Reproduce “STAR: Learning Diverse Robot Skill Abstractions through Rotation-Augmented Vector Quantization” 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
Hao Li, Qi Lv, Rui Shao, Xiang Deng, Yinchuan Li, Jianye Hao, Liqiang Nie
Researchers from Harbin Institute of Technology (Shenzhen) and Huawei Noah’s Ark Lab developed STAR, a framework using rotation-augmented vector quantization and a causal transformer to learn diverse robot skill abstractions and their causal dependencies. This approach achieved a 93.6% success rate on the LIBERO benchmark, an improvement of 12.1% over previous methods, and demonstrated complete codebook utilization.
Key insights
- Encoding geometric relationships via rotation-augmented gradients allows points within the same skill code region to receive differential updates, effectively preventing codebook collapse and fostering diverse skill abstractions.
- An autoregressive mechanism, coupled with hierarchical skill prediction, effectively models the temporal and causal dependencies between skills, enabling coherent action generation for complex sequential tasks.
- Combining discrete skill predictions with a continuous action refinement mechanism is crucial for bridging the gap between abstract skill representations and precise, continuous robot control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.