Rivian will test its Uber robotaxi system in Chicago as well as San Francisco and Miami before year-end, and the hardware characterisation of its first in-house chip should be complete “within the next month and a half,” its autonomy and hardware chiefs told Goldman Sachs’s Communacopia conference on Tuesday.
James Philbin, Senior VP of Autonomy and AI, and Vidya Rajagopalan, Senior VP of Electrical Hardware, supplied the new detail in a session hosted by Goldman Sachs Analyst Mark Delaney that otherwise held to the roadmap Rivian set out in July.
Chicago Added
Philbin said the second milestone in the Uber agreement “involves testing in three urban areas by the end of the year. That’s San Francisco, Miami, and Chicago.”
Chicago was named for the first time.
The March 19 agreement named San Francisco and Miami as the 2028 launch cities, and Rivian’s July 30 earnings call referred to safety-driver testing in those two.
“We already have expert driver teams already driving around, actually collecting data and validating the system that we want to show,” Philbin said. Uber conducted a review “a few weeks ago” and “everyone’s very pleased with the progress.”
He added that Rivian had recently demonstrated its point-to-point system to Uber’s CEO and CFO, “and I think one of those folks posted about it on Twitter.”
Passing the milestone unlocks $250 million, the second of five tranches in Uber’s $1.25 billion commitment, Chip Newcom, VP of Investor Relations, said in the same session.
The first $300 million, released on the first milestone in April, closed on May 4.
Three further milestones, tied to an L4-capable R2 and expansion to up to 25 markets including at least one in Europe, release the final $700 million.
RAP1 Chip
Rajagopalan gave the most specific status yet on RAP1, the 5-nanometre processor Rivian unveiled at its AI Day on December 11.
Rivian has had silicon in-house “for more than a year and a half,” she said.
Automotive-grade characterisation testing “is underway and pretty much wrapping up.”
“Within the next month and a half, we should be pretty much done in sort of the hardware characterization testing piece of it,” she said, a statement about that test phase rather than full production release.
Vehicles carrying the chip are on the road. The Gen 3 compute module built around two RAP1s, demonstrated at AI Day, has been in-house for over a year.
The software has now been proven on it. Rivian’s released features, Universal Hands-Free, Lane Change on Command and Highway Assist, “have been exercised and are working as expected” on the new silicon.
Performance meets the target Rivian set, “at least 4x as performant” than Gen 2, “as yet without fully optimized software.” The compiler and tooling are “all on track.”
Gen 3 remains on schedule for late 2026, the date given at AI Day, when Rajagopalan said the combination of 11 cameras, five radars, one LiDAR and 1,600 TOPS of compute would be “the most powerful combination of sensors and inference compute in consumer vehicles in North America.”
CEO RJ Scaringe said in July that production silicon had entered final testing, EV reported.
Gen 3 Will Come in Tiers
Two statements on Gen 3 were new.
Asked how its cost compares with Gen 2, Rajagopalan said Gen 3 “will enable a few different performance points. The lower performance point will be cheaper than the Gen 2 hardware, and the higher performance point will be on par or slightly higher.”
On a dollars-per-TOPS basis it is cheaper across the board, which was one of the three reasons Rivian built its own chip, with performance and engineering velocity.
She described merchant AI silicon as carrying “a lot of margin stacking” and a premium in the current market, and said an in-house chip lets Rivian “offer more at the same price point.”
Asked when Gen 3 reaches the R1, she said Rivian “haven’t actually announced that yet.”
She also corrected a common reading of the roadmap. Point-to-point driving, due by year-end, “will actually be available on Gen 2 and Gen 3. So what the RAP1 chip and LiDAR enable is really functionality at a later date. The eyes-off functionality we’ve talked of for later. There should be no difference at launch.”
A Chip for Robots, Offered Openly
Rajagopalan said RAP1’s “first application was autonomy, but it is a very capable chip for physical AI. If anybody out there is interested in using our chip for robots to reach out.”
Rivian spun out its robotics work as Mind Robotics, which raised $500 million in March. Rajagopalan said Rivian “cannot speak for Mind, they are another company.”
Philbin was more cautious on whether driving models transfer to robots.
“The jury’s still out a little bit,” he said. Robots have “a lot more degrees of freedom” and it is “not clear how to get the human-derived ground truth,” and they face “more generalization problems” than driving, “which is really the one task.”
Point-to-Point: What Is Left
Philbin put the point-to-point release “towards the end of the year,” the target Scaringe reaffirmed in July.
The team has been “scaling up the model, scaling up the data, and the validation of that model, and then actually putting the vehicles on the road in a test framework.”
The remaining work is “dialing in a good humanistic behavior in all cases,” making sure the model “is not too conservative, that it’s sort of tracking speed as you would expect.” He said it already handles “relatively complex scenarios like construction or narrow negotiation” well. Validation “needs to happen so that we can make sure it’s safe before we ship it.”
Eyes-off in 2027 “is essentially the same system,” he said, gaining confidence through more miles and long-tail validation until “the system architecture” allows the driver to look away.
Two details on training were new. Rivian uses “maybe 10 to one customer data to expert driving data,” with the expert data correcting customer habits such as “not stopping at stop signs fully” or “speeding slightly.”
And training compute comes from Amazon Web Services, “a great supply of GPUs that at least for the next 6 to 9 months is going to support us.” Past that, “of course, we’ll need to see.”
Autonomy+ take rates are “tracking actually better than we expected,” Philbin said, describing the features as “very sticky” and predicting they become “table stakes for any new vehicle.”
Rivian began charging $49.99 a month or $2,500 for the package in April.
Robotaxi as a Step, Not the End
The L4 robotaxi will run “essentially the same” software stack as the consumer car, Philbin said, with “a robust safety architecture” that “comes with the RAP1 silicon” and “some additional sensor changes for the final L4 product that we’re working through the details of now.”
Rajagopalan explained why: “Robotaxi has to start from a stationary spot with nobody in the vehicle, which is not typically a problem you have to solve for a consumer vehicle.” The robotaxi gets more sensing and more compute headroom for that.
Both executives said the robotaxi is not the goal. “We really think personal level 4 is the product that we want to get to,” Philbin said. “Robotaxi is a step along the way.”
Bringing the stack to Europe would mean deploying expert drivers there first, with “relatively modest coverage of new road conditions and new road rules,” Philbin said.
Rivian has set no timeframe for autonomy on its commercial vans.
