About Me
I am currently a fourth-year undergraduate student in Institute for Interdisciplinary Information Sciences (IIIS, Yao Class) at Tsinghua University, majoring in Artificial Intelligence. Prior to this, I was a member of the China National Physics Olympiad Training Team.
I am fortunate to be working with Prof. Yuxiao Dong. Previously, I had the privilege of working with Prof. Mengdi Xu and Prof. Li Yi. My research focuses on LLM-based embodied agents for Physical AI.
Outside of research, I enjoy 🏊♂️ swimming, ⚾ softball, 🏸 badminton, 🎤 singing, 📖 reading, and 🌍 traveling. I truly value meaningful conversations and collaborations—if you share similar interests or would like to connect on research or beyond, I would be delighted to hear from you!✨
Research Interests
My research focuses on LLM-based embodied agents for Physical AI. I am interested in how agents can act in the physical world, construct environments and tasks for learning, and continually improve through interaction.
- LLMs as Embodied Policies: Enabling large language models to make grounded decisions and generate executable actions and policies through closed-loop interaction with their environments.
- Agentic Environment and Task Generation: Building agents that autonomously construct, validate, and refine interactive simulation environments and tasks for embodied learning and evaluation.
- Self-Improving Embodied Agents: Developing agents that learn from interaction and execution feedback to improve their policies, accumulate reusable skills, and generalize to new tasks.
Publications
-
CoRL 2026 Workshop on Continually Self-Improving Robots (CSIR)
-
Chenyu Zhang*,
Yuhang Cao*, Daru Du,
Yingxi Lu, Jing Shao,
Jiajun Liu,
Ruoqu Chen,
Liu Cao,
Yicheng Liu,
Hang Zhao,
Mengdi Xu†
preprint
-
Jiajun Liu, Jieming Li, Zi Zhuang, Hang Yu, Qingli Chen,
Liu Cao,
Yingxi Lu,
Ruoqu Chen,
Yuhang Cao,
Chenyu Zhang,
Yankai Lin,
Mengdi Xu†
CVPR 2026 Workshop on 3D-LLM/VLA
Projects
-
Surveyed the development, variants, and practical performance of approximation algorithms for the Facility Location Problem, highlighting theoretical advances and implementation insights.
-
Proposed STAMP—a dual approach combining targeted prompt engineering with a pluggable memory-augmented adapter—for controllable translation from modern to Classical Chinese.
-
Investigated photorealistic image synthesis through a from-scratch C++ path tracing renderer, analyzing acceleration structures, sampling strategies, and material modeling to evaluate efficiency and quality.
-
It's a platform that allows operators, sellers, and regulators to manage and interact with goods and services in an online marketplace.
Language: Scala, TypeScript, HTML, CSS
-
Introduced MagicDance, a fascinating pipeline that can automatically produce personalized dance videos from an arbitrary music clip and a single reference image of a dancer.
PDF
Team awarded second place!
-
We solved the art-style-classification problem by a two-stage architecture where a DNN extracts features from image patches and an SNN adapter makes the final classification.
Selected Awards
2025
2024
2023
Experience
-
2026.10 - present
Intern
Working on LLM-based embodied agents for Physical AI, with a focus on embodied policies, agentic environment and task generation, and self-improvement through interaction.
-
2026.10 - present
Intern
Working on LLM-based embodied agents for Physical AI, with a focus on embodied policies, agentic environment and task generation, and self-improvement through interaction.
-
2026.02 - 2026.09
Worked with
Prof. Mengdi Xu (Tsinghua University)
Studied action tokenization for vision-language-action models and failure-aware progress estimation for robot learning.
-
2025.6 - 2026.1
Advised by
Prof. Li Yi (Tsinghua University)
Developing a perception-policy disentanglement framework for sim2real post-training to enhance policy performance while keeping generalizability. Extensive experience in both simulation and real-world robot experiments.
-
2025.2 - 2025.8
Algorithm Intern
Enhanced the physics-reasoning capabilities of multi-modal large language models through reinforcement learning.
Powered by Jekyll and Minimal Light theme.
Template source from Yitang.
Last updated: 2026-11-19 04:46:29 +0000