The Ulsan National Institute of Science and Technology (UNIST) announced on the 21st that Professor Han Seung-yeol's team at its Artificial Intelligence (AI) Graduate School has achieved the remarkable feat of having three papers simultaneously accepted at the International Conference on Learning Representations (ICLR 2026). ICLR is considered one of the "world's top three AI conferences," along with the Conference on Neural Information Processing Systems (NeurIPS) and the International Conference on Machine Learning (ICML).
Researchers Lee Sang-hyun, Hwang Jae-bak, and Cho Yong-hyun participated as first authors of the three accepted papers respectively, and all produced research achievements in the field of reinforcement learning, which is the core of next-generation physical AI technology. Reinforcement learning is a learning method in which AI interacts with its environment and finds optimal actions on its own through trial and error. Examples of reinforcement learning include enabling robots or autonomous vehicles to directly confront uncertain and unpredictable real-world physical environments, perceive situations, and cope with unexpected variables.
Professor Han's team presented three technologies as their respective research results: the "Self-Improving Skill Learning (SISL)" technique, which can effectively train AI even with offline data collected directly from industrial sites; the "Strict Subgoal Execution (SSE)" learning technique, which can increase the success rate of robots performing complex tasks; and the "Sequential Subvalue Q-learning (S2Q)" technique, which addresses optimization problems in multi-agent reinforcement learning (MARL) environments where multiple AI agents cooperate.
Professor Han stated, "This research demonstrates the possibility of stably applying reinforcement learning even in environments with limited data and uncertainty," adding, "We expect its expansion into various fields such as autonomous driving, robotics, and smart manufacturing."
