Reproduce “MOVE: A Simple Motion-Based Data Collection Paradigm for Spatial Generalization in Robotic 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
Huanqian Wang, Chi Bene Chen, Yang Yue, Danhua Tao, Tong Guo, Shaoxuan Xie, Denghang Huang, Shiji Song and 2 more
Researchers from Tsinghua University and BAAI introduce MOVE, a motion-based data collection paradigm that enhances spatial generalization in robotic manipulation by injecting dynamic variability into demonstrations. The method achieves superior data efficiency, outperforming static collection by 2-5x in simulation and matching performance with less than half the data in real-world pick-and-place tasks.
Key insights
- Injecting continuous motion into objects and cameras during data collection transforms demonstrations into rich spatial segments, overcoming static data's spatial sparsity.
- This dynamic data acquisition method, MOVE, enables policies to generalize robustly to unseen spatial configurations with improved data efficiency.
- Combining dynamic dimensions through a phased augmentation strategy is crucial for enriching spatial information across multiple variables within a single trajectory.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.