Go to ICLR 2025 Conference homepageHuWo：Building Physical Interaction World Models for Humanoid Robot Locomotion
Reinforcement Learning control has been proved to be an effective approach for quadruped robot locomotion. However, locomotion tasks for humanoid robots are challenging, especially in complex environments. The main reason is that humanoid robots must maintain balance during movement and constantly engage in complex dynamic interactions with the environment. Understanding robot-environment interaction dynamics is key to achieving stable locomotion for humanoid robots. Since there is privileged information that the robot cannot directly access, to expand the observable space, previous reinforcement learning-based methods either reconstruct environmental information from partial observations or reconstruct robotic dynamics information from partial observations, but they fall short of fully capturing the dynamics of robot-environment interactions. In this work, we propose an end-to-end reinforcement learning control framework based on physical interaction$\textbf{Wo}$rld Model for $\textbf{Hu}$manoid Robots (HuWo). Our key innovation is to introduce a physical interaction world model to understand the interaction between the robot and environment, employing the hidden layers of transformer-XL for implicit modeling of this process across temporal sequences. The proposed framework can showcase robust and flexible locomotion ability in complex environments such as slopes, stairs, and discontinuous surfaces. We validated the robustness of this method using the $Zerith1$ robot, both in simulations and real-world deployments, and quantitatively compared our HuWo against the baselines with better traversability and command-tracking.