Xiu Yuan

Xiu Yuan

Embodied intelligence with foundation models

I'm a first-year Computer Science PhD student at Washington University in St. Louis, advised by Prof. Chongjie Zhang. I received my B.S. in Computer Science from UC San Diego, where I was advised by Prof. Hao Su, and my high school study at Nanjing Foreign Language School.

In Summer 2026 I interned at Waymo as a Research Intern on the Pre-training team within AI Foundations, mentored by Kratarth Goel (Senior Staff Research Scientist, Tech Lead), Jonah Philion (Senior Research Scientist), and Qi Zhao (Staff Software Engineer). My work focused on improving the reasoning capabilities of driving foundation models.

Portrait of Xiu Yuan

Research

My research focuses on advancing embodied intelligence with foundation models. I organize my work around three core pillars:

  1. Efficient online adaptation and continual improvement of low-level embodied control models, including vision-language-action models (VLAs) and world-action models (WAMs).
  2. Stronger reasoning capabilities for high-level embodied models, particularly vision-language models (VLMs).
  3. Agentic robotic systems that can autonomously plan, adapt, coordinate, and interact with the physical world.

I'm always happy to collaborate on interesting projects. If you'd like to chat, feel free to shoot me an email.

Industry Experience

Education

Publications & Preprints

Sorted by recency. Highlighted entries are representative papers.

RoboAuto: Autonomous Robot Program Search through Multi-Level Physical Understanding

Buzhao Liu*, Jianing Ye*, Xiu Yuan*, Yangtong Feng, Chengsong Huang, Tyler Dinh, Yevgeniy Vorobeychik, Ning Zhang, Chongjie Zhang (* equal contribution)

  • arXiv 2026

Project Page arXiv

Execution feedback tells a robot program whether it succeeded, but rarely why it failed. RoboAuto guides program search with multi-level physical understanding: a hierarchical knowledge base of object-centric physics and reusable skills, grounded in each scene by measuring the quantities a plan actually needs, and refined through self-designed experiments when that understanding is uncertain. Planning proceeds in verified stages, and the knowledge base transfers across simulators and embodiments, improving over ASPIRE on LIBERO-PRO, LIBERO-Plus, Robosuite, and real-robot long-horizon tasks.

Buzhao Liu is an undergraduate student I proudly mentored.

V-JEPA Policy: Building Effective World-Action Models on Predictive Visual Latents

Yang Zhang, Jiangyuan Zhao, Chenyou Fan, Jiayu Hu, Xiu Yuan, Chenjia Bai, Xiu Li

  • Submitted to ICLR 2027

arXiv Code

World-action models typically inherit a complete pretrained video generator along with its predictive knowledge. V-JEPA Policy instead builds one on the frozen latent space of a V-JEPA 2.1 encoder, jointly learning an instruction-conditioned future-latent predictor and a flow-matching action expert from scratch. At 0.9B parameters, it stays competitive with world-action and vision-language-action baselines on LIBERO, LIBERO-Plus, and RoboCasa-GR1, and transfers to real-world bimanual manipulation. Action-free predictor pretraining on DROID videos further improves out-of-distribution generalization.

MEMORA: Embodied Action Memory from Egocentric Videos for Reasoning and Planning

Zihao Yu, Xiu Yuan, Chongjie Zhang

  • EMNLP 2026
  • Foundation Model for Robot Planning Workshop @ RSS 2026 (Oral)

Long-horizon robot planning needs persistent embodied memory beyond next-action prediction. MEMORA realizes Embodied Action Memory through a formation–consolidation–retrieval lifecycle with four typed stores: Environment, Entity, Activity, and Inferred Knowledge. Online editing tracks object identities and state histories, while offline consolidation turns repeated experience into reusable procedures and participant-specific regularities.

Zihao Yu is a master's student I proudly mentored.

SUN-WM UAV navigation route planned over an aerial semantic map
SUN-WM: A Semantic World Model with Learned Map Dynamics for UAV Navigation

Qiwei Wang, Xiu Yuan, Jianing Ye, Chongjie Zhang

  • Submitted to ICRA 2027

Qiwei Wang is a master's student I proudly mentored.

Unified Latent Steering and Residual Refinement for Online Improvement of Diffusion Policy Models

Zhengbang Zhu*, Ziyan Li*, Xiu Yuan*, Hanbo Zhang, Yuxiao Liu, Chongjie Zhang, Yong Yu, Weinan Zhang, Minghuan Liu (* equal contribution)

  • Submitted to ICRA 2027
  • Decision-Making from Offline Datasets to Online Adaptation Workshop @ ICML 2026

Imitation policies are brittle under distribution shift, and direct finetuning is costly. USR enables finetuning-free online improvement of diffusion policies through unified latent Steering and residual Refinement: a lightweight actor steers diffusion toward promising modes while adding corrections beyond the policy's support. USR achieves state-of-the-art results on benchmarks, MultiModalBench, and real-robot VLA adaptation.

Responsive Noise-Relaying Diffusion Policy: Responsive and Efficient Visuomotor Control

Zhuoqun Chen*, Xiu Yuan*, Tongzhou Mu, Hao Su (* equal contribution)

  • TMLR 2025
  • Generative Models for Robot Learning Workshop @ ICLR 2025

Project Page arXiv Code

We propose RNR-DP, a diffusion-based approach for visuomotor control that ensures responsive and efficient action generation. By leveraging a noise-relaying action buffer and a sequential denoising mechanism, RNR-DP demonstrates advantages on both response-sensitive and regular tasks across the evaluated benchmarks.

Policy Decorator: Model-Agnostic Online Refinement for Large Policy Models

Xiu Yuan*, Tongzhou Mu*, Stone Tao, Yunhao Fang, Mengke Zhang, Hao Su (* equal contribution)

  • ICLR 2025

Project Page arXiv Code

Large policy models learned by offline imitation learning are often limited by the quantity, quality, and diversity of demonstrations. We introduce Policy Decorator, which uses a model-agnostic residual policy to refine large policy models during online interactions. By implementing controlled exploration strategies, Policy Decorator enables stable, sample-efficient online learning.

State-to-Visual DAgger method diagram
When Should We Prefer State-to-Visual DAgger Over Visual Reinforcement Learning?

Tongzhou Mu*, Zhaoyang Li*, Stanislaw Strzelecki*, Xiu Yuan, Yunchao Yao, Litian Liang, Hao Su (* equal contribution)

  • AAAI 2025

arXiv Code

Visual reinforcement learning trains policies directly from visual observations, but faces challenges in sample efficiency and compute cost. This study empirically compares State-to-Visual DAgger — a two-stage framework that first trains a state policy, then uses online imitation to learn a visual policy — against Visual RL across a diverse set of tasks.

Enhanced Contactless Salt-Collecting Solar Desalination

Yue Bian, Zhihao Ye, Gengyou Zhao, Kun Tang, Yan Tang, Si Chen, Lijuan Zhao, Xiu Yuan, Shunming Zhu, Jiandong Ye, Hai Lu, Yi Yang, Lan Fu, Shulin Gu

  • ACS Applied Materials & Interfaces 2022

Paper

A paper-based thermal radiation-enabled evaporation system (TREES) achieves sustainable, highly efficient salt-collecting desalination, featuring a dynamic evaporation front based on the accumulated salt layer where water serves as its own absorber via energy down-conversion. On 7 wt % brine it continuously evaporates water at 2.25 L m−2 h−1 under 1 sun illumination — well beyond the input solar energy limit — for over 366 hours.

Carbonized magnolia fruit evaporator
Carbonized Tree-Like Furry Magnolia Fruit-Based Evaporator Replicating the Feat of Plant Transpiration

Yue Bian, Yang Shen, Kun Tang, Qianqian Du, Licai Hao, Dongyang Liu, Jinggang Hao, Dong Zhou, Xiaokun Wang, Huiling Zhang, Peiye Li, Yimeng Sang, Xiu Yuan, Lijuan Zhao, Jiandong Ye, Bin Liu, Hai Lu, Yi Yang, Rong Zhang, Youdou Zheng, Xiang Xiong, Shulin Gu

  • Global Challenges 2019

Paper

Magnolia fruits, as tree-like living organisms, can become outstanding 3D evaporators through a simple carbonization process. The mini tree possesses omnidirectional high light absorptance with minimized heat loss and gains energy from the environment. Water confined in the fruit has reduced vaporization enthalpy and transports quickly following Murray's law, achieving a record-high vapor generation rate of 1.22 kg m−2 h−1 in the dark and 3.15 kg m−2 h−1 under 1 sun illumination.