Abstract: This talk presents a data-driven approach that leverages real-world motion as a target for imitation, enabling robots to learn natural, adaptive, and diverse behaviors. Moving beyond simple mimicry, our work introduces a novel control framework that learns the underlying patterns within this data. This allows the system to capture multiple distinct behaviors and generate user-steerable and adaptive actions while preserving the stylistic coherence of the source motion. We address three key challenges inherent to this data-driven approach, presenting solutions that leverage a novel integration of model-based control and reinforcement learning. We will detail the core methodologies and present experimental results demonstrating the framework’s effectiveness as a scalable and efficient method for learning complex, natural skills. Finally, we will discuss the potential of this approach as a general control framework for the emerging field of humanoid robotics, aiming to achieve the agile, context-aware locomotion necessary for general-purpose robots.

Speaker Bio: Dongho Kang is a Research Scientist at the Robotics and AI Institute (RAI). His research lies at the intersection of optimal control, reinforcement learning, and data-driven character animation. He received his Ph.D. in Computer Science (2025) and his M.S. in Mechanical Engineering (2019) from ETH Zurich. Previously, he obtained his B.S. from Seoul National University with a double major in Mechanical & Aerospace Engineering and Computer Science & Engineering (2016).