Building embodied agents that remember, learn, and evolve.

Yuanhong Yu

I am a Ph.D. student at Zhejiang University, supervised by Sida Peng and Xiaowei Zhou. My research focuses on persistent embodied intelligence and the foundations of lifelong agency in the physical world.

I am currently a research intern at OneRobotics's Artificial Intelligence Research Institute in Hangzhou, working on embodied agents for the real world. I also build open-source tools for agent collaboration and long-running workflows.

Previously, I received my B.Eng. in Computer Science from Northwestern Polytechnical University, advised by Jiaqi Yang.

Yuanhong Yu
ZJU-3DV ยท OneRobotics
Research

Publications

BoxDreamer: Dreaming Box Corners for Generalizable Object Pose Estimation

ICCV 2025

BoxDreamer: Dreaming Box Corners for Generalizable Object Pose Estimation

Yuanhong Yu, Xingyi He, Chen Zhao, Junhao Yu, Jiaqi Yang, Ruizhen Hu, Yujun Shen, Xing Zhu, Xiaowei Zhou, Sida Peng

We present BoxDreamer, a novel framework that estimates 6-DoF object poses by predicting 3D bounding box corners from a single RGB image. Our approach achieves strong generalization to unseen object categories without requiring CAD models or category-specific training, enabling real-world deployment across diverse scenarios.

Habitat-GS: A High-Fidelity Navigation Simulator with Dynamic Gaussian Splatting

ECCV 2026

Habitat-GS: A High-Fidelity Navigation Simulator with Dynamic Gaussian Splatting

Ziyuan Xia, Jingyi Xu, Chong Cui, Yuanhong Yu, Jiazhao Zhang, Qingsong Yan, Tao Ni, Junbo Chen, Xiaowei Zhou, Hujun Bao, Ruizhen Hu, Sida Peng

We introduce Habitat-GS, a photorealistic navigation simulator that replaces the mesh-based renderer of Habitat with a CUDA 3D Gaussian Splatting rasterizer and dynamic avatar modules, while staying compatible with the Habitat ecosystem so embodied agents can train and evaluate on visually realistic navigation tasks.

Latest

Preprints

EARTalking: End-to-end GPT-style Autoregressive Talking Head Synthesis with Frame-wise Control

arXiv 2026

EARTalking: End-to-end GPT-style Autoregressive Talking Head Synthesis with Frame-wise Control

Yuzhe Weng, Haotian Wang, Yuanhong Yu, Jun Du, Shan He, Xiaoyan Wu, Haoran Xu

We propose EARTalking, a novel end-to-end, GPT-style autoregressive model for interactive audio-driven talking head generation. Our method introduces a frame-by-frame, in-context, audio-driven streaming generation paradigm with Sink Frame Window Attention for variable-length video generation with identity consistency.

Building in Public

Open Source

View all projects

I build practical tools for understanding agent behavior and coordinating AI-assisted work. These are the two projects that best represent that direction today.