
Real-to-Sim-to-Real at the Last Millimeter
Learning image-based policies in simulation for precise drawer and charger insertion on a low-cost robot arm.
I am a Computer Science Ph.D. student at the University of Chicago, advised by Zhiyuan Li, with research at TTIC.
I am interested in coding agents, reinforcement learning, formal theorem proving, and agentic robotics, including Code as Policies. My work explores how agents can write code, reason with formal guarantees, and act in the physical world.
I am looking for a research internship in 2027. Get in touch ↗
Before my Ph.D., I completed my B.S. in Computer Science, Mathematics, and Economics at the University of Chicago.

Real-to-Sim-to-Real at the Last Millimeter is accepted to CoRL 2026 Late-Breaking Results.
Towards Training Transformers as a Universal Computer is accepted to Findings of EMNLP 2026.
Teaching an Agent to Sketch One Part at a Time at ECCV 2026.
Schur Nets at NeurIPS 2024.

Learning image-based policies in simulation for precise drawer and charger insertion on a low-cost robot arm.

A multi-turn visual agent that draws editable vector sketches one semantic part at a time.
@inproceedings{du2026teaching,
title={Teaching an Agent to Sketch One Part at a Time},
author={Du, Xiaodan and Xu, Ruize and Yunis, David and Vinker, Yael and Shakhnarovich, Greg},
booktitle={European Conference on Computer Vision},
year={2026}
}
Learning to execute unseen programs with a small transformer and PENCIL scaffolding.
Preprint title: Training Transformers as a Universal Computer.
@article{xu2026training,
title={Training Transformers as a Universal Computer},
author={Xu, Ruize and Yang, Chenxiao and Li, Yanhong and McAllester, David},
journal={arXiv preprint arXiv:2604.25166},
year={2026}
}
Constructing equivariant operations from the graph Laplacian for higher-order graph neural networks.
@inproceedings{zhang2024schur,
title={Schur Nets: Exploiting Local Structure for Equivariance in Higher Order Graph Neural Networks},
author={Zhang, Qingqi and Xu, Ruize and Kondor, Risi},
booktitle={Advances in Neural Information Processing Systems},
year={2024}
}Compressing formal mathematical proofs in Lean with kernel-checked correctness and reusable lemmas.
* Equal contribution
University of Chicago · Computer Science
Advisor: Zhiyuan Li · Expected June 2028
University of Chicago
B.S. in Computer Science, Mathematics, and Economics
Google · Software Engineer
Gmail Ads auction optimization and production experimentation, applying Google DeepMind’s FunSearch.
Google · Software Engineer
Query execution and vectorized joins in the F1 query kernel.
For research collaborations and opportunities: richardxu@ttic.edu · UChicago email