In 2024, Hyundai Motor Group talked about a car that never stops learning and improving. Two years and eight months later, the same principle resurfaced under a new name: the Data Flywheel. Right now, roughly 40 dedicated vehicles are collecting data. Nvidia-based vehicles are targeted for 2028; Hyundai's own Atria AI, for 2029. What this media day showed wasn't a production-ready autonomous driving system — it was a fresh starting point, one that reorganized the group's past development approach and reconnected its data, its organization, and its production schedule.

By Sang Min Han _ han@autoelectronics.co.kr
한글로보기




Las Vegas, January 8, 2024. Song Chang-hyun, then CEO of 42dot, took the CES stage and described the car as "an AI machine that never stops learning and improving." Framing the shift as "software-defined everything," he said Hyundai was integrating a continuous machine-learning infrastructure called the Big Data Loop into its software-defined vehicles and vehicle data platform.
It was a cycle: the car gathers data on the road, the AI learns from it, the improved software is redeployed to the car, and the car generates new data in turn. His point was simple: in autonomous driving, data volume and learning speed determine how quickly a product evolves.
Two years and eight months later, on September 11, 2026, the venue changed to 42dot's headquarters in Pangyo. There, Hyundai Motor Group held its first autonomous driving media day and announced that the Data Flywheel was now fully operational. Minwoo Park, President and Head of Hyundai Motor and Kia's Advanced Vehicle Platform (AVP) Division and CEO of 42dot, redefined the terms of competition from the outset.
"Autonomous driving is no longer a contest over who has the best individual feature. What determines a company's competitiveness is how much data it can secure, how fast it can learn from it, and how quickly it can put the results into actual products and services."
The name and scope may differ, but the Big Data Loop and the Data Flywheel share the same basic principle: keeping collection, training, validation, deployment, and re-collection connected without interruption. So what actually changed between the loop of 2024 and the flywheel of 2026?
What Hyundai unveiled this time wasn't a concept — it was the current method and the production timeline. At the same time, the announcement showed that the vision it laid out in 2024 has finally started moving, in earnest, through actual vehicles, defined development processes, and a working organizational structure.



What Hyundai Had to Reconnect First

Hyundai currently runs about 40 dedicated data-collection vehicles around the clock. Each car gathers roughly 80 hours of data a week, which — added up simply — comes to about 3,200 vehicle-hours of driving data across the fleet every week. The system automatically flags situations the AI finds hard to judge — road construction, severe weather, sudden lane changes, cars parked on narrow streets — and prioritizes them for training (hard example mining), then feeds vulnerabilities confirmed through on-road validation back into repeated training (continuous training). It also reconstructs real driving data into 3D environments to test dangerous or hard-to-repeat scenarios virtually, and runs a Follow-the-Sun setup, where teams in Korea and the U.S. hand development off to each other across time zones.
But it's too early to read this announcement as a progress report on a finished data cycle. By Hyundai's own account, the core technologies only began feeding into the flywheel earlier this year. The Special Event Recorder (SER), which logs data whenever autonomous driving disengages or a critical situation occurs, is still being integrated in stages. And the Data Union — a system for pooling data from multiple vehicles and organizations under a common standard — is still being established.
The fact that Hyundai and Kia sell more than 7 million vehicles a year across some 190 countries and regions doesn't yet translate into data scale, either. Hyundai itself said as much: scale is only potential, and without the other pieces in place, that massive sales fleet becomes a cost burden rather than a data asset. In fact, more than half of the data collected so far is simply repetitive footage of cars cruising straight down a lane. More units sold doesn't automatically mean more learning value. Turning annual sales volume into real training assets requires closing the loop — production vehicles whose owners have consented to data collection, standardized sensors and data formats, automatic screening, training and validation, and redeployment back to the vehicles.
The real change lies not in any single technology, but in how Hyundai has connected the pieces. Atria AI didn't suddenly appear with Minwoo Park's arrival. Hyundai had already named Atria AI a core SDV technology in its executive reshuffle at the end of 2025. Seonggyun Jeong, Group Lead of 42dot's Atria Group, who has been building the Atria end-to-end system, explained that the model wasn't designed as end-to-end from the start — it evolved into its current form from an earlier modular approach, built function by function.
Atria shouldn't be understood, though, as a single neural network directly controlling the entire vehicle. A shared backbone handles perception tasks like object detection and motion prediction, along with driving decisions such as lane-keeping, lane changes, and path generation — but active safety, driving, and parking algorithms are layered on top of that. A guardrail that translates the AI's judgment into safe vehicle control, and a final safety layer that moves the car into a minimal-risk state if something goes wrong, both operate independently, outside the model itself. It's a design that widens what the learning-based model handles while stacking the deterministic safeguards a production vehicle requires on top of it.
The idea of a flywheel itself has been around for two years and eight months. What was missing wasn't test vehicles — it was a production-scale data fleet large enough to drive that cycle. E2E development was already underway, too. What changed under Minwoo Park was tying the data, the validation environment, the organization, and the production plan into a single timeline.
Behind the need for this overhaul lies the exit of the previous leadership. In December 2025, Song Chang-hyun, who had led autonomous driving and the SDV transition, resigned. The company cited "personal reasons," but some media at the time pointed to sluggish progress commercializing autonomous driving and internal disagreement over development direction as the real backdrop. According to a Hyundai Motor Group source familiar with the situation, 42dot had been developing its own independent autonomous driving system largely on its own, but even within Hyundai and its affiliates, people said it was hard to know exactly what 42dot was working on — a sign of how little its work and its role had been shared internally. Song's departure shook up the existing plan, and after Minwoo Park took over, the group's internal capabilities were recombined into a homegrown platform strategy.
Right after starting in February 2026, Park held an all-hands meeting and stressed "one team." Today, the AVP Division and 42dot have merged their development environments, work processes, and information-sharing systems to the point where they function as a single development team in practice. Pulling together capabilities that had been scattered across 42dot and the rest of Hyundai Motor Group turned out to be work that had to happen before the flywheel could. This media day wasn't Atria's birth announcement. It was an account of how the organization and data system behind an Atria that had already been in development got rebuilt to carry it into actual production.







Opening 2028 with Nvidia, Giving Atria One More Year

The two-track strategy Hyundai unveiled shows the nature of this redesign most clearly. What came out of GTC this past March was the broad direction of the partnership — combining Hyundai's SDV and fleet capabilities with Nvidia's DRIVE Hyperion. This media day broke that down into a concrete execution structure and a 2028–2029 timeline: Nvidia supplying the platform and software, the AVP Division handling production, safety, and quality, and 42dot developing the AI models and IP.
The first track is Nvidia. Hyundai plans to integrate Nvidia's in-vehicle AI computing platform and autonomous driving software into its SDV architecture, targeting production of Level 2+ vehicles in the first half of 2028 and Level 2++ vehicles in the second half. The sensor setups used separately by the AVP Division, 42dot, and Motional will also be standardized in stages around Nvidia's Drive Hyperion 10.
The second track is Atria — an E2E-based, in-house system co-developed by the AVP Division and 42dot, targeted for production in the second half of 2029.
Hyundai hasn't locked this timeline in as fixed, though. It said the detailed schedule will be managed flexibly, depending on how ready the first production vehicles are and on each country's certification procedures for autonomous driving.
The two systems aren't competing to see which wins out by some fixed date. Nvidia is the track that allows Hyundai to bring automated-driving systems into series production in 2028, gaining early experience in vehicle integration, validation, and manufacturing. Atria gets an extra year — time to mature its own model, development system, and data flywheel before reaching production in 2029.
What Hyundai is after in the long run isn't any one feature. It's the ability to keep training and improving its own AI — in other words, a foundation for controlling the pace of its own technological evolution. If Nvidia's role is to accelerate the path to production, Atria's is to keep control of the technology and the data in-house. Rather than choosing between building and buying, Hyundai is buying first to enter the market and giving itself more time to build. How far data from Nvidia-based vehicles will feed into Atria's training, and how the two systems' data formats and usage rights will connect, hasn't been disclosed.
The Data Union, too, starts out focused on improving how data is used within the group — Hyundai, Kia, 42dot, Motional. Success on the two-track strategy won't come down to simply hitting two production dates. The experience gathered from different vehicles has to actually speed up the learning of Hyundai's own AI.



What the Footage Showed — and What It Didn't

Hyundai released footage of an SDV testbed carrying Atria AI driving through Seoul. One video followed Minwoo Park and Seonggyun Jeong riding along through expressways, arterial roads, bridges, and city streets while they walked through the development process and the logic behind the system's decisions. Alongside it came three unedited, one-take clips shot in dense morning rush-hour traffic in Gangnam, in heavy bus and truck traffic in Jamsil, and in the rain in Pangyo. A separate video covered ten representative edge cases, including avoiding parked cars, reacting to sudden cut-ins, clearing unprotected left turns, detecting pedestrians in crowded areas, and spotting oncoming traffic on narrow side streets.
The vehicle operated at Level 2++ capability in the released footage, with no driver intervention shown. That confirms Atria can drive through some complex urban situations by reading vehicles, pedestrians, lane markings, and signals together. It's also the first official look at how the data-collection and iterative-learning system Hyundai described is actually being applied to Atria's urban-driving development.
A handful of clips can't establish how mature the technology really is, though. Hyundai didn't disclose the total test-driving distance, the number of interventions or disengagements, what share of the footage shot made it into the public release, the repeat success rate on the same stretch of road, the sensor configuration, or the exact operational design domain. The footage shows where Atria's development stands and how it handles particular scenarios — it isn't proof of overall production readiness.
VLA is still at the simulation-validation stage. The VLA footage did not show the model controlling a vehicle in real time. Instead, it showed the model analyzing recorded driving logs to evaluate its situational understanding and language-based reasoning. Real-vehicle testing is set to begin between late this year and early next year. Within Hyundai's in-house development, Atria E2E remains the main track aimed at production; VLA is the research track preparing for what comes after.
VLA research isn't the only thing that comes after Atria, either. Hyundai is also using the Gwangju pilot program to prepare for real-vehicle validation of Level 4 technologies built on Atria AI, working with the Ministry of Land, Infrastructure and Transport to deploy an Atria AI-equipped SDV pace car in Gwangju by year-end. The complex domestic road scenarios captured there will feed back into the data flywheel for training and validation. If Level 2+ and Level 2++ are the immediate production track, Level 4 validation is running in parallel on the same data foundation.







After the Flywheel

Which is why, in this announcement, the substance of the flywheel matters more than the footage. Hyundai called it "fully operational." The core technologies only started connecting to it this year, and the SER and Data Union are still being added in stages. The SER logs 15 seconds of sensor data, vehicle signals, and Atria's internal processing results before and after an event — triggered manually, by automatic detection of situations like sudden acceleration or braking, or by autonomous driving disengaging — transmits it wirelessly, and runs it through automated annotation. Hyundai didn't say when, or to how many production vehicles, it plans to expand the roughly 40 collection vehicles it currently runs.
In 2024, Hyundai said it would build a car that never stops learning and improving. In 2026, it laid out how the data flywheel meant to drive that learning actually works, along with Atria's production timeline. Between the two announcements, more than technology changed. Leadership turned over, the organization got rebuilt, and instead of trying to hit its timeline through in-house development alone, Hyundai drew up a roadmap that sequences Nvidia and Atria across 2028 and 2029.
If, as Minwoo Park says, the essence of the autonomous driving race is learning speed, then Hyundai's real test starts after the urban-driving footage it just released. The question is how it connects roughly 40 dedicated vehicles to the potential of 7 million annual sales, and how it turns the production experience it gains from Nvidia first into the evolution of its own technology. Hyundai expects that once its autonomous driving systems reach full-scale production, it can overtake competitors' accumulated data volume within five years. Applying that to the 2028 Nvidia-based production date puts the crossover around 2033 — though that isn't a date Hyundai itself has committed to. How close that forecast comes to reality by the time Atria reaches production in 2029 is the real test Hyundai faces.
The Pangyo media day, then, was not the final answer. That answer has to come from the production vehicles of 2028 — and from Atria in 2029.

AEM(오토모티브일렉트로닉스매거진)