Embodied AI · Robot Learning · Vision-Language-Action

Yushun Xiang 项裕顺

Master's student at RhOS lab, Shanghai Jiao Tong University, advised by Prof. Yong-Lu Li. I work on making robots that see, understand, and act, with a focus on efficient vision-language-action models and benchmarks for embodied agents.

Portrait of Yushun Xiang

News

  • RoboICL is out on arXiv: embodied in-context learning with GPT-6 Astra. Code and project page are available.
  • SpecPrune-VLA was accepted to ICML 2026.
  • We released The Great March 100 (GM-100), a benchmark of 100 detail-oriented tasks for embodied AI. See the project page.
  • Started my master’s at RhOS, Shanghai Jiao Tong University.

Publications

* equal contribution

  1. Preprint 2026

    RoboICL: Embodied In-Context Learning with GPT-6 Astra

    Fangcheng Liu*, Yeqing Shen*, Anda Cheng*, Weishi Mi, Chao Tang, Chenyuan Liu, Yushun Xiang, Tingguang Li, Yong-Lu Li, Yehui Tang

    arXiv preprint, 2026

    Drives a frozen GPT-6 Astra as a few-shot robot learner using demonstration context and a bounded, anchored interaction memory, with no robot-specific training.

  2. Preprint 2026

    The Great March 100: 100 Detail-oriented Tasks for Evaluating Embodied AI Agents

    Ziyu Wang, Chenyuan Liu, Yushun Xiang, Runhao Zhang, Qingbo Hao, Hongliang Lu, Houyu Chen, Zhizhong Feng, Kaiyue Zheng, Dehao Ye, Xianchao Zeng, Xinyu Zhou, Boran Wen, Jiaxin Li, Mingyu Zhang, Kecheng Zheng, Qian Zhu, Ran Cheng, Yong-Lu Li

    arXiv preprint, 2026

    GM-100: a benchmark of 100 long-tail manipulation tasks built from human-object interaction primitives and object affordances, with large-scale teleoperation data, designed to separate today's VLA models.

  3. ICML 2026

    SpecPrune-VLA: Accelerating Vision-Language-Action Models via Action-Aware Self-Speculative Pruning

    Hanzhen Wang, Jiaming Xu, Yushun Xiang, Jiayi Pan, Yongkang Zhou, Yong-Lu Li, Guohao Dai

    International Conference on Machine Learning (ICML), 2026

    A training-free, two-level token pruning scheme for VLA policies: action-level static pruning plus layer-level dynamic pruning, steered by an action-aware controller, for faster inference at near-identical success rates.

About

I am a master’s student at the RhOS lab, Shanghai Jiao Tong University, under the supervision of Prof. Yong-Lu Li. I earned my BEng degree from the School of Software at Northwestern Polytechnical University.

Honors & Awards

  • Meritorious Winner, Mathematical Contest in Modeling
  • Chinese National Scholarship
  • Excellent Intern, a summer internship program organized by ASLP@NPU