Hyundai Motor Group launches 'data flywheel' to sharpen autonomous driving AI

Data flywheel kicks into gear

Over 40 data-collection vehicles run day and night

Special Event Recorder automatically logs edge cases for AI training

Nvidia partnership and Artria AI form two-track strategy

Group-wide sensor standardization underway; Level 4 pilot set for Gwangju

Park Min-woo, head of Hyundai Motor and Kia's AVP division and CEO of 42dot, delivers a welcome address at the Hyundai Motor Group Autonomous Driving Media Day held Friday at 42dot's headquarters in Pangyo, Seongnam, Gyeonggi Province. [Hyundai Motor Group]

Hyundai Motor Group said Friday it is fully activating a "data flywheel" — a self-reinforcing cycle that feeds real-road driving data into AI training and validation before pushing improved models back into vehicles. Rather than simply accumulating large volumes of data, the group is focusing on rapidly identifying and repeatedly training on "edge cases," the unusual situations that trip up AI systems, to accelerate technological advancement.

The group unveiled its autonomous driving development strategy and the current state of its data flywheel at the Hyundai Motor Group Autonomous Driving Media Day, held Friday at 42dot's headquarters in Pangyo, Seongnam, Gyeonggi Province.

"The autonomous driving race is no longer about which company has the better feature," said Park Min-woo, head of Hyundai Motor and Kia's AVP division and CEO of 42dot. "What determines competitiveness is how much data you can secure, how quickly you can learn from it, and how fast you can reflect the results in actual products and services."

Targeting edge cases for repeat training — quality over quantity

The data flywheel works as a virtuous cycle: vehicles collect driving data, which feeds AI training and validation; the improved model is then deployed back into vehicles, generating new data in turn. The more vehicles operate, the more data accumulates, and as AI performance improves, the system encounters a wider range of driving situations — producing yet more data.

Hyundai Motor Group currently operates more than 40 dedicated data-collection vehicles around the clock. The fleet focuses on capturing edge cases that arise on real roads — construction zones, adverse weather, sudden lane changes, and vehicles parked on narrow side streets.

The core objective is intensive training on situations that AI finds particularly difficult. The group uses "hard example mining" to automatically identify edge cases, then runs them through a continuous training pipeline for repeated learning. Situations that are hard to reproduce on actual roads are recreated in virtual environments for validation.

A Special Event Recorder, or SER, is another key component of the flywheel. It automatically logs incidents during autonomous driving — sudden acceleration or braking, unstable following distances, and disengagement of the autonomous driving function — and feeds that data into AI training. The group plans to eventually develop the system so that AI can assess its own weaknesses and automatically collect the data it needs.

Speed through outside tech, control through in-house AI — a two-track strategy

Researchers at 42dot work on autonomous driving technology at the Vehicle Workshop. [Hyundai Motor Group]

Building on this data infrastructure, Hyundai Motor Group is pursuing a two-track strategy that combines external technology with in-house development. The group plans to apply Nvidia-based Level 2+ features to mass-production vehicles in the first half of 2028, then launch Level 2++ vehicles powered by its own Artria AI system in the second half of 2029.

In the near term, the group will leverage Nvidia's automotive AI computing platform and autonomous driving software to accelerate mass production. At the same time, the AVP division and 42dot are jointly developing Artria AI — an end-to-end autonomous driving system — to internalize core technology for the long term.

Data and know-how gained during the earlier Nvidia-based mass production process will also feed into the refinement of Artria AI. "Most of the problems that arise during the Nvidia solution mass production process apply equally to Artria mass production," said Kwon Jeong-hyeon, executive vice president and head of Hyundai Motor and Kia's Autonomous Driving Development Center. "We plan to use that know-how and data in developing Artria."

Alongside the end-to-end model at the heart of Artria AI, 42dot is also developing a next-generation vision-language-action model, or VLA. While the end-to-end approach directly maps sensor inputs — such as camera feeds — to vehicle actions, VLA adds a layer of language-based situational understanding and reasoning. The technology is designed to help vehicles not only decide how to move, but also understand and explain why they made a given judgment.

42dot believes this will improve the system's ability to handle rare situations that are difficult to learn from driving data alone, as well as its overall explainability. "VLA is the core technology for realizing physical AI — where AI goes beyond simply driving to understanding situations, reasoning, and acting," said Lee Hee-seok, group leader of 42dot's Trion Group. "It will serve as a foundation that can expand from autonomous driving into robotics and other fields."

Unified sensor architecture — and Gwangju Level 4 pilot data to feed the loop

Hyundai Motor Group Chairman Euisun Chung (left) and Nvidia founder and CEO Jensen Huang pose for a photo at Nvidia's headquarters in Santa Clara, California, on July 24 (local time). [Yonhap]

To improve data utilization efficiency, Hyundai Motor Group is also gradually integrating the sensor architectures used across Hyundai Motor, Kia, 42dot and Motional. Each organization currently uses different sensors depending on its development goals and history, and differing sensor configurations produce data in different formats — requiring separate processing before the data can be used for AI training.

The group plans to standardize around Nvidia's Drive Hyperion 10 platform. Going forward, it aims to accumulate data from different vehicles under a common standard and, over the longer term, build a foundation that can incorporate data from external partners using the same sensor ecosystem.

The standard sensor configuration under consideration includes 10 cameras, one front-facing radar and 12 ultrasonic sensors. While mounting positions may vary by vehicle class, the goal is to apply the same sensors across as many models as possible.

Nvidia's Orin X and Thor chips are among the in-vehicle computing options being evaluated. "We want to secure a minimum hardware specification that allows customers to receive software updates to the latest model seven or eight years after purchasing the car," Park said. "At this point, Orin X looks like the optimal choice."

Real-road Level 4 autonomous driving pilots are also tied to the expansion of the data flywheel. Working with the Ministry of Land, Infrastructure and Transport, Hyundai Motor Group plans to deploy Artria AI-equipped SDV pace car-based test vehicles in Gwangju by year-end. Driving scenarios and unexpected-situation data gathered during the pilot will feed back into AI training and performance improvement. The group said it intends to use the process to raise the maturity of both its Level 2+ and above mass-production driver-assistance technology and its future Level 4 systems.

"Autonomous driving competitiveness depends less on how much data you can secure and more on how quickly you can connect that data to training, validation and performance improvement," Kwon said.

eyre@heraldcorp.com