Motor Primitive Learning for Humanoid Motion

Overview

This project studies how reusable motor primitives can be extracted from motion data and composed across time and body parts to generate diverse humanoid behaviors.

Motivation

Conventional imitation-learning pipelines often require separate datasets and policies for individual tasks. The goal is to learn a compact primitive library that can be reused and recombined rather than expanding the skill library for every new behavior.

Research Directions

• Learning reusable motion units from humanoid trajectories
• Spatial decomposition across body parts
• Temporal activation and composition of primitives
• Integration with reinforcement and imitation learning

Media and Results

Figures, videos, quantitative results, and related publications will be added as the project progresses.