We introduce DexVLG, a vision‑language‑grasp model trained on a large-scale synthetic dataset that can generate instruction‑aligned dexterous grasp poses and achieves SOTA success and part‑grasp accuracy.

• DexGraspNet3.0, a large-scale dataset containing 170M part-aligned dexterous grasp poses on 174k objects, each annotated with semantic captions.
• DexVLG, a vision‑language model to generate language-instructed dexterous grasp poses in an end-to-end way.
• We curate benchmarks and conduct extensive experiments to evaluate DexVLG in simulation and the real world.

TODO List:

• Release the training/inference code of DexVLG
• Release model weights

@article{dexvlg25,
      title={DexVLG: Dexterous Vision-Language-Grasp Model at Scale},
      author={He, Jiawei and Li, Danshi and Yu, Xinqiang and Qi, Zekun and Zhang, Wenyao and Chen, Jiayi and Zhang, Zhaoxiang and Zhang, Zhizheng and Yi, Li and Wang, He},
      journal={arXiv preprint arXiv:2507.02747},
      year={2025}
    }