Why Rivian Threw Out Its First Self-Driving Stack and Built Its Own Brain

Scaringe says Rivian always saw itself as a mobility company, and autonomy was part of the plan from the start. When R1 launched at the very end of 2021, it used what he calls a 1.0 approach: a third-party front-facing camera feeding a rules-based planner that Rivian built around it.

He says the team knew at launch that this was the wrong design, since work on it had begun well before. At the turn of 2021 and 2022 Rivian decided to reset completely. He describes the decision as easy, because the company expected to spend many billions on the technology and driving itself is at the core of transportation.

The clean-sheet platform reached vehicles in mid-2024 with the Gen 2 cars, sharing no code and no perception or compute hardware with the first generation. Rivian then had to grow its fleet enough to build a data flywheel and start training models. Scaringe says the next three to four years of progress will have a completely different slope from 2021 to 2025, because the new architectures are truly AI-based rather than human-codified rules.

Scaringe says the arrival of transformer-based encoding made it clear that the future was neural nets. That was painful for companies that had invested in classic systems, because he expects the vast majority of that earlier work to be thrown away. The change was a complete rethink of the architecture rather than a gradual shift.

Rivian's philosophy is to vertically integrate whatever is important, including electronics, software and the high-voltage systems. For autonomy he lists the core ingredients. You need full control of the perception platform with no intermediary processing signals, so that raw data can feed the system. The vehicle must trigger and store interesting events, upload them when Wi-Fi is available, and a large GPU cluster must train on them.

Independent autonomy vendors typically lack the mileage a carmaker has, and sensor makers lack the vehicle architecture and car park. Scaringe thinks there are more than one but fewer than five companies outside China with the capital, GPUs and fleet to do this. When pressed, he includes Tesla and Waymo alongside Rivian. He says solutions stuck in the 1.0 framework have effectively no chance of catching up to neural-net approaches.

A neural-net approach needs a lot of inference on the vehicle, which can be bought or built. Rivian decided to build an in-house chip. Scaringe says the reason was cost, not capability, since the aim is for every car to be able to run very high levels of autonomy.

He says people fixate on perception because it is visible, but cameras are designed and built in-house, radars are very cheap and lidar is now cheap too. The onboard compute is the expensive piece, far costlier than the perception stack. Bringing the chip in-house removes cost so the system can go on every car.

On the levels of autonomy, Scaringe says the old split has blurred. A few years ago, level two meant camera-heavy and level four meant lidar-heavy and too expensive for consumer cars. Now the difference in perception and compute is fading, and what matters is how well the system handles extreme corner cases. A driver may not notice the gap over the first nine-nines of driving, and the real difference lies in the fifth, sixth and seventh nine.

Scaringe says the safety question is how confident the team is that the system covers obscure, rare events that could otherwise cause terrible outcomes. Neural nets have made the capabilities much stronger and allowed deployment across far more vehicles. Where the state of the art used to be a test fleet of a few hundred cars, every car on the road can now help identify corner cases, and simulation adds more.

By 2030, he predicts, it will be inconceivable to buy a car and not expect it to drive itself, and Rivian hopes to get there sooner. He compares it to airbags and air conditioning, which were once optional. Within a couple of years he thinks it will be hard to accept a car that cannot drop you at the airport or pick up your kids from school.

The interviewer suggests most big carmakers lack the ingredients Scaringe listed. He answers that no one can easily predict market-share shifts, but he says it is inconceivable for a mass-market car company to operate at scale without a software-defined architecture, even before autonomy enters the picture.

Scaringe explains that, apart from Tesla and Rivian, every car uses a domain or function-based architecture. Each function, such as chassis control, doors or climate, has its own small computer, called an ECU. A modern car can have 100 to 150 of them, each running an island of software written by a supplier, or often a supplier to the supplier.

That makes debugging and updating very hard. A feature like walking up to the car and having it unlock, set the climate, adjust the seats, play a sound and flash the lights spans many ECUs and perhaps ten different players, so carmakers rarely change it. With a zonal architecture of one to three computers running one operating system, Scaringe says such a sequence can be changed in minutes or an hour and shipped over the air.

Rivian ships about one update a month, adding a couple of features and refining others based on customer feedback, and customers now anticipate each release. Scaringe traces the legacy design back to early fuel-injection computers, which carmakers outsourced as non-core, and the pattern spread like a field of weeds over 60 to 70 years. Rivian signed a $5.8 billion software licensing deal with Volkswagen Group to use its network architecture and ECU topology across VW's brands.

Scaringe says carmakers have three choices: shrink, build it themselves despite lacking software DNA, or source from third parties. He says the third-party autonomy options are mostly classic 1.0 solutions, which suit selling a sensor and a function but fit poorly with a constantly learning data flywheel. He thinks vertically integrated stacks will gain big advantages.

Asked whether driving models will converge like language models, Scaringe says LLMs converged because they train on nearly the same internet data. Driving has no equivalent, so you need a strong sensor set and a large enough car park to collect data. Tesla has by far the largest fleet.

Rivian's bet is a perception stack heavier than Tesla's but lighter than Waymo's, with better cameras, more megapixels, radar for object detection and, on R2, lidar. He says lidar is a powerful training tool: it provides ground truth for tiny distant objects that cameras struggle to identify. Traditionally that required a dedicated ground-truth fleet, and putting it on every car turns the whole fleet into a data acquisition machine.

Safe driving is a baseline, he says, so differences will show up in how the car feels and in the interface. Rivian already offers mild, medium and spicy driving modes, and the model may learn personal preferences such as lane choice. He expects such behavior to become a key purchase factor, and notes Tesla may handle it differently.

Scaringe says R1 is a flagship with an average selling price around $90,000. He says the R1S is the best-selling premium electric SUV in the US, outselling the Tesla Model X roughly two to one, but price limits volume. R2 is the first truly mass-market product, starting at $45,000, in a range where the average new US car costs about $50,000.

He notes that roughly half of the EV market is the Model 3 or Model Y, leaving a large opportunity to pull buyers out of combustion vehicles with something different. EV adoption in the US is around 8%, while buyers under $70,000 can choose from well over 300 model lines in every form factor. In EVs, he sees only one to three great choices.

Scaringe says many Tesla alternatives look nearly identical to the Model Y when their side profiles are overlaid. He says he has owned a Model Y and thinks it is a great car, but the world needs another choice. Cars are a form of personal expression, he argues, and adoption will broaden as more brands and form factors appear, which is also a motivation behind the Volkswagen partnership.

He says Rivian designs R2 as the best possible vehicle, not as a good EV, so buyers are drawn into electrification by the car itself. Most R1 customers are first-time EV owners, which he says is the real goal. Closing on the future relationship with cars, he says cars enable freedom and self-expression, and Rivian wants them to inspire as well as enable, for example with a flashlight in the door as an invitation to explore.