现代汽车集团发布自动驾驶长期战略：以“数据飞轮”和VLA模型追赶，2028年量产L2+/L2++、2033年数据量超特斯拉

Hyundai Motor Group plans to accelerate its lagging self-driving development with tools including a "data flywheel" and vision-language-action (VLA) models. The group expects to surpass the accumulated data volume of rivals such as Tesla by 2033.
Hyundai Motor Group unveiled the strategy at an Autonomous Driving Media Day held on the 11th at the headquarters of 42dot in Seongnam, Gyeonggi Province. Key figures in the group's self-driving efforts attended, including Park Min-woo, president and head of the AVP Division at Hyundai Motor and Kia and chief executive of 42dot; Kwon Jung-hyun, executive vice president and head of the Autonomous Driving Development Center at Hyundai Motor and Kia and AD division lead at 42dot; Jung Sung-kyun, Atria group lead at 42dot; and Lee Hee-seok, Trion group lead at 42dot. "Hyundai Motor Group will build differentiated self-driving competitiveness by combining the capability to develop software and AI in-house with the capability to deliver that technology through vehicles and customer experiences on a global scale," Park said.
The core concept the group presented was the data flywheel. It refers to a virtuous cycle in which data is collected from vehicles, used to train and validate artificial intelligence models, and the improved models are then deployed back to vehicles to gather new data. "Competition in self-driving is no longer about which specific feature is better," Park said. "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 reflect those results in actual products and services."
Hyundai Motor Group sells more than 7 million vehicles a year in about 190 countries. Once self-driving systems begin to be installed, that scale gives it an environment for rapidly gathering real-world driving data. "Once mass production of self-driving systems begins in earnest, we will surpass the volume of data accumulated by competitors within five years," Jung said.
The group plans to apply Level 2+ self-driving functions based on Nvidia solutions to mass-produced vehicles in the first half of 2028, and to launch mass-produced vehicles with Level 2++ in the second half of the same year. By the group's own calculation, that would put it past rivals' data volume in 2033. Level 2+ is driver-assistance technology that allows hands-off driving under specific conditions such as highways while the driver supervises, and Level 2++ extends that capability to general roads and urban sections.
The group will expand the technology in stages, including mass-producing Level 2++ in the second half of 2029 based on its own self-driving platform, Atria AI, rather than Nvidia solutions. "By 2027 we will implement a wide range of functions, including driving, parking and active safety, at a high level of completeness," Kwon said. "After that, 2028 should be seen as a period for gathering a large volume of edge-case data — exceptional driving situations — to build a genuinely safe product, and 2029 as a period for preparing to obtain safety certification not only in Korea but also in North America and Europe."
The group also released for the first time video of a software-defined vehicle (SDV) testbed car equipped with Atria AI driving through a complex urban area at a Level 2++ standard. The testbed vehicle, based on the Ioniq 6, is fitted with eight cameras and one radar. Working with the Ministry of Land, Infrastructure and Transport, the group plans to deploy autonomous SDV pace cars in Gwangju, Jeolla Province, from the end of this year to check the performance and stability of the technology on domestic roads.
Hyundai Motor Group currently operates about 40 vehicles dedicated to data collection, running day and night. Since early this year it has been rapidly improving the performance of its AI models through core technologies including hard example mining, continuous training, a special event recorder and a data union. Hard example mining automatically selects driving situations the model struggles with, and continuous training feeds those into repeated learning to shorten development cycles. The special event recorder automatically logs relevant data when the system detects specific situations, such as avoiding a vehicle parked on the shoulder or sudden acceleration and braking, or when the driver or the system disengages self-driving functions. The data union is a framework that standardizes on Nvidia Hyperion 10 so that data can be shared across Hyundai Motor and Kia, 42dot and Motional, the group's U.S. robotaxi joint venture, which currently operate on different sensor specifications.
The group also plans to add VLA models to its self-driving models. VLA combines visual recognition, language-based reasoning and action generation in a single model. Because it introduces language-based situational understanding and reasoning, it can improve judgment and explainability in complex driving situations. "VLA is a core technology for realizing physical AI, in which the AI goes beyond simply driving to understand a situation, reason about it and act," Lee said. "Starting with self-driving, it will be the foundation for expanding into various areas such as robotics."