STAR: Learning Diverse Robot Skill Abstractions through Rotation-Augmented Vector Quantization

Hao Li, Qi Lv, Rui Shao, Xiang Deng, Yinchuan Li, Jianye Hao, Liqiang Nie
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:35562-35580, 2025.

Abstract

Transforming complex actions into discrete skill abstractions has demonstrated strong potential for robotic manipulation.Existing approaches mainly leverage latent variable models, e.g., VQ-VAE, to learn skill abstractions through learned vectors (codebooks), while they suffer from codebook collapse and modeling the causal relationship between learned skills. To address these limitations, we present Skill Training with Augmented Rotation (STAR), a framework that advances both skill learning and composition to complete complex behaviors. Specifically, to prevent codebook collapse, we devise rotation-augmented residual skill quantization (RaRSQ).It encodes relative angles between encoder outputs into the gradient flow by rotation-based gradient mechanism. Points within the same skill code are forced to be either pushed apart or pulled closer together depending on gradient directions.Further, to capture the casual relationship between skills, we present causal skill transformer (CST) which explicitly models dependencies between skill representations through an autoregressive mechanism for coherent action generation.Extensive experiments demonstrate the superiority of STAR on both LIBERO benchmark and realworld tasks, with around 12% improvement over the baselines.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-li25bo, 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}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {35562--35580}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/li25bo/li25bo.pdf}, url = {https://proceedings.mlr.press/v267/li25bo.html}, abstract = {Transforming complex actions into discrete skill abstractions has demonstrated strong potential for robotic manipulation.Existing approaches mainly leverage latent variable models, e.g., VQ-VAE, to learn skill abstractions through learned vectors (codebooks), while they suffer from codebook collapse and modeling the causal relationship between learned skills. To address these limitations, we present Skill Training with Augmented Rotation (STAR), a framework that advances both skill learning and composition to complete complex behaviors. Specifically, to prevent codebook collapse, we devise rotation-augmented residual skill quantization (RaRSQ).It encodes relative angles between encoder outputs into the gradient flow by rotation-based gradient mechanism. Points within the same skill code are forced to be either pushed apart or pulled closer together depending on gradient directions.Further, to capture the casual relationship between skills, we present causal skill transformer (CST) which explicitly models dependencies between skill representations through an autoregressive mechanism for coherent action generation.Extensive experiments demonstrate the superiority of STAR on both LIBERO benchmark and realworld tasks, with around 12% improvement over the baselines.} }
Endnote
%0 Conference Paper %T STAR: Learning Diverse Robot Skill Abstractions through Rotation-Augmented Vector Quantization %A Hao Li %A Qi Lv %A Rui Shao %A Xiang Deng %A Yinchuan Li %A Jianye Hao %A Liqiang Nie %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-li25bo %I PMLR %P 35562--35580 %U https://proceedings.mlr.press/v267/li25bo.html %V 267 %X Transforming complex actions into discrete skill abstractions has demonstrated strong potential for robotic manipulation.Existing approaches mainly leverage latent variable models, e.g., VQ-VAE, to learn skill abstractions through learned vectors (codebooks), while they suffer from codebook collapse and modeling the causal relationship between learned skills. To address these limitations, we present Skill Training with Augmented Rotation (STAR), a framework that advances both skill learning and composition to complete complex behaviors. Specifically, to prevent codebook collapse, we devise rotation-augmented residual skill quantization (RaRSQ).It encodes relative angles between encoder outputs into the gradient flow by rotation-based gradient mechanism. Points within the same skill code are forced to be either pushed apart or pulled closer together depending on gradient directions.Further, to capture the casual relationship between skills, we present causal skill transformer (CST) which explicitly models dependencies between skill representations through an autoregressive mechanism for coherent action generation.Extensive experiments demonstrate the superiority of STAR on both LIBERO benchmark and realworld tasks, with around 12% improvement over the baselines.
APA
Li, H., Lv, Q., Shao, R., Deng, X., Li, Y., Hao, J. & Nie, L.. (2025). STAR: Learning Diverse Robot Skill Abstractions through Rotation-Augmented Vector Quantization. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:35562-35580 Available from https://proceedings.mlr.press/v267/li25bo.html.

Related Material