[Proposal-ML] Teleoperation-free Demonstration Learning for Humanoid Loco-manipulation

Boyuan Zheng, Junming Zhao, Kexian Tang

Keywords: humanoid, loco-manipulation, reinforcement learning, imitation learning

Abstract: This project will explore a $\textbf{teleoperation-free}$ learning framework for $\textbf{humanoid loco-manipulation}$. Existing approaches often rely on teleoperation or motion-capture demonstrations, which are costly and difficult to generalize. We plan to adopt a hierarchical control architecture that incorporates methods such as $\textbf{diffusion models and reinforcement learning (RL)}$, which we have studied in class. The high-level module will use diffusion policy to generate task-level goals from UMI-style human demonstrations, while the low-level module will employ a teacher–student structure to learn deployable control policies.

Submission Number: 44