MotionPyramid: Hierarchical Motion Representation and Residual Interfaces

MotionPyramid introduces a hierarchical representation for motion, enabling the organization of actions from low-level motor commands to higher-level behavioral structures. It employs a recursive stack of latent decoders to create a multi-level action interface, which is frozen post-pretraining for use in downstream reinforcement learning policies. This approach enhances early learning and motion regularity while allowing for precise control, demonstrating that motion can be structured similarly to perception, thus providing practitioners with a reusable framework that balances abstraction and controllability in humanoid control tasks.