Kakao Mobility is upgrading the brain of its self-driving cars with a focus on artificial intelligence (AI). Until now, the driving process in vehicles was handled step by step, but going forward, AI trained on real driving data will grasp surrounding conditions all at once and draw up a driving plan. The company judged that it is difficult for developers to set rules one by one for every unexpected situation that occurs on the road.
Starting at the end of this year, the company will apply this end-to-end (E2E) model and, for safety, have a Rule-based Safety Evaluator verify once more by checking traffic laws and the vehicle's physical limits. The plan is to boost performance by repeatedly training on real driving data collected in southern Seoul's Gangnam through its in-house pipeline.
Kakao Mobility unveiled this Autonomous Driving technology roadmap and data training system at a press briefing held in Gangnam District, Seoul, on the 9th.
Kakao Mobility formed an Autonomous Driving task force (TF) in 2018 and has continued technology development and demonstrations. In 2020, it obtained a temporary driving permit from the Ministry of Land, Infrastructure and Transport with a self-driving car using its own technology, and the following year provided passenger transport service in Pangyo. Since Mar. of this year, in southern Seoul's Gangnam, it has been operating the "Seoul Autonomous Car," which carries real passengers using its in-house Autonomous Driving technology.
◇ from rule-based to E2E… AI draws up the driving plan
Until 2025, Kakao Mobility developed its Autonomous Driving algorithm in three stages: perception, decision, and control. When sensors recognize the surrounding environment such as vehicles or pedestrians, the decision stage sets the driving path, and the control stage converts it into steering and acceleration/deceleration commands. In this method, each stage is separated, making it easy to trace where an error occurred if a problem arises.
However, there is also the problem that developers find it hard to predefine all situations that can occur on real roads. In addition, if vehicles or pedestrians are missed at the perception stage, the wrong information can move to the decision stage and affect driving.
To supplement this, from the first quarter of this year Kakao Mobility has been running in parallel a "planner" that decides which path the vehicle will take next, splitting it into rule-based and AI-based. In complex urban situations that are hard to resolve with an existing rule-based planner, an AI planner trained on driving data decides the route.
On top of that, by the end of this year the company plans to expand the existing AI planner into an E2E model. An E2E model does not split perception and decision into separate modules as before; instead, a single AI model integrates information from sensors to create a driving plan. Im In-ho, leader of the AI Driving part in the Autonomous Driving Development Team at Kakao Mobility, said, "Rule-based is like transcribing a driving manual into a machine line by line, and E2E is like learning driving intuition by watching countless hours of driving up close."
With E2E, people do not have to preset response methods for every situation as rules; it can learn how to cope with situations through real driving data. Even in complex downtowns where the movements of multiple vehicles and pedestrians must be viewed together, it can judge routes by considering the entire situation.
◇ verifying again whether AI's chosen route is safe… advancing AI with driving data
However, the AI's decision is not passed to the vehicle as is. The E2E model generates multiple candidate driving paths—such as deceleration or lane changes—and scores them to select a suitable path, after which the rule-based safety evaluator checks traffic laws and the vehicle's physical conditions. The moment it judges that the AI-selected route could cause discomfort or anxiety for passengers, it opts for the safest method.
The E2E model steadily improves in performance through data. Kakao Mobility is focusing not on simply securing a lot of driving data, but on selecting situations that AI finds difficult and training on them. The Seoul Autonomous Car, which is actually operating in Gangnam, also plays the role of securing this kind of data.
Kakao Mobility began a late-night Seoul Autonomous Car service on Mar. 16 with two vehicles and increased that to six last month. As of the end of last month, the cumulative driving distance exceeded 13,000 kilometers, and the number of completed trips surpassed 2,000. Since Jun., the company has also been separately collecting driving data during daytime hours with heavy traffic and many pedestrians. At night it carries real passengers and runs the service, and during the day it trains on diverse road situations needed to advance the E2E model.
Kakao Mobility's strategy, ahead of the commercialization era of Level 4 fully driverless Autonomous Driving, is to raise its technical capability based on data accumulated through real driving and compete with global Autonomous Driving companies.