"Starting autonomous driving."
On a weekday evening near Maebong Station in Gangnam, Seoul, a self-driving taxi named 'Seoul Autonomous Vehicle' began its journey. Although an autonomous driving manager was seated in the driver's seat, their hands were off the wheel. The steering wheel moved on its own, and the vehicle adjusted its speed according to traffic signals while maintaining distance from the car ahead.
The self-driving taxi, based on the Kia EV6, featured sensors and cameras throughout its exterior to monitor the surrounding roads. Kakao Mobility developed a modular sensor platform called 'AV-Kit,' which includes three LiDAR units, five radar units, and seven cameras to recognize nearby vehicles, pedestrians, and road conditions.
From the back seat, passengers could view the road through an 'AVV (Autonomous Vehicle Visualizer)' display, which provided real-time 3D graphics of the vehicle's surroundings, including other cars and pedestrians.
The test drive of the Seoul Autonomous Vehicle lasted about 20 to 30 minutes around the Maebong Station area, navigating various urban environments, including turns, child safety zones, tunnels, and merging traffic.
Passing Through Child Safety Zones and GPS Shadow Tunnels
During the drive, an announcement indicated, "You are entering a child safety zone." Simultaneously, the AVV display confirmed the vehicle's entry into the designated area.
Since August, Kakao Mobility has expanded the operational range of the Seoul Autonomous Vehicle to include child safety zones and other areas designated for vulnerable road users. In these zones, the vehicle is programmed to slow down to 80% of the speed limit and limit sudden acceleration. For example, if the speed limit is 30 km/h, the vehicle will operate at a maximum speed of 24 km/h.
A Kakao Mobility representative stated, "Given the numerous restrictions in vulnerable road user zones, safety is our top priority."
As the test drive progressed, the vehicle entered the Maebong Tunnel, an area where GPS signals are often unreliable. In such environments, the role of sensors becomes crucial for navigation and awareness. This segment allowed for testing the vehicle's perception and driving capabilities in conditions with limited GPS signals.
While entering the tunnel, the vehicle adjusted its speed and position in response to merging traffic from adjacent lanes. Despite the presence of multiple vehicles, the autonomous driving manager did not need to take control of the steering wheel, and the self-driving continued through the tunnel.
Gangnam as a Learning Ground for 'Driving AI'
Upon reaching a bus stop near Maebong Station, the vehicle encountered a car with its hazard lights on, creating uncertainty about whether it would remain stationary or suddenly move. Further along, vehicles briefly stopped in the road for passenger pick-ups and drop-offs. Illegal parking and jaywalking pedestrians are common variables that self-driving cars frequently encounter on Gangnam roads.
Kakao Mobility's goal in gathering autonomous driving data in Gangnam is to experience and learn from as many 'edge cases' as possible—situations that are infrequent or difficult to predict. The company believes that experiencing diverse scenarios in Gangnam for a month is more beneficial for enhancing autonomous driving models than driving in less complex areas for a year.
Since March, Kakao Mobility has been operating a late-night autonomous driving service in Gangnam, navigating both main and side roads. Since June, the company has expanded its data collection to daytime hours when there are more vehicles and pedestrians.
Expanding E2E with Gangnam Data... Broadening AI Decision-Making
The various situations encountered on the roads day and night are utilized as training data for the autonomous driving AI. Kakao Mobility has established a data pipeline that collects driving data, refines it, auto-labels it, and conducts data mining for model retraining. During this process, high-value edge cases are selected for further AI training.
Currently, the Seoul Autonomous Vehicle employs a hybrid approach that combines an AI planner with a rule-based planner. The AI, trained on real-world data, assesses driving situations while also adhering to established safety rules.
Kakao Mobility plans to increase the share of the end-to-end (E2E) model in perception and decision-making by the end of the year. In this model, the AI will recognize and assess driving situations to formulate a driving plan, which will then be validated by a rule-based 'safety evaluator' in the final stage.
A Kakao Mobility representative noted, "While the AI planner currently makes some decisions, it does not yet handle everything. We plan to increase the E2E share to expand the areas where AI can make decisions."
Kakao Mobility is also preparing for a Level 4 autonomous driving service that requires no driver intervention. Currently, if a passenger changes their destination during the ride, the autonomous driving is halted, and the manager must intervene. In the future, the company aims to enhance technology to manage destination changes, pick-up and drop-off confirmations, and unexpected situations through the vehicle and control systems, reducing the need for human involvement.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.