Reproduce “DynaMimicGen: A Data Generation Framework for Robot Learning of Dynamic Tasks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vincenzo Pomponi, Paolo Franceschi, Stefano Baraldo, Loris Roveda, Oliver Avram, Luca Maria Gambardella, Anna Valente
DynaMimicGen (D-MG) introduces a data generation framework for robot learning of dynamic tasks, producing large-scale, diverse datasets from as few as one or two human demonstrations. It leverages Dynamic Movement Primitives and continuous object pose monitoring to synthesize adaptable Cartesian trajectories, leading to improved policy success rates up to 97.33% in simulation and 96.0% in real-world for tasks like Lift, while significantly reducing human data collection effort.
Key insights
- Scalable robot data generation from minimal human input (1-2 demonstrations) is achievable by leveraging DMPs to generalize object-centric subtask trajectories.
- Integrating dynamic environmental perturbations during data generation is crucial for creating robust and generalizable policies, especially for image-based observations in evolving scenes.
- Continuous monitoring and adaptation of robot trajectories based on real-time object poses enables effective data synthesis for dynamic tasks, a capability largely absent in prior imitation learning data generation frameworks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.