Kakao Mobility pushes toward single end-to-end self-driving model using Gangnam late-night data

Gangnam night pilot frames Kakao Mobility's data strategy

Kakao Mobility, the South Korean ride-hailing unit spun out of Kakao, told reporters that its autonomous passenger service in Seoul's Gangnam district has accumulated more than 13,000 kilometers of driving and roughly 2,000 completed trips since launch, data the company now treats as the core training asset for its next technology step. The service, which began in March, initially ran from 10 p.m. to 5 a.m. with two vehicles and was expanded to six in August, with passengers giving an average rating of 4.97 out of 5 and one minor rear-end collision reported, in which another vehicle struck the autonomous car with no injuries. In-ho Lim, leader of the AI Driving Part at Kakao Mobility's autonomous driving development team, said at a media session on Tuesday that the firm's development focus is shifting from amassing large driving datasets to selecting difficult situations, arguing that "For the AI model, driving for one month in Gangnam is far better than driving for one year in a low-difficulty area."

Five-level taxonomy isolates rare edge cases

To prioritize that selection, the company classifies driving situations into five levels: Level 1 follows signals and rules, Level 2 focuses on defensive driving, Level 3 reads relationships with nearby vehicles and handles turns and lane changes, Level 4 responds to cutting-in vehicles and motorcycles, and Level 5 covers hard-to-predict risks. Lim said Levels 4 and 5 are the segments that distinguish autonomous driving capability and that data from those segments is the rarest, adding that performance gains from additional data flatten beyond a certain point so "it is data that creates that gap." The vehicles automatically detect and store such situations, and the company treats any switch into manual mode as an exceptional situation, recording the segments before and after together; cited examples include a vehicle making an illegal U-turn at dawn, a segment where the car had to cross the center line because of sewer pipe construction, and an intoxicated person darting out onto a main road near Gangnam Station. The collected records are processed through a five-step cyclical pipeline in which AI now handles frame-by-frame review work that previously required human staff, completing the triage within tens of minutes after server upload.

End-to-end integration targeted by year-end

Kakao Mobility said it plans to integrate perception and decision-making into a single end-to-end (E2E) model within this year, replacing the hybrid approach that paired E2E components with rule-based modules. The company explained the rule-based approach divides driving into separate sensing, perception, planning and control modules that follow human-defined rules, which makes decisions auditable but requires developers to anticipate nearly every road scenario, while an E2E model processes sensor inputs through one AI model and directly outputs vehicle control without defining intermediate steps. A simulator is being used to compensate for the absence of accident data in the training set, and the company said it would keep rule-based safety checks in place alongside the new model rather than removing them.

Origins inside Kakao Mobility's autonomy team

Kakao Mobility formed its autonomous driving team in 2018 and began testing the technology in 2020, building the in-house stack now deployed in Gangnam rather than partnering with an outside robotaxi operator. The press event was held at the Korea Automobile Manufacturers Association headquarters in southern Seoul, where the company also showcased a self-driving test vehicle alongside its data and platform roadmap, positioning the late-night Gangnam corridor as the controlled environment in which its E2E model is being validated before any potential move into broader everyday service.

Unresolved rollout questions

The company has not disclosed a timeline for expanding the service beyond the current 10 p.m. to 5 a.m. Gangnam window, scaling the fleet beyond six cars, or setting a date for a commercial robotaxi launch, and it remains to be seen how the E2E model will perform once rule-based safety checks are reduced or removed. The next verifiable milestone is the year-end target for integrating perception and decision-making into a single E2E model, alongside continued accumulation of late-night Gangnam edge-case data.