"Experiencing various unexpected situations on complex roads like Gangnam is more beneficial for enhancing autonomous driving AI performance than driving for a long time on easy roads."
Im In-ho, leader of the AI driving team at Kakao Mobility, made this statement on September 9 during a Q&A session at a Kakao Mobility autonomous driving study held at the Korea Automotive Mobility Industry Association in Seocho, Seoul. He outlined the goal of introducing an end-to-end (E2E) model by the end of the year, which will process everything from sensor input to driving control using a single AI model.
Kakao Mobility is currently collecting autonomous driving data in the Gangnam area from 10 p.m. to 5 a.m. The late-night service, which began in March with two vehicles, has expanded to six vehicles as of last month, accumulating over 13,000 kilometers of driving distance and completing 2,000 rides. Since June, additional daytime driving has been conducted to gather data during peak traffic and pedestrian times.
The reason for collecting data in Gangnam is due to its complex road environment. Frequent unpredictable situations provide an advantageous setting for the autonomous driving AI to learn from rare edge cases.
Based on the data collected, Kakao Mobility is pushing for the introduction of the E2E model by year-end. Unlike the existing method, which divides the process into perception, judgment, and control stages, the E2E model will handle everything from sensor input to vehicle control with a single AI model.
Im explained, "This will reduce errors that occur in intermediate stages and allow for a comprehensive assessment of the entire driving situation."
However, Kakao Mobility plans not to rely solely on AI for vehicle control. Instead of directly applying the E2E model's decisions to vehicle control, a separate rule-based 'safety evaluator' will be implemented. This system will revalidate the driving path chosen by the E2E model based on traffic laws, vehicle physical conditions, and safety standards, ensuring that rules such as speed limits in school zones and stop signs at intersections take precedence over AI decisions.
Kakao Mobility emphasizes that the key to E2E performance lies in how efficiently 'difficult data' is selected rather than the sheer volume of data.
When the vehicle collects driving data and uploads it to the server, edge cases with high learning value, such as illegal U-turns, unexpected construction, and merging, are identified. AI-based automatic labeling processes the surrounding object information, such as vehicles and pedestrians, into a learnable format. Subsequently, data mining using vision-language models (VLM) analyzes and classifies driving situations based on factors like weather, time of day, road structure, and vehicle movements, which are then reintegrated into AI model training.
AI has also been applied to the data selection process, which previously required human review of driving footage. The AI analyzes uploaded driving data within minutes to identify necessary situations for learning, which are then utilized for model training.
A platform and monitoring system are also being developed to connect driving technology to actual unmanned autonomous driving services. Based on Kakao T, the system will manage not only the calling and dispatching of autonomous vehicles but also the operational zones and routes suitable for each vehicle's capabilities. Through 24-hour monitoring, real-time conditions of vehicles, passengers, and road situations will be monitored. In the absence of a driver, the platform and monitoring system will take over roles previously handled by human drivers, such as managing pick-up and drop-off points and changing destinations.
VLM-based monitoring and remote support features are currently under development and testing. The remote support will not involve operators directly controlling the vehicle but rather providing necessary information to the vehicle in unexpected situations like construction or road closures, with the vehicle's AI making the final driving decisions and controls.
Kakao Mobility plans to gradually expand the number of vehicles, operating hours, and regions for its autonomous driving service in Gangnam after thoroughly verifying safety.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.