Existing modular autonomous driving systems are designed by separating functions such as perception, prediction, planning, and control into individual modules. In this structure, the interfaces and rules between modules must be manually defined by humans, which limits the system’s ability to flexibly handle complex and unpredictable situations in real-world driving environments.
End-to-End (E2E) autonomous driving is a next-generation approach that directly learns the entire process from sensor inputs to vehicle control commands using a single deep neural network. By learning complex relationships between sensory observations and driving actions from large-scale driving data, the system can better handle diverse driving situations.
Our laboratory conducts research on End-to-End autonomous driving based on imitation learning and reinforcement learning. In addition, we investigate multimodal models that utilize diverse sensor inputs, as well as simulation-based training and Sim-to-Real transfer techniques, with the goal of developing autonomous driving systems that operate reliably in real-world environments.
Research Competitiveness
We conduct learning-based End-to-End autonomous driving research that combines imitation learning and reinforcement learning, enabling the system to learn human driving patterns while also improving driving policies through interaction with the environment.
We also develop multimodal End-to-End autonomous driving models that integrate diverse input sources such as camera data and map information, enhancing both situational understanding and decision-making capabilities of autonomous vehicles.
In addition, we perform research on large-scale simulation-based training and Sim-to-Real transfer techniques, enabling the development of autonomous driving systems that achieve stable performance in real-world environments while significantly reducing the cost and risks associated with real-world data collection.
Our team participated in the 2025 Hyundai Autonomous Driving Challenge, where we developed an End-to-End autonomous driving model for urban environments and ultimately achieved 2nd place in the competition.
Through these efforts, we aim to establish next-generation End-to-End autonomous driving technologies that integrate data-driven learning, multimodal perception, and simulation-based training, and to advance autonomous driving systems with strong scalability and generalization capabilities that can operate reliably in diverse real-world environments.