I don't understand how people keep falling for this. Sure, it seemed realistic enough at first but how many cracks in the facade are too many to ignore? In 2015 Musk said two years. In 2016 he said two years. In 2017 he said two years. Tesla did a 180 on the fundamental requirements of FSD and decided it doesn't need lidar just because they had a falling-out with their lidar supplier. That level of ego-driven horseshit is dangerous.
Look, the message is clear enough already: FSD is two years away, and always will be! You've got to admire that kind of consistency.
• In January 2021 he said end of 2021 [2]. • In May 2022 he said May 2023 [1].
So he moved from around 2 years to 1 year.
Looking at the rate of progress from Tesla FSD videos on YouTube, I wouldn't bet on 2023.
Whenever people talk about Tesla FSD, I always like to point to Cruise who just got a permit to carry paying riders in California [3]. Honestly, their software looks significantly smarter than Tesla's -- Highly recommend their 'Cruise Under the Hood 2021' video for anyone interested in self-driving![4]
[1] https://electrek.co/2022/05/22/elon-musk-tesla-self-driving-...
[2] https://www.cnet.com/roadshow/news/elon-musk-full-self-drivi...
[3] https://www.reuters.com/business/autos-transportation/self-d...
I've been a believer in Tesla for a decade now but it is starting to seem like the competition is catching up.
Every successful fraud has people it's tuned for. For example, consider how terribly written most spam is. That selects for people who are not fussy about writing. Conversely, a lot of the people doing high-end financial fraud is done by people who are very polished, very good at presenting the impression of success. Or some years back I knew of a US gang running the Pigeon Drop [1] on young East Asian women in a way that was tuned to take advantage of how they are often raised.
Telsa's only has ~3% of the US car market, so they're definitely in the "fool some of the people all of the time" bucket. Musk's fan base seems to be early adopters and starry-eyed techno-utopians [2]. He's not selling transportation. He's selling a dream. They don't care that experts can spot him as a liar [3] because listening to experts would, like, totally harsh their mellow.
Although it's much closer to legal fraud, I don't think that's otherwise hugely different than how many cars are marketed. E.g., all of the people who are buying associations of wealth when they sign up for a BMW they can't afford. Or the ocean of people buying rugged, cowboy-associated vehicles that never use them for anything more challenging than suburban cul de sacs.
[1] https://en.wikipedia.org/wiki/Pigeon_drop
[2] https://www.theverge.com/2018/6/26/17505744/elon-musk-fans-t...
[3] e.g.: https://twitter.com/kaifulee/status/1126238951960993792
I am cautiously optimistic about future of FSD in general now.
Also, by “as good or better than a human”, for me the biggest things are:
- Involved in the same number, or fewer, accidents
- Does not piss off other drivers any more than a human (like stopping way back from stop signs, getting “stuck” on a decision and not making progress, etc.)
In this vid the Tesla was nowhere close to the above standard. Still really impressive, but lots of work to do. Hard to say how close it is - maybe a few quarters away, but could also be a decade or more.
Google/Waymo is closer in terms of safety and pissing off other drivers, but it’s also much more conservative from the rider’s point of view. Like it will do 3 right turns to avoid a tricky left, will take side streets over the highway, etc. Waymo vids seem much safer and more predictable, fewer clear bugs (e.g. all the times the Tesla FSD makes the wrong decision on where to stop at an intersection are bad, “hard dealbreaker” bugs, that Google/Waymo don’t have), but I think it would just get you to your destination too much slower than a human driver, so buyers wouldn’t like it.
But it misses that the kinds of erros AI can make are so wild, other drivers and pedestrians can't even imagine or anticipate them. Imagine all kinds of different responces the AI might trigger on sudden appearence of a car, where a dog should be.
Human drivers could be drunk, but they don't go from perfect driving to batshit crazy in a split second.
No basic mistakes. No shut downs in the middle of the road. And is approved for use as a taxi service.
And you clearly see the benefits of LiDAR by the stability and consistency with which it identifies other objects e.g. cars, pedestrians on the road. The inability of Tesla FSD to accurately identify the bounding box of the truck at 6:06 is extremely concerning for example.
Yeah, phantom braking is an absolute plague when it comes to systems that depend on cameras or radar.
I doubt the same issue would appear with stereo depth.
(You need to solve vision anyway, because for that object, of which LIDAR tells you is exactly 12.327 ft away, you still need to figure out whether it is a trashcan or a child. And if it is a child, whether it is about to jump on the road or walking away from the road. LIDAR does not tell you these things. It can only tell you how far they are away. It is not some magical sensor which Tesla is just too cheap to employ.)
Why? First of all don't drive into anything that moves, period. Secondly, if you can avoid bumping into anything at all, do that.
That is not to say Lidar doesn't have its issues (and there are quite a few), we likely will need a combination of sensors including cameras, lidar and radar.
Tesla doesn’t use their radar anymore in any case. Only monocular cameras.
Every self driving car company (except for Tesla) uses both LiDAR and vision.
But they all rely on LiDAR for bounding box detection which is exactly the main problem Tesla FSD has.
You've retreated from this as the reason that leobg is frightfully ignorant on behalf of cycomanic. That means the next most contentious claim is the claim that builds upon the foundation of the first claim - that vision being required makes LiDAR irrelevant. The problem for you though is that when you make the concession that vision is necessary, you run into a problem. The sensor situations under which LiDAR is much better than vision tend to involve a vision failure through a lack of light or due to heavy occlusion. There is definitely and necessarily a cut off point at which leobg's claim becomes somewhat true. This denies the right to call him ignorant, because the law of charity demands that his point be the thing that maximizes the truth of his comments. So the claim of ignorance - which amounts to a character attack - becomes unjustified.
>of which LIDAR tells you is exactly 12.327 ft away, you still need to figure out whether it is a trashcan or a child. And if it is a child, whether it is about to jump on the road or walking away from the road. LIDAR does not tell you these things.
That is ignorant, because LIDAR together with processing obviously can tell you if the thing is a trashcan or a child. The post by is ignorant, because to my understanding it implies that LIDAR does not provide enough information to make that determination, which is untrue and not how LIDAR works.
Now if they mean we still need some way to process this information and make decisions of what the different things are, that's a bit disingenuous because that's completely orthogonal to LIDAR vs cameras vs RADAR and using that argument we could dismiss any of the other technologies ignoring the fact that more (and different) data typically allows you to make better decisions.
Bellman gave us the bellman equations, but also gave us the term curse of dimensionality. The equations he shared and the modeling problems he encountered are fundamentally related to modern reinforcement learning. More data doesn't come without cost. So often I hear people speak of the introduction of lower resolution as equivalent to the introduction of error, but this is a false equivocation. Latency in decision making means the introduction of a lower resolution can increase the resolution error, but still decrease the solution error. This is so fundamental a result that it applies even to games which don't have latency in their definition. Consider poker. The game tree is too big. Operating with respect to it as an agent is a mistake. You need to create a blueprint abstraction. That abstraction applied to the game introduces an error. It is lower resolution view of the game and in some ways it is wrong. Yet if two programs try to compete with each other, the one that calculated with respect to the lower resolution version of the game will do better than the one that did its calculations with respect to the higher resolution view of the game. High resolution was worse. The resolution without error was worse. Yet the game under consideration was orders of magnitude simpler than the game of a self driving car.
I've been paying some attention to this debate and I'm not convinced yet that the situations under which LiDAR is superior are sufficient. I think we agree on that already. For me, this reduces the set of situations under which LiDAR is able to be considered superior - if vision is bad, but you need vision, then better to avoid the situation then use the wrong thing [1]. So the situations under which LiDAR becomes superior becomes a subset of the situations that it is actually superior. That subset doesn't seem very large to me, because both LiDAR + vision and vision alone are both necessarily going to be reducing the dimensionality of the data so that the computation becomes more tractable.
[1]: This isn't exactly uncommon as an optimization choice. It'll get dark later and you'll stop operating for a time. Then light will come. You'll resume operation. This is true across most of the species on this planet. If you are trying to avoid death by car accident you could do worse than striving to operate in situations where your sensors will serve you well.
[1] https://www.researchgate.net/figure/LiDAR-point-cloud-a-high...
Which Luminar has solved for Volvo:
https://www.forbes.com/sites/samabuelsamid/2021/06/24/next-g...
https://techcrunch.com/2022/01/05/volvo-partners-with-lumina...
No you don't. You just need to avoid hitting it.
The problem with a vision-only system is that you need to know what an object is to determine the bounding box and thus how to avoid hitting it. Which is the problem we've seen with FSD where if you combine two options e.g. a boat on a truck then it gets confused because it hasn't been trained on it yet.
If you can accurately determine the 3D geometry of the scene(e.g. with LiDAR), the 3D object detection task becomes much easier.
That being said, most tasks for self-driving such as object detection can be robustly “solved” by LiDAR-only (to the extent that the important actors will be recognized) but adding in cameras obviously helps to distinguish between some classes.
Trying to do camera only (specifically monocular) runs into the issue of no 3D ground truth meaning it’s a lot more likely to accidentally not detect something in frame (say a white truck).
That’s why you can have LiDAR and partially-“solved” vision but need fully solved vision if it’s the only input.
Their Ai confuses a dog and a car. If confuses a parked semi and an underpass. It drove into a solid concrete wall on a few occasions.
Even when they solve vision, that does not tell you what a pedestrial intends to do - even another human can't always be sure.