Rivian Autonomy & AI Day! Lidar, Hands-Free Driving, Rivian Assistant & more announced

[mkhuffman]

So with everything we know do we think we can expect automatic lane change before point to point? Was kind of expecting auto park and auto lane change to be announced with universal hands free

It seems like it to me:

Sponsored

 

[iansriv]

My R1S is my first EV. I was not fully aware of OTA updates and am thankful for what I have received so far. I'm more used to ICE cars being "static." You get what you bought. So, if I miss out on all the latest and greatest I will either do with what I have or get the latest.

 

[dleepnw]

“Gen 3 Autonomy hardware including ACM3 and LiDAR is currently undergoing validation and we expect it to ship on R2 models starting at the end of 2026.”

so a reason not to buy an R2 early on? That could be really bad for the unveiling where strong sales are needed most

Yeah, lets hope the hardware is actually on the early models, otherwise whats the point of buying those unless you really don't care about ADAS or the future enhancements your vehicle can get? And it definitely will be a sales disaster early on if that's the case.

 

[portdirect]

Your inability to understand the output doesn’t make it slop. It just means you’re struggling with the concepts. Let me help you:
Rivian is adopting the same 'Winning Stack' (Multimodal World Models + Lidar + Thor-class Compute) that every actual leader in autonomous driving uses. That is true of Baidu, Waymo, and XPeng.
In certain Reality Distortion Fields, a company running 29 test vehicles with full-time babysittersin Austin (Tesla) is somehow considered 'ahead.' But in the real world of Cabbages and Kings, the company delivering 450,000 paid autonomous rides per week (Waymo) and the one delivering 250,000 paid rides per week (Baidu) are the actual leaders.
The company delivering zero autonomous rides in zero places and holding an accident rate higher than the autonomous leaders is behind.
Rivian is smart enough to mimic the real-world leaders, not the imaginary one.

I don’t think today’s launch gave us enough data to draw strong technical conclusions: beyond the fact that Rivian now has full end-to-end control over its autonomy stack, which is genuinely significant.

Using the 1,600 TOPS figure as an example, Rivian is quoting sparse INT8 performance. Sparse TOPS is a valid and relevant metric, but it represents an aggregate, peak system value under specific sparsity assumptions. For meaningful hardware comparisons, we would need dense INT8 throughput (and ideally bandwidth and latency), which were not disclosed.

It’s also reasonable to assume we won’t see those numbers unless Rivian intends to position RAPx as a merchant competitor to platforms like NVIDIA Thor; publishing dense figures offers little upside for an internal, vertically integrated system.

edit: I do agree that this brings Rivian inline with the rest of the industry and state of the art globally - however in the usa we are seeing a race to a subscription based model , while china etc are viewing this as becoming a commodity tech. Will be a really interesting next 10 years.

 

[captainjp]

Here is the thing though - it doesn’t matter. Rivian is years behind the curve on the basics. They haven’t even yet reached the limits of the original MobileEye system, let alone their current Gen 2 platform. Sensor stack isn’t remotely a limitation in their foreseeable future.

What?! Lol

 

[UnsungZero_OldTimeAdMan]

Here is the thing though - it doesn’t matter. Rivian is years behind the curve on the basics. They haven’t even yet reached the limits of the original MobileEye system, let alone their current Gen 2 platform. Sensor stack isn’t remotely a limitation in their foreseeable future.

 

[ATLRivvy]

What?! Lol

The current Gen 2 R1S self-driving capability and stability is less than GM / Ford systems running on MobileEye

They aren’t facing hardware limitations in making the system better than it is today

 

[captainjp]

The current Gen 2 R1S self-driving capability and stability is less than GM / Ford systems running on MobileEye

That’s great.
Do I give a shit?
Not at all.

As long as I can take my hands off the wheel long enough to pick my nose, roll up the boog and throw it out the window. I’m happy.

 

[ATLRivvy]

That’s great.
Do I give a shit?
Not at all.

Then I’m not really sure why you are in an autonomy thread but you do you

 

[mkhuffman]

What if Rivian lets people add the Lidar based autonomy package to all L2 models built in 2026, with the expectation that a service center will add it when it becomes available?

They could charge $60k for the base model, and $65k for the Lidar model. If you want to take delivery early, you can but it may be a while before you can schedule a time to add it in.

When I thought of this idea it sounded good, but the more I think about it, the more I think the service centers could be overloaded and that approach could piss a lot of people off.

Just a thought. It would let people who really want a Gen3 get an R2 earlier than the end of 2026, or early 2027.

After reading all the negative posts, I have to say I am not negative about this at all. Autonomy capabilities are required. If Rivian can't do it, they cannot survive. And if Rivian cannot do it cheaply, they cannot compete. This announcement puts them on a path to be able to compete in a market full of self-driving vehicles.

 

[sunydrm]

FSD v10 was a hybrid stack (NN perception + C++ heuristic control) running on HW3. It hit a ceiling because you can't hand-code the planning rules for L4.
Three reasons why your '3 years behind' math is off:
1. The Compute Floor: You can run a 'good enough' L2 demo on 72 TOPS. You cannot run a redundant, safety-critical L4 system on that. Rivian is deploying Thor-class compute (1,600 TOPS) in 2026. Tesla is stuck on HW4 (~300-500 TOPS) until AI5 arrives in 2027. Rivian has the hardware headroom today that Tesla is waiting two years for.
2. Model Architecture: Rivian isn't using 'End-to-End' in the sense of 'Imitation Learning' (Monkey See, Monkey Do). They are using World Models (predicting physics/future states). This is the architecture Waymo uses, and it requires massive compute. It looks like v10 now because it’s a baby, but it scales linearly with compute—unlike heuristics which plateau.
3. Maps aren't Cheating, they are Safety: Calling maps 'cheating' is 2019 thinking. Using priors (maps) + real-time perception is how you get to 99.999% reliability. Trying to derive the speed limit from pixel analysis every single frame is an unnecessary compute tax.
Rivian’s demo wasn’t about 'Look how smooth it is today.' It was about 'Look at the massive silicon and architecture we built to solve the edge cases that stalled Tesla for the last 5 years.

Xpeng ceo just tried fsd v14 and declared it was better than their upcoming release.

https://cnevpost.com/2025/12/11/xpeng-ceo-revisits-tesla-fsd-silicon-valley-heres-his-take/

who is distorting what realit

Tesla has 60ms reaction time. This is insanely fast. You only get there through hyper optimization.


Rivians current demo is worse than fsd v10
Tesla fsd computer has almost double the memory bandwidth of NVIDIA Thor. They have this because they use high power gddr6x memory and not the low power type. Ai5 has over 2000 tops and has 5x the memory and memory bandwidth

by comparison Waymo has 320gb of high bandwidth memory and 12000 TOPS if their usage of h100 is to be believed. Even a Waymo engineer many years ago said Tesla wouldn’t have the level of driving they have now with the cameras and compute they have.

 

[mkhuffman]

Xpeng ceo just tried fsd v14 and declared it was better than their upcoming release.

who is distorting what reality?

Tesla fad has 60ms reaction time. This is insanely fast. You only get there through hyper optimization.

Tesla fsd computer has almost double the memory bandwidth of NVIDIA Thor. They have this because they use high power gddr6x memory and not the low power type. Ai5 has over 2000 tops and has 5x the memory and memory bandwidth

Thanks for providing all your references.

 

[sunydrm]

I do recall at some point this one saying that Rivian vehicles were not capable of autonomous driving. Some crap having to do with the steering wheel motors and not being able to turn 90 degrees.

Streering is fed through Adas commands. You can upgrade the firmware of the steering column to accept higher torque limits. You cannot do this on consumer cars because the car is not approved for this. The device would not be certified for this purpose. That’s why point to point driving is reserved for R2

500 tops found in NVIDIA orin could deliver fsd v14 performance already

 

[Lisa & Bob]

That's true for some people. There will always be some people that need the latest & greatest thing.

The type of person that keeps a car for ~10 years doesn't really care. Neither do the people who don't care to pay some subscription fee for self-driving whatever.

With Tesla's uptake of FSD reportedly in the ~15% range, I'm skeptical this is a market that will have mass consumer adoption without dramatically lower prices. Tesla consumers are among the most tech-interested of any brand.

I’m so old school that adaptive cruise control is overwhelming. Hopefully I’ll live long enough for L4 so I can be carried to the car and it will know where I want to go..

 

[sunydrm]

Thanks for providing all your references.

https://cnevpost.com/2025/12/11/xpeng-ceo-revisits-tesla-fsd-silicon-valley-heres-his-take/



Tesla has ridiculous reaction time through a neural network optimization and having high bandwidth memory.

Sponsored