DynaMimicGen: A Data Generation Framework for Robot Learning of Dynamic Tasks

Learning robust manipulation policies typically demands large and diverse datasets, whose collection is time-consuming, labor-intensive, and impractical in dynamic environments. We present DynaMimicGen (D-MG), a scalable data-generation framework that enables policy learning from minimal human supervision while uniquely supporting data synthesis in dynamic task conditions. Starting from only a few human demonstrations - potentially as few as a single example - D-MG segments the demonstrations into sub-tasks and uses Dynamic Movement Primitives (DMPs) to generalize each skill to novel and continually changing scenes. Unlike prior methods that assume static environments or rely on simple interpolation, D-MG synthesizes smooth, physically plausible Cartesian trajectories that remain task-consistent while adapting in real time to variations in object pose, robot configuration, and scene geometry. Our method supports basic, contact-rich, and long-horizon tasks in both static and dynamic settings. DynaMimicGen increases dataset-generation success rates by 3.6% (Stack), 59.5% (Square), 124.7% (HammerCleanup), 232.7% (MugCleanup), and 4.60% (NutAssembly) over MimicGen (which requires ten demonstrations), while remaining robust to perturbations. By injecting dynamic perturbations during data generation, D-MG increases dataset variability, enabling Diffusion Policy and Behavior Cloning agents to achieve strong performance across all tasks and consistently outperform MG-trained policies. By reducing the need for large human demonstration sets and enabling adaptation in dynamic environments, D-MG provides a scalable and effective foundation for autonomous robot learning.