla. That page may change if Tesla does update their system.
Based on that page the radar looks ahead 160 meters, but the narrow angle forward camera can see 250 meters. If that’s true, then the radar may actually be redundant.
Based on my experience, the radar is the first thing to foul in snow. The cameras are kept clean by the windshield wipers, and the front ones have their own mini defrost coils. The radar has nothing. People suggested putting a hydrophobic coating on the bumper to keep the radar clear, but I found it didn’t do anything. Maybe the coating isn’t meant to go 70 MPH into the wind, and just wore off. Supposedly some cars put a defroster in front of the radar, but Tesla chose not to do that.
I think Tesla’s argument is that there are certain things that LIDAR will always struggle with (fog, reading text on a sign), so if your goal is fully autonomous driving, it will necessarily need to have vision system combined with a smart AI too.
But if you have a vision system and a smart AI, you don’t need LIDAR. So from a long-term perspective, LIDAR is a developmental dead-end. What’s the point in adding it, especially since it adds substantial cost?
There’s a similar argument for removing radar. Radar is terrible if you want to distinguish if that object up ahead is an aluminum can in the middle of the road, a stopped vehicle, or a highway overpass. Sure, it has its advantages, and it’s not that expensive to add, but what do you do when vision and radar disagree as to whether the object up ahead represents a danger?
If you just err on the side of caution and stop the car anyway, you’re going to have a lot of false positives where the car suddenly lurches to a halt in the middle of the highway because of a coke can. If you’re going to always defer to vision, what’s the point of radar?
Vision is much better at identifying objects, but it is not as good at determining distance (particularly in the dark). I agree that when vision (and particularly AI) becomes advanced enough, lidar and radar become superfluous. But we aren’t there today, so the question is what to do for the next 10-20+ years until we reach that point.
What do SDVs do when there’s an object in the lane, whether that’s a road alligator or a box in the lane that fell off of an earlier passing truck that requires them intentionally deviating from their lane to get around said object? What about a cyclist, who is not only moving at x mph but which by law you must give either 3’ or 4’ (depending upon the state) of space when passing. That could mean your car needs to go into oncoming traffic but can’t come back into it’s own lane until it’s moved past the cyclist who could be moving anywhere up to 40+ mph (downhill)?
Decent-ish review article today from the New York Times (likely paywalled), that reflects the growing public awareness over the challenges faced in autonomous vehicle development and how AV abilities have been overhyped by the media, although it doesn’t say much that us Dopers who have been following this thread don’t already know.
IMO the key takeaway from the article:
Page above no longer mentions radar, and it looks like radar has been removed from latest vehicles.
Lyft customers in Miami will be able to hail a robotaxi from Argo AI, an AV startup backed by Ford and Volkswagen, by the end of the year, the companies announced Wednesday. Next year, riders in Austin, Texas, will be able to do the same.
Continuing slow progress.
One sentence from the link contradicts the title of the article;:
So the “self-driving” cars have not just one, but two drivers.
Yes, slow progress. Very slow, I think.
In this thread, two years ago, (post #309), I posted the same comment, about the same issue (two safety drivers required.)
But that was about a company called Waymo, this one is called Argo.
So…different company…exact same problem.
Presumably the second safety driver is there is watch the first safety driver to ensure that he’s attentive enough.
Way back in the 1990s when these technologies were being defined, under ISTEA, this phase was to be labeled co-piloting. I believe the use of the term co-piloting rather than self-driving would have avoided some problems we are observing.
Good news! Elon Musk says that they’ll probably have level 5 figured out by the end of the year. (Again)
“I remain confident that we will have the basic functionality for level 5 autonomy complete this year."
I mean, the canard here is that he can define for himself whatever he means by ‘basic functionality.’ It just won’t mean what most of us mean by it.
Yeah, I’m going to call bullshit on that. There’s no way Tesla will provide any kind of true autonomous driving by the end of the year. Even if that year is 2025. Or probably 2035.
You can probably say that (title of OP - yes, I know I’m many years late in some ways) for pretty much any AI technology that is attempting to access meat space. Car driving is simply the lowest-hanging fruit. AIs can perform exceptionally well when they don’t need meat space sensors, or only need one that’s able to read atmospheric conditions or similar. Ones that need a strong command of the placement and movement of objects in the world around them are going to take a very long time to program correctly, as there is very innate “software” in our brains for doing that, stuff that it much harder to program than chess AI because we do it constantly without knowing how it’s done.
“Green means go. Red means stop.”
…yellow light go very fast.
I don’t think he can define it for himself, not as long as he’s throwing around phrases like Level 5 autonomy, given that has specific meaning.