Hyundai Motor Company: “Data Flywheel Gap to Widen in 3–4 Years”... Four Group Companies Put 'Data Union' Into Operation
- Input
- 2026-11-20 07:00:00
- Updated
- 2026-11-20 07:00:03

Lee made the remarks on the 19th during a special lecture at the autumn conference marking the 20th anniversary of the Korean Auto-vehicle Safety Association, held at the Kim Daejung Convention Center in Seo-gu, Jeonnam-Gwangju.
A data flywheel is a cycle that accelerates as the process of acquiring and analyzing data, then deploying AI models trained on it in vehicles, is repeated. “A flywheel is hard to get moving at first, but the more it turns, the more momentum it gains,” Lee said. He emphasized that “a large gap will emerge between companies that have properly built a data flywheel and those that have not, and the longer time passes, the harder that gap will be to close.”
■“99% of Driving Data Is Routine... Its Value Lies in the 1%”
Lee explained that four elements—scale, quality, efficiency, and operations—must all work together for the data flywheel to run properly. If even one of them does not function as it should, the entire flywheel cannot turn, he said.In terms of scale, he cited as a strength the company's sales of at least 7 million vehicles a year across at least 190 regions. “If we can extract data from this large a fleet of vehicles and use it for AI, the synergies would be enormous,” Lee said.
Hyundai Motor Group operates a Data Union in which four companies—Hyundai Motor Company, Kia Corporation, 42dot, and Motional—collect autonomous-driving data using the same sensor and data-acquisition systems. Lee also said the Data Union could expand globally in the future.
On quality, Lee emphasized “1% data.” “If you travel from Seoul to Busan, nothing unusual happens on at least 99% of the route,” he said. “The truly valuable data comes from less than 1% of the entire route.” Hyundai Motor Company has built a “Special Event Recorder (SER)” system to capture it.
There are four collection methods. Test drivers and others can press a button to record a special situation themselves (voice recordings are automatically converted to text); the system automatically records the segments before and after events such as hard braking or sharp steering; it automatically records when autonomous driving is disengaged because of a system problem during a test or demonstration; and it collects situations selected as necessary under internal policies.
The collected data is not used for training right away. An “issue triage center,” made up of professional test drivers and autonomous-driving developers, classifies root causes and sends them to the teams responsible. Only after those teams resolve an issue through data training, algorithm improvements, expansion of virtual data, or other measures is the solution incorporated into a model and deployed in vehicles. To improve efficiency, real-world data is reconstructed in a three-dimensional (3D) environment, then expanded as much as possible by changing conditions such as rain or nighttime, or by adding other objects. For operations, global research centers in Korea and the United States, among others, use time differences to resolve issues around the clock in a “Follow the Sun” approach.
■“Layered Safeguards Even With Rapid Deployment... Closed-Loop Simulation Is Important”
Lee cited as a key performance indicator (KPI) for the data flywheel “the time from when the 1% data comes in until the first root cause is resolved.” “The time varies depending on how difficult the problem is, but what matters is quickly analyzing the cause when an issue arises,” he said.Lee raised the question of whether rapid deployment of models is safe, then explained the multiple layers of safeguards. Even after all internal validation stages—including tests in real vehicles, on a proving ground (PG), and in simulations—the company assumes that unexpected situations may arise because of the nature of data-driven AI models. Accordingly, it adds a “guardrail” based on a scaled-down version of the existing rule-based approach on top of the AI model and applies additional driving-safety algorithms. At Level 4, it adds safeguards such as remote assistance.
Based on his experience, Lee identified simulation as the area that most urgently needs to be advanced. Data on events such as pedestrian-vehicle collisions are difficult to obtain in real life, and collecting them also carries a risk of accidents. He explained that “closed-loop simulation,” which runs physics simulations based on real-world data, is crucial for safety validation—not an open-loop approach that sets a correct answer in advance and makes the vehicle follow it.
Prefacing his remarks as “a personal opinion,” Lee said, “When buying a car today, people look at its model year, engine, or design. But because the version of the data flywheel and the data used directly affect a vehicle's performance and safety, I think the data flywheel could become one of the criteria for buying a car in the future.” He added, “Data Unions seem to be forming in greater numbers in China and the United States,” and “by looking at which parties are connected through data partnerships, we will be able to see which side has the technological advantage.”
[email protected] Kim Dong-chan Reporter