Sovereign 8s.is mirror of standardgalactic/skill-abstraction
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| assets | ||
| README.md | ||
STAR: Learning Diverse Robot Skill Abstractions through Rotation-Augmented Vector Quantization
Hao Li1,2 · Qi Lv1,2 · Rui Shao1† · Xiang Deng1† · Yinchuan Li2 · Jianye Hao2 · Liqiang Nie1
1Harbin Institute of Technology (Shenzhen) 2Huawei Noah's Ark Lab
†corresponding author
🔥 Updates
- [06/2025] Arxiv paper released.
- [05/2025] STAR has been accepted by ICML 2025 Spotlight!
🔥 Introduction
This is the github repository of STAR: Learning Diverse Robot Skill Abstractions through Rotation-Augmented Vector Quantization. In this work, we presents STAR, a framework that advances both skill learning and composition to complete complex behaviors.
The whole framework of STAR:
🔥 Performance
Here we present the performance comparison between our method and baseline models on the LIBERO benchmark suite.
| Model | LIBERO-Object | LIBERO-Spatial | LIBERO-Goal | LIBERO-Long | LIBERO-90 | Avg. |
|---|---|---|---|---|---|---|
| VLAs | ||||||
| Octo† | 85.7 ±0.9 | 78.9 ±1.0 | 84.6 ±0.9 | 51.1 ±1.3 | - | 75.1 ±0.6 |
| OpenVLA† | 88.4 ±0.8 | 84.7 ±0.9 | 79.2 ±1.0 | 53.7 ±1.3 | - | 76.5 ±0.6 |
| Baselines | ||||||
| ResNet-T | 78.9 ±1.4 | 75.7 ±1.9 | 52.7 ±2.4 | 45.0 ±1.1 | 83.9 ±1.5 | 67.3 ±0.9 |
| Diffusion Policy | 62.6 ±2.8 | 69.5 ±1.8 | 54.6 ±0.5 | 51.2 ±3.0 | 75.3 ±0.7 | 62.6 ±0.6 |
| ACT | 78.8 ±1.2 | 82.0 ±0.5 | 66.1 ±1.6 | 44.0 ±0.5 | 63.4 ±5.8 | 66.8 ±1.1 |
| Skill-based Methods | ||||||
| VQ-BeT | 90.3 ±1.5 | 88.7 ±2.0 | 61.3 ±1.0 | 59.7 ±0.2 | 84.2 ±0.3 | 76.8 ±0.5 |
| QueST | 90.0 ±1.1 | 84.5 ±0.2 | 76.7 ±0.9 | 69.1 ±1.0 | 87.4 ±0.4 | 81.5 ±0.6 |
| Ours | 98.3 ±0.2 | 95.5 ±0.6 | 95.0 ±0.7 | 88.5 ±0.3 | 90.8 ±0.2 | 93.6 ±0.1 |
🔥 Details will be released. Stay tuned.
🔥 Citation
If you find this work useful for your research, please kindly cite our paper:
@article{li2025star,
title={STAR: Learning Diverse Robot Skill Abstractions through Rotation-Augmented Vector Quantization},
author={Li, Hao and Lv, Qi and Shao, Rui and Deng, Xiang and Li, Yinchuan and Hao, Jianye and Nie, Liqiang},
journal={arXiv preprint arXiv:2506.03863},
year={2025}
}