|
Email /
Github /
Google Scholar
|
 |
2024.10 - Present Undergraduate Research Intern Advisor: Jiaolong Yang （杨蛟龙）, Baining Guo（郭百宁）, Harry Shum（沈向洋） |
 |
2024.11 - 2025.11 Undergraduate Research Student Academic Advisor: Prof. Gao Huang （黄高） |
 |
2023.08 - 2027.07 Undergraduate Student, Department of Physics GPA ranking : 1/87 (4.00/4.00) |
My research focuses on the convergence of Physics, Computer Vision, Natural Language Processing, and Robotics, with specific interests in few-step image/video generation, robotic dexterous manipulation with tactile sensing, and multimodal AI. My objective is to create robust, generalizable robotic and generative systems that can operate effectively in complex, real-world settings.
|
MoGe-3: Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement
Lingyu Kong1,3,*,
Ruicheng Li1,3,*,
Ruicheng Wang2,3,*,
Sicheng Xu3,
Chengtang Yao3,
Jianfeng Xiang1,3,
Jiaolong Yang3
1Tsinghua University, 2USTC, 3Microsoft Research · *Equal contribution
A monocular geometry estimation model that recovers fine-grained 3D details via a self-guided sparse volumetric refinement (SSR) module, iteratively correcting distorted structures for high-fidelity point maps. |
|
FM-VLA: Force-based Memory for Vision-Language-Action Models in Contact-Rich Manipulation
Ruicheng Li1,2,†,
Qixiu Li1,2,†,
Ruichun Ma2,
Yu Deng2,
Lin Luo2,
Zhiying Du3,2,†,
Jianfeng Xiang1,2,†,
Huizhi Liang1,2,†,
Ruicheng Wang4,2,†,
Jiaolong Yang2,*,
Baining Guo2
1Tsinghua University, 2Microsoft Research, 3Fudan University, 4USTC · †Work done during internship at Microsoft Research · *Corresponding author
A Vision-Language-Action framework that leverages force-based memory to tackle contact-rich manipulation tasks, enabling robots to reason over interaction history for tasks such as hidden-block search and repeated pressing. |
|
Robot Parkour Enhancement and Sim2sim DeploymentOur work reveals key limitations in current visual control policies and offers insights into enhancing policy robustness and cross-domain generalization. |
|
SAMTM: SAM-based Trajectory Modeling for Efficient Policy LearningWe propose a trajectory modeling pipeline that incorporates semantic segmentation to guide point sampling. |
|
Can LLMS DO SAMPLING?Our work reveals a consistent failure mode in LLMs' probabilistic sampling: prompt-dependent determinism rather than true randomness. We propose a binary sampling benchmark to evaluate this issue and demonstrate that targeted fine-tuning can effectively mitigate it, showing this instability is not inherent but learnable. |
|
Intelligent Platform for Automated Management and Analysis of Computational ExperimentsOur project develops a dedicated Model Context Protocol (MCP) server for experiment tracking, interactive data visualization, and other practical functionalities, streamlining the entire machine learning workflow. |
|
Holographic Optical Tweezers and Real-time Phase ImagingWe developed a low-cost experimental platform that integrates holographic optical tweezers (HOTs) with real-time quantitative phase imaging. This work demonstrates the feasibility of combining these technologies for multi-trap manipulation and dynamic phase analysis of biological samples, offering a new approach for frontier biophysical experiments. |
|
Beyond the Bridge: Holistic Strategies for Baltimore's Transportation FutureWe modeled the Baltimore's transportation network and developed a comprehensive strategy to address traffic congestion, aging infrastructure, and the collapse of the Francis Scott Key Bridge. |
|
A novel data acquisition system tailored for dexterous hand manipulationBuilt a teleoperation system from scratch for dexterous hand data collection at Microsoft Research Asia. Focused on enabling efficient data acquisition for robotic manipulation. |
|
|