Tesla FSD head to head against Mercedes Driver Assist on the same road [video]
youtube.com
youtube.com
The Consumer Reports comparison at the beginning of the video is Mercedes "Driver Assistance" vs Tesla "Autopilot". These are technologies for lane assistance, speed limit assistance, etc. "Advanced cruise control" is a better term for these. This video does not compare these.
Mercedes self driving system is called "Drive Pilot". Tesla's is called "FSD" (Full Self Driving)
This video compares FSD to Drive Assist, these are two different types of technologies.
You are really understating this. This account is one of the few accounts that Elon directly responds to on a regular basis on X.
> 2. They were using an non-geofenced level 2 highway adaptive cruise system for Mercedes, NOT the level 3 system which is only available in Germany. The level 2 system was never designed for city streets and have limitations on sharp turns, if this would an honest video they would compare autopilot vs drive pilot on the highway, which they won't do because drive pilot is better.
>> Mercedes-Benz plans for further customer deliveries of DRIVE PILOT equipped MY2024 EQS Sedan and S-Class models in early 2024 through participating authorized Mercedes-Benz dealers in California and Nevada
Speaking from personal experience, I've tested the most recent version of Supervised FSD, v12.3.6, and it's much better than most people realize.
Once Tesla stops requiring human supervision, I expect FSD adoption rate to go "hockey-stick."
---
[a] Based on a conversion rate for Tesla "Supervised FSD" free trials of ~2%: https://old.reddit.com/r/RealTesla/comments/1cp5pgj/tsla_fsd...
Very significantly improved over the last few updates.
The Mercedes driver seemed to be engaged and attentive, and therefore not likely to play a video game, watch a movie, and stop paying attention causing a crash.
It would be interesting to see Mercedes Drive Pilot (SAE Level 3) compared to Tesla FSD. That feature is not generally available in the US.
Tesla has the most training data consisting of video input + corresponding driver behavior. And they have the most compute power.
Which twist of fate could happen, so that Tesla falls behind in the future?
I could imagine that there is an invention which lets AI learn about the world by just observing and thinking. Without the need for "training" in the form of "if you see this, do this". But as far as I know, this is not on the horizon yet.
Here's why Lidar makes sense. You want independent information streams that are cross checked. If both streams tells you there is no object in front of you, that's a way lower probability of being wrong than just one stream telling you that. You go from probability p to p^2 of being catastrophically wrong. It's the kind of probabilistic added safety of being in an aircraft that's capable of flying with only one engine. Even if one blows up with probability p, you're fine. You need a p^2 event for both engines to malfunction, assuming errors aren't correlated, which they probably are a bit, but not so much that the added benefit isn't massive.
Self-driving, as a mission critical endeavor, needs to be leveraging the same statistical phenomenon in order to reduce the exceedingly rare but highly impactful mistakes.
And by "something it didn't see" I'm talking about CNN/ViT mistakes/hallucinations. I don't see any evidence that it occurs at a level that's much smaller than humans. Any safety data we have is with human oversight. So the majority of errors are caught by humans and won't show up in FSD crash frequency or other publicly available data.
Uh... why? Shouldn't the goal be to be safer than people?
Basically that's just bad engineering: If your argument is not to deploy an existing system that is better than the current installed base because some other technique might be better still, you are harming your metric (safety) and not helping.
If Tesla had one that's safer than humans, then yes they should be allowed to deploy it. My argument was that they're making it harder for themselves to exceed that safety threshold by insisting on vision only.
Waymo and Cruise may have beat them to "first robotaxi" but I'm not convinced that they have anywhere near a scalable, sustainable model.
The advantage of machines is not that they have better brains than humans, but that they have better senses than humans, which is an advantage that goes out the window when you obstinately refuse to use better sensors than merely cameras "because that's what humans use".
I work on a synthetic aperture radar system that is high-resolution enough to "see" the painted stripes on roads through fog and a thin layer of snow.
It's not for automotive use and would increase the price of every car by several hundred thousand dollars but a fusion of multi-spectrum sensors should be the direction we are headed-- not a minimally-viable mono-sensor system.
In other words, failing to realize that raising minimum requirements (and therefore prices) makes it unaffordable for an increasing number of people.
Human-parity (in terms of accident rate, not failure mode) seems a reasonable minimum bar.
For proof I offer the price of the 1969 Volkswagen Beetle, the least expensive new car for sale in the US market for almost its entire sales history: $1800. That’s for one with zero options. A rolling chassis with four seats and a motor.
That’s $16k today. Coincidentally the same price of the Nissan Versa or Mitsubishi Mirage: with backup camera, air conditioning, and airbags.
People don’t WANT the cheap cars, though.
I know the average price of a new car has exploded.
That is a conscious choice by the consumer.
When production of autonomous tech scales it won’t increase costs as much as people assume.
Blanketing the frequency slice with echos seems like a non-trivial problem. Or would that help?
I’ve seen the “human-like” behaviour of FSD 12.x praised a lot by channels like this, particularly where the car is breaking the rules in a way they consider “normal”. And it’s a fair argument that predictable behaviour improves safety.
However, behaviour that is common in the US - like making a turn into a side street while a pedestrian is beginning to cross - would be considered exceptionally aggressive and reckless here in Australia. It’s a cultural difference I’ve adapted to when moving back and forth.
At the end of the day though, when I walk across a street, I don’t want to have to worry if Tesla has fine tuned their model correctly to match our local expectations of yielding. I’d rather they just followed the law as closely as possible - because that’s the most predictable behaviour of all.
I mean, let's be blunt: the "LIDAR vs. cameras" debate has been settled at this point, and the fancy sensors lots. Of all the things one can complain about with FSD, sensor fidelity is not one. Teslas don't hit things. The problems remaining to be solved are in the planning regime: my car still misses turns at a rate somewhat higher than I do, usually because it's in the wrong lane (and that often because it thought that the backup in the turn lane was something it could go around).
Compute power is just money, that is about the thinnest advantage you can have.
The truth is no one has the physical AI algorithm yet. Having cars equipped with cameras, big GPUs connected to can bus is the easy part.
Tesla has some advantage since they are spending the most on automation compared to other car makers.
However compared to Waymo, Tesla is quite behind. Waymo has done a million rider-only rides.
Tesla isn’t even certified to do driverless rides. The driver has to be at the wheel and all accidents are the driver’s fault.
Tesla also doesn’t take safety as seriously as Waymo. So Elon may be a net negative to the org.
If the competent Tesla engineers got similar car hardware to Waymo, perhaps they could have been competitive. LiDAR solves a whole bunch of vision problems and gives extra high quality signal for processing.
But SpaceX didn't have infinite amounts of capital during the years it developed Falcon 9 and Dragon. Until Tesla's market cap blew up during the COVID-19 era, Elon Musk had a "mere" few tens of billions of dollars. ULA's pockets were and are gigantic, too.
In any case, infinite capital guarantees absolutely nothing. Jeff Bezos has been among the world's wealthiest men for far, far longer than Musk's entry into that group. Let me paraphrase an excellent comment I saw on Reddit, in response to one of the usual lies about how the only reason SpaceX is a decade ahead of the rest of the world is that it got zillions in subsidies from the US government:
>If large amounts of funding is the only thing required to succeed, Blue Origin would now have a nuclear-powered spacecraft orbiting Pluto.
Mercedes can get them. So can any major car maker. Making cars is extremely capital intensive.
https://www.statista.com/statistics/502208/tesla-quarterly-v...
And I would expect that those generate more data than they can crunch, so that the bottleneck is the number of GPUs they have.
They recently reported that they have the equivalent of 40k H100 GPUs:
https://www.autoevolution.com/news/elon-musk-explains-the-la...
they will be using 85K H100s (up from 35K now) for training by the end of 2024:
https://www.shacknews.com/article/139619/tesla-ai-training-n...
> This vastly expanded the FSD testing pool, with the Tesla fleet adding 1 billion FSD miles every 2-3 months
It doesn't mean they collect the video feed and all input from every second of those miles. Just that the FSD drives those miles.
There's also a matter of deduplication where they really don't need to use ~250 days a year of the same commute. They probably have more data than others, but I'm asking for a proper confirmation of this common knowledge.
Seems like that is the real debate. Tesla weasels and whines, Mercedes steps up.
If Tesla (or any other manufacturer) actually has faith in their implementation, they will accept legal liability for its actions.
Contrapositively, if Tesla (or any other manufacturer) do not accept legal liability for the actions of their autopilot, then they don't have faith in its implementation.
Everything else is irrelevant snake-oil marketing fluff. Put your money where your mouth is.
It's mostly a stunt, basically. It's an answer to "What's the fastest path to SAE Level 3 in any context", not "How to engineer an autonomous driving solution?"
Mercedes is incrementally improving a system the exact way you’re supposed to engineer such a system. It is not an abuse of the SAE Levels framework, it is exactly why the levels are what they are! It is Tesla that is pulling off stunts.
The NHTSA reported in April that most fatal FSD crashes involved no driver intervention, or driver intervention only in the last 1 second, and in all cases an attentive driver could have perceived and taken steps to lessen the severity of damage and injuries in the crash.
For anything less than SAE Level 4, the human MUST take over driving when requested. The question is, can we? Sign up for the beta test and find out, I guess!
Personally I would much prefer a new Mercedes E class wagon with Driver Assist (Distronic) to a Tesla with FSD.
Drunks and the drowsy. Sure, they should have called an Uber or chosen not to drive at all. But the road is safer for having that multi-thousand pound hunk of steel steered by a more competent and attentive driver (FSD)
I would absolutely prefer any driver using FSD, but especially one physically and mentally inhibited by drugs or alcohol.
> I include 'thinking the Tesla will just handle it' as under the influence.
I've commented on this before, but: no one using FSD for even a short period of time is "under the influence" of the branding. The lived experience of the technology is that is amazing and useful, even if it flawed. YMMV
Also, a little funny to see a random Model Y appear just ahead of the Mercedes at the end of the video.
It seems like Consumer Reports is comparing Mercedes Driver Assistance with Tesla Autopilot and then the video goes on to compare it to Tesla FSD.
Obviously, the average person doesn't know there's a difference and may be misinformed by CR to believe that Tesla sucks, but we can acknowledge here that Tesla Autopilot is very meh compared to FSD, and it's entirely feasible that Mercedes has a justified better rating between these systems, even if FSD is light years ahead.
Tesla may have been first to seriously try but currently there's no evidence they will be first in anything compared to their competitors.
Waymo works because they have mapped out cities with maps and millimeter precision 3D scans. It only works in places with good weather. The rules of the road are hardcoded in millions of lines of fragile code. To top it all off, they are basically driven entirely by remote operators.
FSD works anywhere and doesn't rely on millions of dollars worth of equipment and pre-mapped environments. Tesla's technology is scalable in a way that Waymo never will be. You can buy a Tesla with FSD today for less than 50k. Google will eventually realize that Waymo is burning cash with no path to recovery, and shut it down just like every other Google product.
As much as European cars have a favorable reputation among American consumers: They are awful!
If you do benchmarking, you are going to find European cars have lots of appearance parts that are high quality, but the actual components are poor.
A more realistic comparison would be to compare Tesla vs Ford or GM. These are 3 companies that have similar tech/culture standards. This video is like comparing AMD to Apple.
What do you mean by this?
They are comparing self driving to driver assistance.
How many McLaren, Ferrari, Lamborghini, Porsche, etc equivalents does America produce? Beyond most of those examples there is good reason “German Engineering” is an oft-repeated phrase. See, I can cherry pick too (Europe also produce Fiats, lol, “fix it again tomorrow”).
EDIT: and tbh I don’t even agree with your cherry picked example either, every Tesla I’ve been in has had terrible build quality and been a horrible ride.