Tesla’s Full Self-Driving (Supervised) may not actually be autonomous, but it is still one-of-a-kind in the U.S. even years after its initial launch. For now.
While advanced driver-assistance systems for the highway are a dime a dozen in 2025, Tesla’s FSD aims to tackle all the complexities of driving from A to B: traffic lights, roundabouts, highway interchanges, four-way stops and everything in between. Whether it’s because of Tesla CEO Elon Musk’s conviction that personal autonomous cars are almost here, or because of the company’s infamous tolerance for risk (or a little of both), FSD is the only system like that in America.
But the competition is heating up to build the best car you don’t really have to drive. At Rivian’s Autonomy and AI Day earlier this month, company CEO and founder RJ Scaringe said point-to-point automated driving would come to Rivians sometime in 2026. (A different new feature, Universal Hands-Free, launched in December. That brings adaptive cruise control and lane-keeping to 3.5 million miles of roads.)
At the event, I got a roughly 20-minute ride in an R1S that was running a unreleased driver-assistance tech that previews the future point-to-point software. And the experience was uneventful in a good way. The drive was mostly smooth sailing save for some minor incidents I’ll describe.
It’s tough to take away any real insights from this kind of controlled demo—unless things go very poorly or amazingly well. Still, the ride-along and chat with a Rivian autonomy engineer was a fascinating look into the challenges of training a car to drive itself. Check out the video above to watch the full drive, or read on for the highlights.
Overall, things went well. The R1S smoothly came to a stop behind other cars, respected red lights, successfully executed turns and changed lanes much like a human would. At the same time, nothing out of the ordinary happened that the Rivian would’ve needed to react to. I did sense maybe a bit more jerkiness in the steering wheel than I saw in Tesla’s FSD when I used it recently. But again, this was a short test ride so it’s hard to draw conclusions.
The driver kept his hands off of the steering wheel most of the time. The only unplanned interventions were to nudge the accelerator when the system was being too hesitant. First, it happened after a speed bump, when the SUV inexplicably slowed to a crawl—maybe because of pedestrians on the sidewalk nearby. Later, the same thing happened at a green light. And again during a righthand turn with some people milling about on the closest corner.
Nick Carlevaris-Bianco, the Rivian autonomy team’s senior director of perception, sat next to me in the back seat and explained how the system works and how it’s being trained. What I was experiencing was Rivian’s new “Large Driving Model,” which the EV startup says works a lot like the large language models we’ve all come to know through chatbots like ChatGPT.
Rivian feeds the model enormous amounts of real-world driving data, mainly from customers, and tries to get it to replicate only the good driving behavior. While previous generations of self-driving systems relied on many hard-coded rules (if you see a stop sign, stop) Rivian’s Large Driving Model is trained end-to-end. It absorbs driving data on one end and spits out driving outputs on the other (with some structured guardrails too). That’s also the approach Tesla uses for Full Self-Driving.
The idea is to make a system that’s much more generalizable, and one that you don’t have to teach explicitly to handle every situation. For example, Carlevaris-Bianco said speed bumps weren’t something the team told the model about explicitly. The Rivian slowed down for the ones on my drive solely because that’s what happened in the training data.
A Rivian R1S charging at an Ionna Rechargery.
Photo by: Ionna
The same thing goes for traffic lights, interestingly enough. He said there’s no “explicit logic” in the system that says “if the traffic light is there and it’s red, don’t go.” The output is based on what the model decides.
The approach brings challenges too. Because humans aren’t perfect drivers, the raw data is full of bad habits. Carlevaris-Bianco said that early on the team noticed that the model would stomp the accelerator on an open road. Rivian also had to train out the tendency to roll through stop signs and stop partway into crosswalks.
Especially for a feature that’s still months away from public consumption, Rivian’s software performed admirably. Accounts of other drives at the event were largely positive as well, though some people experienced some hard braking and other disengagements.
Over time, Rivian says its self-driving system will get a lot better—especially once lidar-equipped R2 crossovers hit the road in late 2026, bringing both better perception and the ability to gather higher quality data for the LDM.
Rivian has big plans beyond point-to-point too, with ambitions to reach eyes-off driving in some situations followed eventually by full-blown Level 4 autonomy someday. For now, all I can say is it looks like Rivian is off to a decent start.
Contact the author: Tim.Levin@InsideEVs.com
