Hyundai doubles down on autonomous driving with NVIDIA now, Atria AI in 2029

As the global automotive industry moves deeper into the age of artificial intelligence, Hyundai Motor Group has outlined its latest autonomous-driving plans at an HMG Autonomous Driving Media Day in South Korea.

The group has pushed back the production timeline for its in-house Atria AI system, but says it will keep development moving on two fronts while building a much larger pool of real-world driving data.

Dual-track strategy: NVIDIA now, Atria AI later

Under the latest plan, Atria AI, an end-to-end driving-assistance system being developed by Hyundai together with software company 42dot, is now scheduled to enter production in the second half of 2029.

Rather than waiting for the in-house system to be ready, Hyundai will use a two-track approach in the meantime.

The first track centres on NVIDIA technology. Hyundai plans to integrate NVIDIA's validated vehicle AI computing platform and autonomous driving software into its software-defined vehicle (SDV) architecture.

The first software-defined vehicle using the system is planned for the first half of 2028, with Level 2+ capability and hands-off driving on motorways. An upgraded Level 2++ system is due in the second half of the year, extending the capability to more demanding urban roads.

At the same time, Hyundai's AVP (Advanced Vehicle Platform) Division and 42dot will continue developing Atria AI.

The system does not rely on LiDAR, instead combining cameras, millimetre-wave radar and ultrasonic sensors. It is due to reach production in the second half of 2029, with Hyundai aiming to keep the core technology under its own control and eventually move towards L4 autonomous driving.

The dual-track strategy gives Hyundai a way to bring more advanced driving assistance to customers in the near term while continuing to develop its own system for the longer term.

Building a Data Flywheel

Hyundai believes the next stage of autonomous-driving development will depend as much on data and how quickly the AI can learn as on individual functions.

The group is building what it calls a Data Flywheel, linking real-world data acquisition with AI training, validation and deployment.

Hyundai and Kia sell more than 7 million vehicles a year across more than 190 countries and regions, giving the group a huge potential source of driving data.

Around 40 dedicated data-collection vehicles are currently on the road around the clock, gathering information from difficult situations such as roadworks, severe weather, sudden lane changes and cars parked illegally on narrow streets.

Hyundai is also using several tools to turn that data into better-performing AI.

Hard Example Mining identifies difficult scenarios where the system is most likely to make mistakes, allowing those cases to receive more attention during training.

For continuous training and virtual validation, Hyundai is using technologies such as 3D Gaussian Splatting to recreate real roads in a virtual 3D environment, where edge cases can be tested without putting a vehicle on the road.

The group is also using a Follow-the-Sun development model linking its Korean and US teams, allowing data analysis and model optimisation to continue across time zones.

A Data Union Ecosystem brings together Hyundai, Kia, 42dot and Motional, with the aim of sharing standardised data across different vehicle programmes rather than keeping it within separate organisations.

By the end of this year, Hyundai also plans to begin Level 4 road testing in Gwangju, South Korea, working with government authorities to gather more validation data from real traffic.

VLA model and physical AI

42dot is also working on a next-generation visual-language-action (VLA) model.

Conventional driving systems take what the vehicle sees and turn it directly into an action. A VLA model adds a language-based reasoning layer, allowing the system to interpret unusual situations and explain what it is seeing.

Hyundai sees this as part of the technology needed to move towards physical AI and robotics, rather than simply improving existing driver-assistance functions.

The system is currently being tested in simulation, with full vehicle testing planned to begin between late 2026 and early 2027.

One-take Seoul drive puts Atria AI to the test

Hyundai has also released a one-take driving video on its official YouTube channel to show how its Atria AI system handles real urban traffic.

Senior executives joined the drive themselves, riding in the test car through central Seoul, including motorways and major bridges. The footage also captures the morning rush in Gangnam, heavy traffic around Jamsil and wet-weather driving in Pangyo.

The video highlights 10 challenging urban scenarios, including avoiding illegally parked vehicles, dealing with sudden cut-ins, making unprotected left turns and passing on narrow residential roads.

With its dual-track strategy and data flywheel working together, Hyundai Motor Group is moving towards a more AI-focused mobility business, while aiming to bring advanced driving technology to customers without compromising on safety.