About
I am an undergraduate student in Automation at Tsinghua University. I work on embodied AI and robot learning, multi-agent coordination and inter-robot communication, reinforcement learning, and diffusion-based generative models.
I am currently a research intern with Prof. Wei-Chiu Ma at Cornell University, asking when two robots need to share what they see — and what they should send. I also work with Prof. Ge Liu at UIUC on the theory of diffusion models that generate continuous and discrete quantities at once.
News
- Jun 2026Joined Prof. Wei-Chiu Ma's group at Cornell as a research intern, working on shared perception between humanoids.
- Mar 2026Started working with Prof. Ge Liu at UIUC on noise schedules in joint continuous–discrete diffusion.
- Jan 2026Wrapped up PlugRL, a distributed RL framework for vision–language–action policies, at Tsinghua IIIS.
Manuscripts
-
Humanoid2Humanoid: Shared Sensing for Cooperative Loco-Manipulation.
Under review, CVPR 2027. -
Relative Noise Schedules in Joint Continuous–Discrete Diffusion.
In preparation; target ICML 2027.
Research Experience
Cornell University Jun 2026 – Nov 2026
Research Intern, Dept. of Computer Science · Advisor: Prof. Wei-Chiu Ma- Studying how two humanoids can share what they see in real time, so that a robot walking backward climbs stairs hidden from its own cameras by relying on its partner's view.
- Established when such sharing matters: robots usually infer hidden terrain from gait and contact, and only in specific conditions does the partner's view become indispensable, recovering 20–40 success points.
- Built decentralized policies that act on their own sensing plus messages from the partner, halt safely when the link drops and resume when it returns; demonstrated a simulated pair climbing stairs together.
University of Illinois Urbana-Champaign Mar 2026 – Present
Research Intern, Dept. of Computer Science · Advisor: Prof. Ge Liu- Studying how a diffusion model that generates continuous and discrete quantities at once — atom positions with atom types and bonds — should pace noise across its two channels, a choice now hand-tuned per dataset.
- Characterized when the likelihood bound is indifferent to that pacing and what the best schedule looks like when it is not, and confirmed the predictions on public molecular models.
- Showed that reported likelihoods are dominated by a single coarse step of the sampler, and built an estimator reaching the same precision with roughly 2,600× less computation.
Tsinghua University, IIIS Jul 2025 – Jan 2026
Research Assistant · Advisor: Prof. Huazhe Xu- Architected PlugRL, a distributed framework for reinforcement learning on vision–language–action policies.
Tsinghua University, Dept. of Automation Sep 2024 – Jun 2025
Research Assistant · Advisor: Prof. Yebin Liu- Developed a diffusion model that generates physically plausible hand–object interaction for dexterous manipulation, with the simulation environments and demonstration pipelines it was trained on.
Education
Tsinghua University Sep 2023 – Expected Jun 2027
B.Eng. in Automation (Xinya College & Dept. of Automation)- GPA: 3.8/4.0
- Tsinghua Comprehensive Scholarship (2024, 2025); Luo Yuehua Scholarship (2024)
Technical Skills
| Programming | Python (PyTorch), C++, CUDA, MATLAB, Shell |
| Robotics & simulation | IsaacGym, Isaac Lab, MuJoCo, Legged Gym, RSL-RL |
| Infrastructure | Linux, Slurm clusters, distributed multi-GPU training, Weights & Biases |