Reproduce “BooST: Bridging Semantics and Motions for Efficient Skill Transfer” 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
Jusuk Lee, Daesol Cho, Jonghun Shin, Seungyeon Yoo, Jonghae Park, Taekbeom Lee, H. Jin Kim
Researchers at Seoul National University and Georgia Institute of Technology developed BooST, a framework that learns unified skill representations by explicitly bridging semantic intent and motion dynamics. This approach enables efficient few-shot skill transfer to real robots, demonstrating up to a 140% relative improvement over baselines in low-data simulation settings and successful cross-embodiment transfer with limited real-world demonstrations.
Key insights
- A unified skill representation that explicitly bridges high-level semantic intent and low-level motion dynamics results in stronger generalization and robustness.
- Supervising skill learning primarily through action reconstruction, rather than pixel-level reconstruction, compels the model to ignore task-irrelevant visual information, enhancing robustness to distractors.
- Decoupling large-scale skill pretraining from lightweight downstream policy distillation enables both broad applicability and efficient real-world deployment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.