About Me
I am currently an M.S. student in the Machine Learning program at Carnegie Mellon University, advised by Prof. Fernando De la Torre. Previously, I studied Computer Science (Elite Class) at Guohao College, Tongji University, where I ranked 2nd among 52 students.
My research interests lie in multimodal learning and embodied intelligence, with a particular focus on vision-language models (VLMs) and vision-language-action models (VLAs). I am interested in building intelligent systems that connect perception, language, reasoning, and action in the physical world.
You can find my full CV here. Feel free to contact me via email if you are interested in my work.
Selected Work
💃 Text-Driven 2D Human Motion Generation
- Developed a model to generate 2D human motion from textual descriptions, focusing on realism and adaptability.
- Proposed a novel dataset for two-person interactions, addressing a gap in existing research.
- Designed and implemented a Transformer-based model with a custom attention mechanism and integrated deep reinforcement learning strategies to optimize performance.
🤖 Real-Time Vision System for Robotics
- Designed and implemented automated targeting and shooting mechanisms for robotic vehicles, focusing on real-time object detection, tracking and fire control.
- 🏆 RoboMaster 2025 Super Confrontation: Champion (Eastern Region) & National Top 8 (National Final), Ranked #1 in overall auto-aim accuracy (Eastern Region).
- 🏆 RoboMaster 2025 University League: 2nd Runner-up (3rd Place) (Shanghai Regional).
📍 Real-Time 3D LiDAR SLAM and Relocalization System
- Developed a robust 3D localization framework in ROS 2 integrating Point-LIO with wheel odometry.
- Engineered a global relocalization pipeline based on Generalized ICP for fast and precise point cloud registration.
- Implemented dynamic point cloud processing modules, including moving object detection and ROI optimization using PCL.
🐧 Discrete Unix V6++ Operating System
- Addressed inefficiencies in the original Unix V6++ kernel, such as high external fragmentation from contiguous memory requirements and slow `fork()` performance.
- Implemented a comprehensive discretization of the process image, enabling non-contiguous memory allocation for processes.
- Introduced a copy-on-write (COW) mechanism to minimize overhead during process creation, significantly improving efficiency.
💻 MockMaster - An AI & HCI-Powered Mock Interview System
- An AI-powered mock interview system using computer vision and NLP to provide an immersive experience.
- Core features include real-time analysis of facial expressions and gestures, plus intelligent resume feedback.
- Supports both 1v1 interviews with an AI agent and simulations of multi-person group interview scenarios.