Hyundai Motor Group has put its autonomous driving Data Flywheel into full operation, backed by a fleet that sells more than 7 million vehicles a year and a roadmap that targets NVIDIA-based Level 2+ production vehicles in the first half of 2028. The Group presented the system at its HMG Autonomous Driving Media Day, held this week at 42dot headquarters in Gyeonggi Province, South Korea, alongside its first public showing of Level 2++ capability. According to the Group, the flywheel links data collection, AI training, validation and deployment in a continuous loop intended to shorten development cycles.
Highlights
- NVIDIA-based Level 2+ production vehicles targeted for the first half of 2028, with Level 2++ following in the second half of 2028
- Production of Atria AI-powered Level 2++ vehicles targeted for the second half of 2029
- Approximately 40 dedicated data-collection vehicles operating around the clock, drawing on sales of more than 7 million vehicles annually across roughly 190 countries and regions
- Real-world Level 4 pilot using the Atria AI-equipped SDV Pace Car planned for Jeonnam-Gwangju Special Metropolitan City by year-end
What Is the Dual-Track Strategy?
In March, the Group announced a collaboration with NVIDIA built on a dual-track approach: rapid deployment of NVIDIA’s autonomous driving technologies on one track, development of proprietary AI capabilities on the other.
Track one integrates NVIDIA’s vehicle AI computing platform and autonomous driving software into the Group’s software-defined vehicle (SDV) architecture. The Group says this track prioritizes accelerated deployment while building a foundation for scalable, data-driven development. It includes progressive standardization of sensor systems across Hyundai Motor, Kia, 42dot and Motional around NVIDIA DRIVE Hyperion 10, which the Group expects to yield more consistent data collection for AI training and validation.
Track two is Atria AI, a proprietary end-to-end (E2E) autonomous driving system jointly developed by the Group’s Advanced Vehicle Platform (AVP) Division and 42dot under an integrated development framework. Capabilities are to progress in phases based on real-world driving data collected from production vehicles.
| Milestone | Technology base | Production target |
|---|---|---|
| Level 2+ | NVIDIA solutions | First half of 2028 |
| Level 2++ | NVIDIA solutions | Second half of 2028 |
| Level 2++ | Atria AI | Second half of 2029 |
“Autonomous driving competition is no longer about comparing specific features. Competitiveness is determined by how much data you secure, how quickly you learn and how effectively you can reflect those results in actual products and services. At its core, autonomous driving competitiveness comes down to having systems that enable continuous, rapid learning. Hyundai Motor Group will develop autonomous driving technology that customers can trust, based on a virtuous cycle of data, AI and validation. Our goal is to ensure the safety and quality levels customers can trust while we learn and improve rapidly,” said Minwoo Park, President and Head of the AVP Division at Hyundai Motor Group and CEO of 42dot.
How Does the Data Flywheel Work?
The Data Flywheel operates as a loop in which data collected from vehicles trains and validates AI models, and improved models are deployed back to vehicles to generate new data. The Group points to data acquisition as its starting advantage: Hyundai Motor and Kia sell more than 7 million vehicles annually across approximately 190 countries and regions, and the Group currently runs about 40 dedicated data-collection vehicles around the clock.
The datasets cover routine driving as well as edge cases, including road construction zones, severe weather, abrupt lane changes and emergency maneuvers, vehicles parked on side streets and narrow roads, and dense urban traffic.
The Group notes that data volume alone does not determine AI performance, and says it has been integrating new techniques into the flywheel since earlier this year:
- Hard Example Mining automatically identifies driving situations that AI models find difficult to recognize or interpret and prioritizes those scenarios for training.
- Continuous Training Pipeline feeds newly acquired real-world and validation data into model training on an ongoing basis, with vehicle evaluation findings fed back into data collection and model development.
- Virtual Validation Technology reconstructs real-world driving data into three-dimensional environments, using techniques such as 3D Gaussian Splatting, to recreate scenarios that are difficult or unsafe to reproduce on the road and to verify that newly trained models do not degrade existing performance.
- Follow-the-Sun Development connects development centers in South Korea and the U.S. so teams sequentially carry out data collection, issue analysis and model improvement across time zones for continuous 24-hour development.
- Special Event Recorder (SER) Integration is being phased into the flywheel. SER automatically records and stores significant events that occur during autonomous driving, with the data used primarily for model training. The Group is also exploring ways for SER to identify challenging edge cases and automatically secure the relevant datasets.
- Data Union Ecosystem establishes a framework based on standardized sensor architectures and data structures so data generated across multiple vehicles and organizations can be accumulated under common standards and used for AI training. The initial focus is on Hyundai Motor, Kia, 42dot and Motional, with future expansion planned.
Why Is Real-World Level 4 Validation Part of the Plan?
Alongside mass-production development, the Group is pursuing real-world Level 4 validation. In partnership with South Korea’s Ministry of Land, Infrastructure and Transport, it plans to deploy the Atria AI-equipped SDV Pace Car in Jeonnam-Gwangju Special Metropolitan City by year-end.
The pilot will operate autonomous vehicles on Korean roads with complex traffic dynamics and unpredictable variables rather than on controlled test tracks. Driving scenarios and contingency situations encountered during deployment will be captured and fed back into the Data Flywheel. The Group says the approach supports both the completeness of its Level 2+ driver assistance technology and the validation of Level 4 capabilities.
“Autonomous driving competitiveness depends not on how much data you secure, but on how rapidly you can connect data to learning, validation and performance improvement. Through the data flywheel system now in full operation, we expect continuous improvement of issues identified on real roads and rapid technology advancement,” said Junghyun Kwon, Executive Vice President and Head of Hyundai Motor Group’s Autonomous Driving Development Center, and 42dot Autonomous Driving Division Lead.
“Good autonomous driving AI ultimately starts with high-quality data. Hyundai Motor Group continuously improves the performance and maturity of Atria AI through an integrated development cycle that spans data collection, model training and real-world vehicle validation,” said Seonggyun Jeong, Group Lead of 42dot’s Atria Group.
How Does VLA Technology Fit In?
Beyond its E2E work, 42dot is developing Vision-Language-Action (VLA) technology, which combines visual recognition, language-based reasoning and action generation in a single framework. 42dot plans to develop VLA models alongside its existing E2E models to improve stability and scalability while offering solutions tailored to different vehicle hardware specifications and customer needs.
According to 42dot, conventional E2E models connect driving inputs directly to vehicle actions, while VLA models add language-based situational understanding and reasoning, which the company expects to improve decision-making and explainability in complex scenarios. 42dot also expects VLA to help handle rare driving situations by drawing on large-scale pre-trained knowledge that is difficult to learn from driving data alone.
42dot’s VLA-based technology is currently in the simulation-based model validation stage. The Group plans to activate the full development process, including real-vehicle testing, from late 2026 through early 2027, with issues found on the road fed back into learning and validation. 42dot released development footage showing the VLA model outputting its reasoning in natural language, with on-screen text describing both the action taken and the rationale behind it.
“VLA is a core technology for implementing Physical AI where AI goes beyond simply driving to understand situations, reason through them and act. Starting with autonomous driving, it will provide the foundation to expand into diverse fields such as robotics and beyond,” said HeeSeok Lee, Group Lead of 42dot’s Trion Group.
What Does the Seoul Urban Driving Footage Show?
The Group released video of an Atria AI-equipped SDV Testbed navigating Seoul traffic without driver intervention at Level 2++ capability. The footage falls into three categories, all published on the Group’s YouTube channel.
- Executive Ride-Along: Minwoo Park and Seonggyun Jeong ride through central Seoul across expressways, major thoroughfares, bridges and urban streets while discussing Atria AI’s development process, capabilities and decision-making. The ride-along video also shows how the vehicle perceives surrounding traffic and responds in real time.
- One-Take Urban Driving: Three unedited sequences, each roughly two to four minutes long with no edits other than playback speed, cover congested morning rush hour in Gangnam, high-density traffic with frequent bus interactions in Jamsil, and rainy conditions in Pangyo.
- Edge Case Handling: The edge-case video covers 10 representative scenarios, including avoiding vehicles parked along the roadside, responding to sudden cut-ins, navigating unprotected left turns, detecting pedestrians in congested areas and identifying oncoming vehicles on narrow neighborhood roads.
“Autonomous driving is the flagship Physical AI technology and sits at the center of the automotive industry’s transformation into an AI industry. By combining the Group’s global mass production capabilities with 42dot’s AI and software technology and data flywheel-based learning system, we will implement trustworthy autonomous driving technology for our customers. Development speed and safety are not conflicting values. The more exceptional situations we discover and learn from, the safer autonomous driving becomes. Hyundai Motor Group will enhance both development velocity and safety, maintaining the principle of applying only thoroughly validated technology to our vehicles,” Park said.
