Rivian May Be Much Closer to FSD Than You Think

[azbill]

I keep seeing this Bolt locally, this was last Thursday.

That looks just like the Cruise Bolts they test in Phoenix and SF. Interestingly, Tesla just got approved by ADOT in Arizona to use their Robotaxis here.

Sponsored

 

[SJC]

Before my R1S gen2 I drove a 2018 Model X with FSD. It worked fairly well point to point except you had to have a pretty good grip on the steering wheel or you got punished and locked out of FSD for a week. Any kind of inclement weather would also cause an alert to manually take over. I believe that Rivian's use of better cameras and radar will do a better job in that regard. Tesla has been receiving billions of bits of data for years and compared to Rivian I would liken it to a kindergartener vs. a high school graduate.

 

[CrazyOne]

Before my R1S gen2 I drove a 2018 Model X with FSD. It worked fairly well point to point except you had to have a pretty good grip on the steering wheel or you got punished and locked out of FSD for a week. Any kind of inclement weather would also cause an alert to manually take over. I believe that Rivian's use of better cameras and radar will do a better job in that regard. Tesla has been receiving billions of bits of data for years and compared to Rivian I would liken it to a kindergartener vs. a high school graduate.

The highschool graduate, while more capable, can get you in a lot worse trouble. Like, life and death trouble, given how people use it. There is plenty of evidence and if that isn't enough, nothing will change some opinions.

 

[captainjp]

Edge cases don't negate the net positives to me.

I’m sure the family of the deceased would concur and find solace in knowing their loved ones sacrificed their lives so that FSD is where it is today. You know, for the greater good.

I don’t have absolute faith in autonomy + however, I’ve yet to hear of a fatality involving it. I know its limitations and I’m aware 100% of the time it has control. I work and drive in all five boroughs of NYC every day. I’ve done so for 30-plus years. I wouldn’t have confidence in ANY self driving tech around here. WAY too many variables and unpredictable scenarios.

 

[nsuglerio]

I’m sure the family of the deceased would concur and find solace in knowing their loved ones sacrificed their lives so that FSD is where it is today. You know, for the greater good.

I don’t have absolute faith in autonomy + however, I’ve yet to hear of a fatality involving it. I know its limitations and I’m aware 100% of the time it has control. I work and drive in all five boroughs of NYC every day. I’ve done so for 30-plus years. I wouldn’t have confidence in ANY self driving tech around here. WAY too many variables and unpredictable scenarios.

Let's be honest with ourselves and look at statistics: " Globally, about 1.35 million people die each year in car accidents, according to the World Health Organization. The leading causes of these fatalities include speeding, distracted driving, impaired driving, and failure to use seatbelts "

We don't ban driving because of that and unfortunately not embracing ADAS systems isn't going to make things better.

The examples you provided namely the driving under the semi was using mobileye and that driver was absolutely not paying attention ( I believe he was watching harry potter ).

While anyone being taken down railroad tracks is no different to me than someone putting on cruise control and heading towards a cliff and just doing nothing. You tap the brake and the entire system turns off and you are still in control of the vehicle. You don't turn on FSD or Driver+ and lose control of the vehicle.

In my case my R1T veered into an exit lane at like 50+ mph and I had to regain control as these systems are not there yet for complete unsupervised use. I don't go around telling people my rivian almost killed me, but I also tell people to use any of these ADAS like a more advanced cruise control.

I think you and I might differ as I work in an area here Waymo has grown and now we will have teslas taxis doing the same things. Riding in a waymo is a surreal experience and I would bet all my money in 5-10 years will only get even better. After my 25 years of driving I will say my faith in human drivers dwindles by the day. Drivers seem more distracted than they ever were when I started driving.

 

[TexasBob]

I am working on an SAE committee that is writing the standard for certification of AI in aviation. One of the key issues is that AI always has uncertainty, as you have pointed out.

One of the ways to get around uncertainty is to use dissimilar sensors and dissimilar AI algorithms. Sensor fusion is one of the newer terms for doing this. If you are using two sensors, such as radar and cameras, if they disagree you can hand control back to the driver or pilot, that provides a level of safety but not continued functionality. If you have three dissimilar sensors; lidar, radar and cameras, if one disagrees you can ignore it and continue. The case of three in disagreement at the same point in time becomes much lower in probability (but it does not become zero).

Minority Report

 

[hudesigns]

Based on where Rivian is at with their highway assist there is no way Rivian is even close to anything full FSD. Their highway assist is very poor and glitchy with lots of nuisances.
A few months back on vacation I rented a Hyundai Elantra and drove from the the tip of the Florida keys all the way up to Orlando. It had 2 independent buttons, 1 for adaptive cruise control and 1 for lane keep assist. Activating both of those together had me cruising for hours without ever touch the wheel and peddles. The performance on this crappy $25k car was way, way better than what I've ever achieved in my R1S.

Totally agreed about Hyundai/Kia ADAS implementation being much better and practical for real world driving than Rivian. It does not simply quit whether mapped or not and that’s huge in relieving long drive tedium. Rivian needs to understand not just to ‘demonstrate’ but to solve real issues for real people.

But with Hyundai/Kia ADAS you can not “not touching steering wheel” for more than few seconds, or it’ll beep you with a big orange circle on driver display and if you disregard it’ll disable ADAS. I’ve owned Ionic 5 for two and half years. I just touch lightly with one hand when using it, that’s it. Mine is a MY 2023 and it does not have monitor on steering wheel. I heard it’s extremely annoying for new MY 2025! Beware.

 

[TexasBob]

Tesla has been receiving billions of bits of data for years and compared to Rivian I would liken it to a kindergartener vs. a high school graduate.

Rivian has access to the NVIDIA Training system for vehicles. Tesla is behind not ahead as the multi-billion-dollar failed DOJO experiment (now shutdown) made obvious. There were a series of critical missteps that left them behind.

Lots of interesting stuff in the link below:
https://developer.nvidia.com/drive

 

[captainjp]

Let's be honest with ourselves and look at statistics: " Globally, about 1.35 million people die each year in car accidents, according to the World Health Organization. The leading causes of these fatalities include speeding, distracted driving, impaired driving, and failure to use seatbelts "

We don't ban driving because of that and unfortunately not embracing ADAS systems isn't going to make things better.

The examples you provided namely the driving under the semi was using mobileye and that driver was absolutely not paying attention ( I believe he was watching harry potter ).

While anyone being taken down railroad tracks is no different to me than someone putting on cruise control and heading towards a cliff and just doing nothing. You tap the brake and the entire system turns off and you are still in control of the vehicle. You don't turn on FSD or Driver+ and lose control of the vehicle.

In my case my R1T veered into an exit lane at like 50+ mph and I had to regain control as these systems are not there yet for complete unsupervised use. I don't go around telling people my rivian almost killed me, but I also tell people to use any of these ADAS like a more advanced cruise control.

I think you and I might differ as I work in an area here Waymo has grown and now we will have teslas taxis doing the same things. Riding in a waymo is a surreal experience and I would bet all my money in 5-10 years will only get even better. After my 25 years of driving I will say my faith in human drivers dwindles by the day. Drivers seem more distracted than they ever were when I started driving.

Whatever it takes for you to justify using it, that’s fine. You do you. I’m not convinced.

 

[nsuglerio]

Rivian has access to the NVIDIA Training system for vehicles. Tesla is behind not ahead as the multi-billion-dollar failed DOJO experiment (now shutdown) made obvious. There were a series of critical missteps that left them behind.

Lots of interesting stuff in the link below:
https://developer.nvidia.com/drive

No disrespect but I believe this is a gross oversimplification of things. It's like how people think strapping LIDAR on a car suddenly means you have solved level 4 autonomy. The real secret sauce is the actual software that knows what to do with that sensor data and the odd edge cases that are impossible to predict until they happen. Nvidia's system is not some panacea to self driving.

I would argue Tesla's biggest advantage is the insane amount of real world training data they have and the edge cases that only real life can put forth. Even after Elon basically firebombing the brand with his antics, the model Y still sells a shit ton of vehicles even in the most recent quarter and that's just more drivers providing data.

I am rooting for Rivian but when it comes down to it a consumer can buy a tesla today that has quite an impressive ADAS suite that only is improving. Excited for Dec 11th and hoping its not just announcing we have to pay monthly without any tangible improvements.

 

[nsuglerio]

I agree with the current state of highway assist. The release of the Rivian system was worse than the Mobileye system that was there before it, but Rivian has since surpassed it by a significant degree. I drove for a few hours using it in heavy traffic and was impressed by how well it navigated the traffic and the decisions it made while switching lanes, adapting its behavior to the situation. Heavy freeway traffic jams are among the most challenging traffic situations to navigate, and Rivian performs remarkably well.

I agree - the first update that switched the gen2 over to Rivians in house model was night and day for me. My daily commute was always the systems shutting off when there was a minor sun glare and since then it just stays locked in and hasn't had any sudden disengagements. Even the biggest scare with my R1T was when the mobileye system suddenly veered off an exit while I was going highway speeds. Lastly the previous system handled lane cut ins terribly. It would keep speeding up until a vehicle was 100% in the lane in front of me where the new system is predicting cut ins better.

I think its a huge improvement.

 

[TexasBob]

No disrespect but I believe this is a gross oversimplification of things. It's like how people think strapping LIDAR on a car suddenly means you have solved level 4 autonomy. The real secret sauce is the actual software that knows what to do with that sensor data and the odd edge cases that are impossible to predict until they happen. Nvidia's system is not some panacea to self driving.

I would argue Tesla's biggest advantage is the insane amount of real world training data they have and the edge cases that only real life can put forth. Even after Elon basically firebombing the brand with his antics, the model Y still sells a shit ton of vehicles even in the most recent quarter and that's just more drivers providing data.

I am rooting for Rivian but when it comes down to it a consumer can buy a tesla today that has quite an impressive ADAS suite that only is improving. Excited for Dec 11th and hoping its not just announcing we have to pay monthly without any tangible improvements.

I think you did not look at the link. The NVIDIA system is not about LiDAR or even TOPS. The thing that is interesting is the end-to-end development system OEMs are buying into including the much more advanced simulation learning. NVIDIA’s Omniverse and Drive Sim platforms are its key advantage. They are selling OEMs access to a massive virtual fleet. And here is the thing - it is probably better than anything Tesla has from the real world and here is why:

If it is ultimately proven that robust self-driving requires additional sensors like high-definition radar or LiDAR for safety and reliability, then Tesla's current data set is vastly less useful because you can't train a LiDAR-based perception module (or mm radar) with camera-only data. If Tesla is wrong about the required sensors (it is) then its training advantage becomes a disadvantage. Tesla has less radar and lidar data than even a bit player and (this is the important part) no way to integrate the visual and other sensors data into a multi-sensor training set for a worldview.

That is why Tesla dumped so much money into DOJO and it is why it is such a big deal that it got shut down a few months ago.

Worth 4 minutes on this:

 

[sunydrm]

I think you did not look at the link. The NVIDIA system is not about LiDAR or even TOPS. The thing that is interesting is the end-to-end development system OEMs are buying into including the much more advanced simulation learning. NVIDIA’s Omniverse and Drive Sim platforms are its key advantage. They are selling OEMs access to a massive virtual fleet. And here is the thing - it is probably better than anything Tesla has from the real world and here is why:

I think you're simplifying a lot of things in the wrong way.

TOPS refers to integer operations, which tesla runs 8-bit neural networks.

Tesla might be the only self driving car using 8-bit.



The tesla FSD AI4 computer has 50% more memory bandwidth than Nvidia's Thor computer. High bandwidth memory is very expensive and power hungry. New AI5 computer will have 5x the memory bandwidth so clearly it is a limitation.



So purely looking at TOPS doesn't show the whole picture. Tesla FSD is very optimized for what it has, and AI4 and the new AI5 computer are designed exactly for how FSD works and they know what the compute limitations are which is why tesla can get good equivalent performance for low power draw.

The new AI5 computer has 5x memory bandwidth, like 8x memory, etc. this will blow past the performance of Nvidia Thor.

If it is ultimately proven that robust self-driving requires additional sensors like high-definition radar or LiDAR for safety and reliability, then Tesla's current data set is vastly less useful because you can't train a LiDAR-based perception module (or mm radar) with camera-only data. If Tesla is wrong about the required sensors (it is) then its training advantage becomes a disadvantage. Tesla has less radar and lidar data than even a bit player and (this is the important part) no way to integrate the visual and other sensors data into a multi-sensor training set for a worldview.

There is this misconception that the reason tesla does not have self driving is because they use camera only. Tesla is pursuing end-to-end self driving without HD maps, and these self driving solutions use vision as a backbone sensor. Waymo EMMA (end-to-end research model) is vision only. Wayve AI who is arguably #2 in this behind tesla and is backed by Nvidia is also vision-only. Many cars with lidar in their ADAS systems, the lidar is likely not even used. Lucid is not using their lidar and many Chinese and western systems it is turned off.



No one has ever demonstrated end-to-end, self driving without HD maps and Tesla is easily #1 in this. A lot of the other companies that you see thrown around like Mobileye, nuro, motional, etc. are very far from even doing robotaxi with a safety monitor.



You have companies doing traditional self driving (ala waymo) like zoox who have driverless with a small ODD and geofence, but they look to be very far from even a 100 fleet driverless solution without a super restricted ODD. They are not exactly #2



More sensors, higher resolution cameras, lidar, etc. all increase memory bandwidth requirements which affect latency. It is not a guarantee that more sensors makes driving easier, because it increases compute requirements which makes driving HARDER if you are already compute limited.

Rivian is likely going to sell the idea that rivian R2 will have the best sensors and therefore they will be able to achieve autonomy because of the sensors. Most likely rivian will not use most of their sensors and cameras in the long run and they are very compute limited already. Tesla has had 10 years of experience in optimizing their AI models and being extremely efficient with memory and compute.

Current FSD v14.2 is probably 30-50K miles per safety critical intervention. In order for rivian to show a notable improvement they would have to achieve 100K+ miles per safety critical intervention. Not only is that not happening, but by the time rivian has any semblance of even what FSD v12 looked like, it will be 2+ years, which tesla will be on FSD 16 or 17 and have the AI5 computer.

There are going to be limitations for using cameras only, but these will be ODD limitations and limitations in how the cars will drive. It is not known at scale whether camera only will not achieve the safety levels for unsupervised. If tesla removes the safety monitor like they plan to in December, then I believe the answer will have revealed itself. XPeng in China is also pursuing vision only self-driving.

There was a DCar test many months back in china and tesla destroyed the competition in a safety test, many of the other cars containing lidar and HD radar.


Now if you want to say that sensors are needed deliver autonomous driving, Rivian does not have equivalent sensors to ANY Robotaxi vehicle on the market. The cameras in a rivian and all of the sensors do not appear to have any sort of cleaning. The radars around the car are cheap, low resolution cruise control radar. By comparison, every waymo camera has cleaning, every lidar has cleaning, and all of the radars around the car are imaging radars. Even the lidar that rivian will use is solid state, low resolution, forward facing lidar with around 120 FOV. This is nothing like the high resolution 360 degree (5 lidars) that Waymo would use.

That is why Tesla dumped so much money into DOJO and it is why it is such a big deal that it got shut down a few months ago.

Tesla spent about 1 billion on dojo. Most of that money went towards actual training compute. It wasn't a failure. Nvidia makes better gpus and Tesla purchase a lot of them. Training compute is important. Rivian lacks training compute. Storage is also important. Camera framerates are important. Rivian appears to be running their cameras at low framerate because of memory bandwidth limitations. They also appear to be collecting and training on low framerate video because they can't afford the proper amount of storage.

Rivian also advertises the megapixels of all their cameras but it makes no sense to have high resolution cameras all around. Some cameras do not need the same resolution, and resolution affects photosite size and low light performance.

My guess, is honestly with maybe 5% of what tesla is spending on FSD, it will go about like you would expect. I would expect in 2-3 years, Rivian has something approaching the level of FSD 12 performance. Can they get to the underlying safety of FSD 14 any time soon? I doubt it.





But the most telling issue of all is that even if Rivian does have an impending solution for self driving that works. They simply do not have the money to deliver such a system. The actual process for tesla to scale their robotaxi solution is very expensive. Billions of dollars expensive. To build the infrastructure needed for consumers to use their cars in unsupervised fashion (level 3 or 4) is also billions of dollars a year.

Self driving is a dangerous game and you can spend billions of dollars only to never have self driving. You can also not have the scale of vehicles to afford to deliver self driving to customers. There are just so many problems.

 

[ATLRivvy]

I’m on a road trip now and the handsfree highway assist has actively tried to murder us twice in the first 150 miles so I’m gonna say no they aren’t close to autonomy.

First time it slammed on the brakes in the middle of 75 mph traffic flow. I assume because there was an 18-wheeler with its caution signs out 3 lanes of traffic away from us.

The second time it tried to take a hard left from the far left lane of the highway and nearly slammed us into the wall. I assume this is because it got confused by a stub lane nearby to connect to future construction. Don’t know why the cars decision was to hard left though.

Despite all the hype from RJ - the Rivian system is still significantly behind the Bluecruise in my Mach-E despite it having far fewer sensors / cameras and barely any updates for over a year now. Most importantly because in situations of (increasingly rare) confusion the Mach-E aggressively alerts the driver to take control far in advance of any decision it makes. The Rivian has very poor decisioning logic right now

 

[Viet658]

China cars are making rapid progress so hopefully Rivian will be as fast as them at least.

Sponsored