(Seongnam = Yonhap News) Reporter Hong Gyu-bin = One-way three-lane road in Songpa-gu, Seoul. When a sedan tried to turn right and enter the main road, the car behind slowed down and gave up its space. The sedan driver showed his gratitude to the car behind him by flashing his emergency lights.
However, the person receiving the thanks was not the driver of the car behind. This is because the decision to give way was not the Hyundai Motor Group employee in the driver's seat, but the autonomous driving brain 'Atria AI'.
Hyundai Motor Group held an autonomous driving media day at the 42dot headquarters in Seongnam, Gyeonggi-do on the 11th and unveiled the first video of a test bed vehicle equipped with Atria AI driving in downtown Seoul.
The test bed vehicle, which ran in Gangnam-gu and Songpa-gu, Seoul, accurately recognized surrounding vehicles and pedestrians not only on major roads congested with traffic and signals, but also in narrow alleys, and skillfully adjusted direction and speed.
It predicted incoming vehicles, preemptively slowed down, and secured a safe distance from vehicles stopped on the shoulder. Even when another car jumped out between the stopped cars, it recognized this and slowed down.
I turned right safely according to the right turn-only signal, which is rare, and made a stable unprotected left turn after all oncoming vehicles with flashing lights had passed. At the T-shaped intersection, I checked the surroundings and turned left.
There was no blockage even in the crowded alleys. It accurately avoided stopped trucks and accurately recognized bicycles and pedestrians. At an intersection, I made a smooth left turn by identifying the available space and the path of the car on the other side.
Park Min-woo, head of the AVP division (president) of Hyundai Motor Company and Kia Motors, also evaluated the performance of Atria AI without hesitation in a video released together.
When the car joined the main road only at the end of the acceleration lane, he said, "I think it was a mistake in judgment, but I don't like it. If I had driven it, I would have slowed down (in advance) and gone behind the car next to me."
Regarding the scene where the car stopped and restarted when another car cut in front, he said, "It was the car that was going to come in anyway, so I wonder what it would have been like if there had been less competition with that car (in the first place)."
There are places these Atria AI vehicles must pass through before getting on the road. This is the ‘Vehicle Workshop’ at the 42Dot Pangyo office building.
Here, a 'smoke test' is performed to check whether the autonomous driving function is running normally, communication with the vehicle controller is smooth, and basic driving functions are working. 42Dot operates workshops in two locations: Pangyo headquarters (1,818 m2) and Sujeong-gu, Seongnam (9,889 m2).
At the workshop I visited that day, there was an SDV test bed vehicle that traveled around downtown Seoul. The appearance is not much different from the mass-produced IONIQ 6, but there are 8 cameras and 1 radar installed throughout the vehicle. Atria AI makes driving decisions based on the information recognized by these sensors.
Next to the test bed vehicle, there was an Ioniq 5-based vehicle with LiDAR equipment on the roof. Forty2dot operates about 40 such vehicles, which collect actual road data to train Atria AI.
In addition, Hyundai Motor Group is building a 'data flywheel' system that uses the data collected in this way for AI learning and verification, and connects the results to model improvement and new data acquisition.
Although the amount of accumulated data collected is evaluated to be small compared to global competitors, Hyundai Motor Group emphasized that the performance of autonomous driving AI is not determined solely by the amount of data.
The explanation is that model performance can be determined by how well AI selects difficult situations (edge cases) from massive driving data and focuses on learning them.
Hyundai Motor Group is accelerating the data flywheel through technologies such as 'Special Event Recorder' (SER), which automatically records important situations that occur during autonomous driving, 'Hard Example Mining', which automatically selects driving situations that AI finds difficult, 'Continuous Training', which repeatedly improves AI by reflecting new data, and '3D Gaussian Splatting', which reproduces video data in a 3D virtual environment.
We are also building a 'data union' system that expands the scope of data use within and outside the group based on a standardized sensor system and data structure.
