Hyundai bets slow and steady will win self-driving race, expects to overtake Tesla in five years

PANGYO, Gyeonggi – Hyundai Motor Group is “deliberately delaying” the timeline of its autonomous driving system, betting that a slower rollout focused on data collection and machine learning will let it overtake rivals such as Tesla's Full Self-Driving (FSD) within five years.

The Korean automaker on Friday disclosed development details and a timeline including its self-driving AI, called Atria AI, which relies on a method akin to a student's error notebook — cataloging mistakes and retraining the system to correct them.

“We believe Atria AI could move faster if we pushed with everything we have,” said Park Min-woo, head of Hyundai Motor Group's Advanced Vehicle Platform Division and chief executive of 42dot, the group's autonomous driving affiliate, at a press event in Pangyo, Gyeonggi.

“But we made a strategic decision to slow down. Rushing into mass production now risks delivering a middling Level 2++ system, far short of our real target,” Park said. 

Hyundai joined with Nvidia to develop a 2++ system — which refers to the six levels of automation, from 0 to 5, as defined by global standards organization SAE International — for production in the second half of 2028. A Level 2++ system is roughly equivalent to Tesla's current FSD. A separate line of Hyundai vehicles equipped with its in-house Atria AI is slated to follow in the second half of 2029.

Atria AI is built on what the company calls a “Data Flywheel,” a concept that as more vehicles hit the road, more data accumulates, and that data sharpens the AI, which in turn improves the vehicles. 

Hyundai currently runs about 40 dedicated data-collection vehicles around the clock to build out its driving dataset.

The most distinguishing feature is its ability to isolate edge cases — the unpredictable scenarios that most frequently challenge autonomous systems such as construction zones, severe weather, abrupt lane changes, vehicles parked haphazardly on side streets — and feed them back into training on a loop.

Engineers use a technique called Hard Example Mining to automatically flag driving situations the model struggles with, then apply a Continuous Training pipeline to fold those cases back into the learning process. 

A parallel system, the Special Event Recorder, automatically logs data when it detects abrupt acceleration or braking, erratic following distances or moments when a driver or the system itself disengages autonomous mode.

“Hyundai sells 7 million vehicles a year across 190 markets, which means we can accumulate data at scale and at this pace, we can overtake what our rivals have built in five years,” Park added. 

“The competition in autonomous driving is no longer about the speed. It comes down to how much data you can gather, how fast you can learn from it and how quickly that translates into real products and services.”

Separately, 42dot said it has begun developing a next-generation autonomous driving system built on a vision-language-action, or VLA, framework that integrates visual perception, linguistic reasoning and vehicle behavior generation.

In one clip shown by the company, as the vehicle changed lanes, on-screen text displayed the system's explanation of the maneuver and the logic behind it.

“Through 2027, we'll focus on bringing driving, parking and active-safety features to a high level of polish at the product level,” said Kwon Jung-hyun, who leads the autonomous driving division at 42dot. “2028 will be devoted to gathering vast amounts of edge-case data to build a genuinely safe product. And 2029 is when we'll prepare for safety certification, not just in Korea but in North America and Europe as well.”

BY SARAH CHEA   [[email protected]]