A Unified Representation of Different Dynamics Using Deep Koopman Operator

In recent years, dynamic modeling based on Koopman operator theory has emerged as a significant area of focus. However, discrepancies in dynamic models arise under varying conditions for the same system, as well as between different systems, rendering the applicability of a single dynamic model challenging. In contrast, a unified dynamic model enables control across diverse systems and lays the foundation for transfer learning in future tasks. In this work, we propose an end-to-end deep learning framework designed to learn dynamic models for multiple robotic systems, including those with different control dimensions. Experimental results demonstrate that our approach effectively establishes a unified dynamic representation for complex, high-dimensional fully-actuated systems. Additionally, using this unified model, we successfully performed basic trajectory tracking tasks in torque control mode across all tested robotic systems. Subsequent research could incorporate relevant constraints into this framework to facilitate transfer learning.