Xiangju Mi 米相举

Xiangju Mi standing by the sea

Hi there! I am a third-year undergraduate student at Peking University. And I am fortunate to be advised by Prof. Shanghang Zhang at the HMI Lab.

Currently, my research interests center on humanoid robotics, dexterous manipulation, and agentic learning.

Beyond research, I enjoy exploring physics and playing basketball.

Research Interests

I am interested in building general-purpose embodied agents that learn and act reliably in the physical world.

A central question for me is how robots can acquire reusable skills from diverse experience, compose them for new tasks, and improve through interaction. I hope to develop policies that generalize across tasks and environments, adapt to unfamiliar situations, and recover from failure.

News

RoboBench was accepted to ECCV 2026.
Released LaST-HD, aligning human-hand and robot demonstrations through world-model latents for efficient cross-embodiment learning.
LaST0 was accepted to ICML 2026 as a Spotlight (top 2.2%).
Released LaST0, a token-efficient framework for latent physical reasoning before action generation.
Released RoboMIND 2.0, a large-scale multimodal dataset for bimanual and mobile manipulation.

Selected Publications

  1. ECCV 2026 Equal Contribution

    RoboBench: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models as Embodied Brain

    Yulin Luo*, Chun-Kai Fan*, Menghang Dong*, Jiayu Shi*, Xiangju Mi*, Mengdi Zhao*, Bo-Wen Zhang*, Cheng Chi*, Jiaming Liu, Gaole Dai, Rongyu Zhang, Ruichuan An, Kun Wu, Zhengping Che, Shaoxuan Xie, Guocai Yao, Zhongxia Zhao, Pengwei Wang, Guang Liu, Zhongyuan Wang, Tiejun Huang, Shanghang Zhang
    Project Paper Code Dataset

    A real-world benchmark for evaluating multimodal large language models across five cognitive dimensions, 14 capabilities, 25 tasks, and 6,092 question–answer pairs.

    Overview of the RoboBench embodied-brain evaluation benchmark
  2. arXiv 2026 Equal Contribution

    LaST-HD: Learning Latent Physical Reasoning from Scalable Human Data for Robot Manipulation

    Jiaming Liu*, Yinxi Wang*, Chenyang Gu*, Siyuan Qian*, Xiangju Mi*, Hao Chen, Jiawei Chen, Qingpo Wuwu, Xiaoqi Li, Nuowei Han, Yiming Zhang, Xuheng Zhang, Yang Yue, Yeqing Yang, Lei Wang, Peng Jia, Hao Tang, Shanghang Zhang
    Project Paper

    Aligns human-hand and robot demonstrations through world-model latents, enabling efficient cross-embodiment learning from scalable human data.

    Overview of LaST-HD latent alignment and human-to-robot learning

* Equal contribution.

Education

Peking University

– Present

B.S. in Theoretical and Applied Mechanics (Minor in Robotics Engineering)

Research Experience

Awards