Earlier quoted context omitted.
In that video they mention doing localization based upon a prebuilt map of the route by matching images to the model 10 times per second. That is by definition, not FSD. That is like the system announced today, a limited route autonomy. For comparison, FSD v3 (they are shipping v4 in every vehicle now) performs localization 2,000 times per second based upon a hybrid model of every road in open street maps and a gener…
They aren’t shipping v4 in the model 3, so no on the “every vehicle”
Mercedes beats Tesla to autonomous driving in California
71–80 of 778 posts
Re: Mercedes beats Tesla to autonomous driving in California
#72Friendly reminder that this system is HEAVILY limited, with the following restrictions: - Must be under 40 mph - Only during the daylight and only on certain highways - CANNOT be operated on city/country streets - CANNOT be operated in construction zones - CANNOT be during heavy rain or fog or flood roads And for comparison: https://www.youtube.com/watch?v=DjZSZTKYEU4 Tesla FSD navigating the complex city streets of…
Would you bet your life on a Tesla Twitter promise? They are not even betting one cent on their car safety.
Re: Mercedes beats Tesla to autonomous driving in California
#73Earlier quoted context omitted.
All of that is technically correct but isn’t the point the OP was making. These are products with wildly different aspirations and at different stages of achieving those aspirations.
Yes. One product works and other doesn’t. And one that doesn’t work is better and for sure will improve and work in the future? And one that does work won’t improve?
It’s bizarre and absurd to claim a system that’s driven billions of miles “doesn’t work”.
I really like how Simondotau put it. These are wildly different systems. Benchmarks can always be rigged and I see this move by Mercedes no differently.
Having failed to compete with Tesla on actual self driving technology, they concocted a minimally useful system that could check a box in their marketing material. Mercedes, here, is an ostrich.
My primary benchmark for how useful the system is would be how many passenger-miles its driven. We’ll be waiting a long time for Mercedes to be bragging about that metric.
Re: Mercedes beats Tesla to autonomous driving in California
#74Earlier quoted context omitted.
> but the system as a whole is far less capable than Tesla FSD, the renowned “exaggerator” told me so We actually have no idea what Benz et al.’s systems are capable of if they were to deploy them with the same recklessness as Tesla.
This statement makes no logical sense. Their systems are “capable of” what they can be used to achieve in actual operation. It’s not like it’s this massively sophisticated system for FSD with most of the code commented out. These are purpose-built and heavily tested to a specification - that’s how engineering works. The Mercedes system specification is extremely limited in practice, but gives them the marketing win o…
Systems that are required to have ~100% performance within an operational design domain generally have more than 0% performance beyond that domain.
For example, Tesla FSD seems like a pretty good highway cruise control, but it can be dangerously deployed in city driving too. Of course this is a little difficult to argue about since AFAIK Tesla has never defined its ODD and their lead autopilot software guy said in his deposition that he didn’t even know what an ODD was?
I suppose one man’s “marketing win” to “claim” L3 is another man’s “actually deploying L3.”
Re: Mercedes beats Tesla to autonomous driving in California
#75Friendly reminder that this system is HEAVILY limited, with the following restrictions: - Must be under 40 mph - Only during the daylight and only on certain highways - CANNOT be operated on city/country streets - CANNOT be operated in construction zones - CANNOT be during heavy rain or fog or flood roads And for comparison: https://www.youtube.com/watch?v=DjZSZTKYEU4 Tesla FSD navigating the complex city streets of…
Took me less than a few minutes, to find in the same channel, FSD failing and the human having to intervene. Here it is at the proper time and for a video uploaded 10 hours ago... - https://youtu.be/KG4yABOlNjk?t=995
And even worst failures...looking at all these I simply do not believe the videos posted are not edited or a selection of a success out of many failed ones.
https://youtu.be/KG4yABOlNjk?t=1252
https://youtu.be/WowhH_Xry9s?t=77
https://youtu.be/WowhH_Xry9s?t=196
Since the video was live on YouTube no editing was done...
These videos are only posting the successful events, we need to see them all. That is why probably Tesla is not putting their money where their mouth is.
Re: Mercedes beats Tesla to autonomous driving in California
#76Earlier quoted context omitted.
Detecting lane is mostly solved problem. Detecting AND predicting what other things will do on the road (not only cars, but people, animals, trash, etc) is not.
> predicting what other things will do on the road… [like] trash At least we do have a model for this one (physics). I wonder what additional signals the car would need to accurately simulate trash movement? Wind?
I don’t think you could ever accurately simulate trash movement.
Trash is irregularly shaped, even with the highest resolution cameras, it’d be impossible to know what the far side of the trash is shaped like. Then on top of that there’s no way to predict highly localized gusts of wind.
The best way to deal with trash is to just have an ML model deal with lots of it so it can make predictions about what it is likely to do (which is basically what Tesla is already doing).
Re: Mercedes beats Tesla to autonomous driving in California
#77Friendly reminder that this system is HEAVILY limited, with the following restrictions: - Must be under 40 mph - Only during the daylight and only on certain highways - CANNOT be operated on city/country streets - CANNOT be operated in construction zones - CANNOT be during heavy rain or fog or flood roads And for comparison: https://www.youtube.com/watch?v=DjZSZTKYEU4 Tesla FSD navigating the complex city streets of…
In the end, Mercedes assumes liability for its system, where Tesla doesn’t. Tells you all you need to know.
How exactly is Mercedes accepting liability? I mean to say, how does this work in practice? Who absorbs the demerit points if your car is accused of going 40 in a 30 zone?
Re: Mercedes beats Tesla to autonomous driving in California
#78Friendly reminder that this system is HEAVILY limited, with the following restrictions: - Must be under 40 mph - Only during the daylight and only on certain highways - CANNOT be operated on city/country streets - CANNOT be operated in construction zones - CANNOT be during heavy rain or fog or flood roads And for comparison: https://www.youtube.com/watch?v=DjZSZTKYEU4 Tesla FSD navigating the complex city streets of…
Very serious limitation of Tesla:
The human has to keep the hands on the wheel at all times and eyes on the road.
Edit: I looked at these FSD videos before and am very sure it will not work that well in an average European city. Unfortunately, because I really would love to buy a real self driving car that legally allows me to watch a movie while driving (highways is enough).
Re: Mercedes beats Tesla to autonomous driving in California
#79It doesn’t take a lot to retro-fit roads with autonomous guardrail markers when we repave. It’s going to take a federal-level effort to accomplish this (which should have been apart of pandemic infrastructure stimulus). I strongly believe this would improve safety for autonomous vehicles by 1000x.
I used to think that V2X road infrastructure was a major necessary component for self driving. I no longer hold that opinion. The “self-driving” infrastructure would not be magically any better than the “human-driving” infrastructure, and how to handle conflicting data from the infra in terms of sensor fusion is not at all clear. In short, it creates as many problems as it solves, and doesn’t really solve the problem…
I definitely agree state of the art is good enough for 97% of driving scenarios but they’re also not fail-safe and robust. In order for autonomous to be truly acceptable, they need to be 100x more safe. No, nearly flawless. In order to accomplish that you need physical and hardcoded methods as well. In addition, and equally as important, is that road infrastructure gives you improved ability to coordinate cars on the road without relying on individual and disparate compute units.
Re: Mercedes beats Tesla to autonomous driving in California
#80Earlier quoted context omitted.
Yes. One product works and other doesn’t. And one that doesn’t work is better and for sure will improve and work in the future? And one that does work won’t improve?
It’s bizarre and absurd to claim a system that’s driven billions of miles “doesn’t work”. I really like how Simondotau put it. These are wildly different systems. Benchmarks can always be rigged and I see this move by Mercedes no differently. Having failed to compete with Tesla on actual self driving technology, they concocted a minimally useful system that could check a box in their marketing material. Mercedes, her…
Because it’s very very simple. One system is self driving and other isn’t, no matter how you try to spin it.