Mercedes MB.DRIVE ASSIST PRO Public Demo (or what Lucid's Dream Drive Pro might look like)

What we are talking about is the total time it takes to process an input and send a command to the car. The less memory bandwidth you have (the more you use for sensors), and running VLA models means that your output latency is very, very high.

High latency means bad reaction time, which prevents you from braking for a pedestrian that pops out from behind a car.

This is simply not how VLA models work. As mentioned before, the VLA model acts as a planner, charting out the next few seconds of driving, with adjustments 10x per second on Nvidia's latest model. Notably, that is about 3x quicker than the avergage person's reaction time. Also, this means there wouldn't be "ping-ponging" in a lane because there is a planned trajectory.

This planner is completely separate from the low latency systems that are coupled with VLA models and adjust as much as 100x per second. This is what is actually executing the VLA model's planned trajectory. For the brief moments where the planned VLA path is no longer feasible, like due to something jumping in front of your car, the separate system computes a new trajectory.

In other words, this concept that systems utilizing VLA models have "very high latency" and "bad reaction times" is not grounded in reality.
 
Sure, I can make up numbers too. Doesn't make it a valid argument.
Current auto industry has 1 employee for 66 cars. That means for 2 million vehicles, you need 30K+ full time employees. Now because you own the entire driving stack and car, you are responsible for all of this

Remote support, service, etc. is about 1.8 billion a year at an average of 60K each. How much does it cost to run service centers in every city in america? Billions of dollars

Tesla currently uses a tech that the car drives from satellite images and plans routes. Assuming this tech is still used at scale, buy satellite images of the entire U.S. at high resolution. It's not cheap. That's easily 2-3 billion and you need to do this regularly

Currently AI is expensive and developing new software is also expensive. That doesn't change

Even just building a cortex 1 level training cluster is basically impossible by a regular automaker. You need 10 billion and it would take at least 4 or 5 years.
 
Last edited:
Notably, that is about 3x quicker than the avergage person's reaction time.
You can compare FSD to every system already in china. The reaction times are nonexistant/slow. It's night/day

Nvidia will not be any different

3x faster than a human is like 300ms where tesla is like 100ms
 
Current auto industry has 1 employee for 66 cars. That means for 2 million vehicles, you need 30K+ full time employees. Now because you own the entire driving stack and car, you are responsible for all of this

Remote support, service, etc. is about 1.8 billion a year at an average of 60K each. How much does it cost to run service centers in every city in america? Billions of dollars

Tesla currently uses a tech that the car drives from satellite images and plans routes. Assuming this tech is still used at scale, buy satellite images of the entire U.S. at high resolution. It's not cheap. That's easily 2-3 billion and you need to do this regularly

Currently AI is expensive and developing new software is also expensive. That doesn't change

Even just building a cortex 1 level training cluster is basically impossible by a regular automaker. You need 10 billion and it would take at least 4 or 5 years.
WTF are you even taking about and what does it have to do with my statement about you making up latency numbers?
 
WTF are you even taking about and what does it have to do with my statement about you making up latency numbers?
Those are the costs of enabling a system to drive at scale. Which is an important part of actually delivering on unsupervised.

The latency numbers are not made up. You can simply test these systems and see how quick they react to objects appearing out of nowhere. There is a night/day difference

We will have to see across the next cycles of NHTSA reports but based on current findings it seems that above 5mph, tesla's accident rate is lower than waymo in texas

This would indicate a latency difference and/or better accident avoidance.

Waymo used to have jerky steering which is an indicator of high latency. At least that seems to be better.
 
You fail to see the point and advantage of complimentary sensors is that each have strengths and weaknesses and having multiple sensor types minimizes the weaknesses.
In security, which I know quite a lot about, and which is a domain in which we care quite a lot about robustness, we call this layering "defense in depth."

but you know, what do I know
 
 
You can compare FSD to every system already in china. The reaction times are nonexistant/slow. It's night/day

Nvidia will not be any different

3x faster than a human is like 300ms where tesla is like 100ms
As said above, the reaction time is not measured by the VLA model. It's measured by the secondary safety layer which can be 100ms or less. In other words, as good if not better than the made up Tesla you're describing.
 
Last edited:
The latency numbers are not made up. You can simply test these systems and see how quick they react to objects appearing out of nowhere. There is a night/day difference

We will have to see across the next cycles of NHTSA reports but based on current findings it seems that above 5mph, tesla's accident rate is lower than waymo in texas

This would indicate a latency difference and/or better accident avoidance.
Your latency numbers were absolutely made up. Please give me evidence suggesting the contrary. Your anecdotal experience of feeling one is slower than the other is not evidence.

Also, accident rates above 5 mph can be driven by a lot of things. You assuming it's latency is unsubstantiated.
 
Back on the MB demo - very promising. Didn't Nvidia also do Mobileye - what the Tesla first gen autopilot was based on?
 
Your latency numbers were absolutely made up. Please give me evidence suggesting the contrary. Your anecdotal experience of feeling one is slower than the other is not evidence.

Also, accident rates above 5 mph can be driven by a lot of things. You assuming it's latency is unsubstantiated.
Asking for evidence? Godspeed. 🫡
 
Your latency numbers were absolutely made up. Please give me evidence suggesting the contrary. Your anecdotal experience of feeling one is slower than the other is not evidence.

Also, accident rates above 5 mph can be driven by a lot of things. You assuming it's latency is unsubstantiated.
There's videos of people testing the car's reaction to people/balls, dummies, etc. popping out from behind a car where you can't see them.

The car reacts in about 1.5 frames at 30fps of video

XPENG VLA 2.0 by comparison is extremely slow

By comparison, if Nvidia claims 3x faster than a human, that would be pretty slow. Human reaction time is just over 1 second if you calculate both reaction and actually moving your foot to press the brake.


 
My concern is that here in Houston, where Lucid/NVIDIA is doing a lot of their robotaxi testing, a large portion of my around town driving is on a mix of surface streets and highways. Having two separate systems, one that is point to point for surface streets and then another that is only basic hands free (without point to point) for highways, that have to be engaged and disengaged separately, will give this system limited usability for much of my and most other peoples driving.

My 4 times a week drive to the gym is typically a 2 mile drive on surface streets, get on highway and drive 3-4 miles, exit and drive another 2 miles on surface streets. In this scenario, the NVIDIA system (as it is now) is basically useless. I could do that drive using just surface streets but that would add 10 minutes to drive.
 
My concern is that here in Houston, where Lucid/NVIDIA is doing a lot of their robotaxi testing, a large portion of my around town driving is on a mix of surface streets and highways. Having two separate systems, one that is point to point for surface streets and then another that is only basic hands free (without point to point) for highways, that have to be engaged and disengaged separately, will give this system limited usability for much of my and most other peoples driving.

My 4 times a week drive to the gym is typically a 2 mile drive on surface streets, get on highway and drive 3-4 miles, exit and drive another 2 miles on surface streets. In this scenario, the NVIDIA system (as it is now) is basically useless. I could do that drive using just surface streets but that would add 10 minutes to drive.
FSD was this way for many years but they were connected systems that behaved as one. The only ever genuine at-fault death of FSD happened when they were using the old system on freeway where presumably it was not designed to stop for pedestrians

I think it's because it's hard to make the car do what you want (mercedes has no speed control with E2E just like tesla) which people don't want on the freeway

One solution is to make the car go well above the speed limit and then put a limiter so that the consumer can limit the car to what ever speed they want. The downside is the neural network may plan a maneuver for a speed it is not able to go.
 
Kim Java's new video on MBZ's NVIDIA self driving. Looks very good. She compares it to Tesla's current FSD on HW3.
 
Back
Top