Reproduce “MoMaGen: Generating Demonstrations under Soft and Hard Constraints for Multi-Step Bimanual Mobile Manipulation” 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
Chengshu Li, Mengdi Xu, Arpit Bahety, Hang Yin, Yunfan Jiang, Huang Huang, Josiah Wong, Sujay Garlanka and 6 more
MOMAGEN is a framework developed by researchers at Stanford University, The University of Texas at Austin, and Amazon that automates the generation of diverse, high-quality synthetic demonstrations for multi-step bimanual mobile manipulation. It formulates data generation as a constrained optimization problem, explicitly incorporating reachability and object visibility to produce data that improves imitation learning policy performance and facilitates sim-to-real transfer from a single human demonstration.
Key insights
- Framing data generation as a constrained optimization problem, explicitly modeling hard and soft constraints, provides a robust and generalizable approach for complex robotic tasks, extending prior X-Gen methods.
- Explicitly enforcing reachability and object visibility as constraints during data generation is essential for creating behaviorally valid and perceptually relevant demonstrations for mobile manipulators, overcoming fundamental limitations of prior methods.
- High-quality, diverse synthetic data generated from minimal human input (a single demonstration) can effectively train imitation learning policies, leading to improved performance, enhanced generalizability, and more efficient sim-to-real transfer.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.