The robot swerved through the cafeteria of Rivian’s Palo Alto office, shelves adorned with chilled canned coffees — until it didn’t. Five minutes later, a man carefully pushed it out of everyone’s way, the words “I’m stuck” flashing yellow on the poor droid’s screen.

It was an inauspicious start to Rivian’s “Autonomy & AI Day,” a showcase for the company’s plans to make its vehicles capable of driving themselves. Rivian doesn’t make the cafeteria robot and isn’t responsible for its abilities, but there was a familiar message in its foibles: this stuff is hard.

Hours later, as I rode in a 2025 R1S SUV during my 15-minute demo of Rivian’s new self-described “Large Driving Model,” I was reminded of that message.

The EV equipped with the automated-driving software drove myself and two Rivian employees on a switchback route near the company’s campus. As we glided past Tesla’s engineering office, I noticed a Model S in front of us slow to turn into the rival company’s lot. The R1S eventually noticed this, too, braking hard just before the Rivian employee nearly intervened.

During my demo drive, there was one actual disengagement. The employee in the driver’s seat took over as we passed through a one-lane section of road due to some tree-trimming. Minor stuff overall. But it wasn’t exactly rare either; I spotted multiple other demo rides that had disengagements too.

The rest of the drive went well enough for software that is not ready to be shipped, especially when you consider that Rivian threw out its old rules-based driver assistance system and adopted an end-to-end approach — which is how Tesla developed Full Self-Driving (Supervised). It stopped at stoplights, it handled turns, it slowed for speed bumps, all without programmed rules telling it to do these things.

A quiet pivot in 2021

Rivian’s old system “was all very deterministic, and it was all very structured,” CEO RJ Scaringe said in an interview Thursday. “Everything that the vehicle did was the result of a prescribed control strategy written by humans.”

Scaringe said that when Rivian saw transformer-based artificial intelligence taking off in 2021, he quietly “reconstituted the team and started with a clean sheet and said, let’s design our self-driving platform for an AI-centric world.”

After spending “a lot of time in the basement,” Rivian launched the new ground-up driving software in 2024 on its second-generation R1 vehicles, which use Nvidia’s Orin processors.

Scaringe said it was only recently that his company started to see dramatic progress “once the data started really pouring in.”

Rivian is betting it can train its Large Driving Model (LDM) on fleet data so quickly that it will allow the company to roll out what it calls “Universal Hands-Free” driving in early 2026. That means Rivian owners will be able to take their hands off the wheel on 3.5 million miles of roads in the U.S. and Canada (so long as there are visible painted lines). In the back half of 2026 Rivian will allow “point-to-point” driving, or the consumer version of the demo we received Thursday.

The ‘eyes off’ to ‘hands off’ challenge

By the end of 2026, after Rivian has started shipping its smaller, more affordable R2 SUVs, it will ditch the Nvidia chips and outfit those vehicles with a new custom autonomy computer unveiled Thursday. That computer, plus a lidar sensor, will eventually allow drivers to take their hands and eyes off the road. True autonomy — where a driver doesn’t have to worry about re-taking control of the vehicle — lies well beyond that and will largely depend on how fast Rivian can train its LDM.

This rollout introduces a near-term challenge for Rivian. The new autonomy computer and lidar won’t be ready until months after the R2 goes on sale. If customers want a vehicle that can handle eyes-off driving (or more), they’ll have to wait. But the R2 is a crucial product for Rivian, and the company needs it to sell well — especially in the wake of declining sales of its first-generation vehicles.

“When tech is moving as fast as it is, there’s always going to be some level of obsolescence, and so what we want to do here is to be really direct” about what’s coming, Scaringe said. The early R2s will still get Rivian’s promised “point-to-point” driving, which will be based on the new software and will be hands-off but not eyes-off.

“So [if] you’re buying an R2 and you buy it in the first nine months, it’s just going to be more constrained,” he said. “I think what will happen is some customers will say ‘that matters a lot to me, and I’m going to wait.’ And some will say ‘I want the newest, best things now, and I’m going to get the R2 now, and maybe I’ll trade it in a year or two, and I’ll get the next version later. Fortunately, there’s so much demand backlog for R2 that we think, by being upfront with this, customers can make the decision themselves.”

“In a perfect world, everything times at the same time, but the timeline of the vehicle and the timeline of the autonomy platform are just not perfectly aligned,” he said.

When I first interviewed Scaringe in 2018, before Rivian even showed what its vehicles looked like, he shared a goal that still rattles around my head. He wanted to make Rivian’s vehicles so capable of driving themselves that: “if you go for a hike, and you start at one point and you finish at another point, you have the vehicle meet you at the end of the trail.”

It was the kind of pie-in-the-sky promise about self-driving cars that was all the rage seven years ago, but it stuck with me at least because it was something that felt true to Rivian’s whole brand of aspirational adventure.

Scaringe told me Thursday he still thinks it’s possible for Rivian to enable a use case like that in the next few years. It certainly won’t happen until the company tests and builds its more-capable R2 vehicles, which is at least a year away in a best-case scenario.

“We could [do that]. It’s not been a huge focus,” he said. That could change as the company gets closer to level 4 autonomy, though, since by then the company will have its LDM trained on trickier roads without guiding features like lane lines.

“Then, it becomes a bit of a like, what’s the ODD [operational design domain]? Dirt roads, off road? Easy,” he said. Just don’t expect a Rivian driving itself up Hell’s Gate in Moab.

“We’re not putting any resources into rock crawling autonomously,” he said. “But in terms of getting to the trail head? For sure.”