GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data

Shengliang Deng, Mi Yan, Songlin Wei, Haixin Ma, Yuxin Yang, Jiayi Chen, Zhiqi Zhang, Taoyu Yang, Xuheng Zhang, Heming Cui, Zhizheng Zhang, He Wang
Proceedings of The 9th Conference on Robot Learning, PMLR 305:1004-1029, 2025.

Abstract

Embodied foundation models are gaining increasing attention for their zero-shot generalization, scalability, and adaptability to new tasks through few-shot post-training. However, existing models rely heavily on real-world data, which is costly and labor-intensive to collect. Synthetic data offers a cost-effective alternative, yet its potential remains largely underexplored. To bridge this gap, we explore the feasibility of training Vision-Language-Action (VLA) models entirely with large-scale synthetic action data. We curate SynGrasp-1B, a billion-frame robotic grasping dataset generated in simulation with photorealistic rendering and extensive domain randomization. Building on this, we present GraspVLA, a VLA model pretrained on large-scale synthetic action data as a foundational model for grasping tasks. GraspVLA integrates autoregressive perception tasks and a flow-matching-based action generation into a unified Chain-of-Thought process, enabling joint training on synthetic action data and Internet semantics data. This design helps mitigate sim-to-real gaps and facilitates the transfer of learned actions to a broader range of Internet-covered objects, achieving open-vocabulary generalization in grasping. Extensive evaluations across real-world and simulation benchmarks demonstrate GraspVLA’s advanced zero-shot generalizability and few-shot adaptability to specific human preferences. We will release SynGrasp-1B dataset and pre-trained weights to benefit the community.

Cite this Paper

BibTeX
@InProceedings{pmlr-v305-deng25a, title = {GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data}, author = {Deng, Shengliang and Yan, Mi and Wei, Songlin and Ma, Haixin and Yang, Yuxin and Chen, Jiayi and Zhang, Zhiqi and Yang, Taoyu and Zhang, Xuheng and Cui, Heming and Zhang, Zhizheng and Wang, He}, booktitle = {Proceedings of The 9th Conference on Robot Learning}, pages = {1004--1029}, year = {2025}, editor = {Lim, Joseph and Song, Shuran and Park, Hae-Won}, volume = {305}, series = {Proceedings of Machine Learning Research}, month = {27--30 Sep}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v305/main/assets/deng25a/deng25a.pdf}, url = {https://proceedings.mlr.press/v305/deng25a.html}, abstract = {Embodied foundation models are gaining increasing attention for their zero-shot generalization, scalability, and adaptability to new tasks through few-shot post-training. However, existing models rely heavily on real-world data, which is costly and labor-intensive to collect. Synthetic data offers a cost-effective alternative, yet its potential remains largely underexplored. To bridge this gap, we explore the feasibility of training Vision-Language-Action (VLA) models entirely with large-scale synthetic action data. We curate SynGrasp-1B, a billion-frame robotic grasping dataset generated in simulation with photorealistic rendering and extensive domain randomization. Building on this, we present GraspVLA, a VLA model pretrained on large-scale synthetic action data as a foundational model for grasping tasks. GraspVLA integrates autoregressive perception tasks and a flow-matching-based action generation into a unified Chain-of-Thought process, enabling joint training on synthetic action data and Internet semantics data. This design helps mitigate sim-to-real gaps and facilitates the transfer of learned actions to a broader range of Internet-covered objects, achieving open-vocabulary generalization in grasping. Extensive evaluations across real-world and simulation benchmarks demonstrate GraspVLA’s advanced zero-shot generalizability and few-shot adaptability to specific human preferences. We will release SynGrasp-1B dataset and pre-trained weights to benefit the community.} }
Endnote
%0 Conference Paper %T GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data %A Shengliang Deng %A Mi Yan %A Songlin Wei %A Haixin Ma %A Yuxin Yang %A Jiayi Chen %A Zhiqi Zhang %A Taoyu Yang %A Xuheng Zhang %A Heming Cui %A Zhizheng Zhang %A He Wang %B Proceedings of The 9th Conference on Robot Learning %C Proceedings of Machine Learning Research %D 2025 %E Joseph Lim %E Shuran Song %E Hae-Won Park %F pmlr-v305-deng25a %I PMLR %P 1004--1029 %U https://proceedings.mlr.press/v305/deng25a.html %V 305 %X Embodied foundation models are gaining increasing attention for their zero-shot generalization, scalability, and adaptability to new tasks through few-shot post-training. However, existing models rely heavily on real-world data, which is costly and labor-intensive to collect. Synthetic data offers a cost-effective alternative, yet its potential remains largely underexplored. To bridge this gap, we explore the feasibility of training Vision-Language-Action (VLA) models entirely with large-scale synthetic action data. We curate SynGrasp-1B, a billion-frame robotic grasping dataset generated in simulation with photorealistic rendering and extensive domain randomization. Building on this, we present GraspVLA, a VLA model pretrained on large-scale synthetic action data as a foundational model for grasping tasks. GraspVLA integrates autoregressive perception tasks and a flow-matching-based action generation into a unified Chain-of-Thought process, enabling joint training on synthetic action data and Internet semantics data. This design helps mitigate sim-to-real gaps and facilitates the transfer of learned actions to a broader range of Internet-covered objects, achieving open-vocabulary generalization in grasping. Extensive evaluations across real-world and simulation benchmarks demonstrate GraspVLA’s advanced zero-shot generalizability and few-shot adaptability to specific human preferences. We will release SynGrasp-1B dataset and pre-trained weights to benefit the community.
APA
Deng, S., Yan, M., Wei, S., Ma, H., Yang, Y., Chen, J., Zhang, Z., Yang, T., Zhang, X., Cui, H., Zhang, Z. & Wang, H.. (2025). GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data. Proceedings of The 9th Conference on Robot Learning, in Proceedings of Machine Learning Research 305:1004-1029 Available from https://proceedings.mlr.press/v305/deng25a.html.

Related Material