I am a junior undergraduate student majoring in Artificial Intelligence at Tongji University. My core research interest lies in Embodied AI, specifically in developing adaptive control frameworks and reinforcement learning agents that enable robots to navigate and interact with highly constrained, dynamic environments.

我是范于翔，同济大学人工智能专业大三在读。我的研究方向为具身智能，致力于开发自适应控制框架与强化学习智能体，使机器人能够在受限且动态的环境中完成高精度的导航与交互任务。

• Autonomous Manipulation in Constrained Environments
  • Developed an Asymmetric Hybrid Control strategy for precision manipulation of flexible instruments, integrating PPO-based Reinforcement Learning with interpretable PID controllers.
  • Utilized the SOFA Framework for physics-based modeling and behavioral extraction to bridge the simulation-to-reality gap.
  • Note: This work is currently under double-blind review (IEEE T-ASE); details are kept general to maintain anonymity.
  • 针对受限空间操作开发了基于 PPO 强化学习与可解释 PID 的异步混合控制策略；利用 SOFA 框架进行物理建模与行为提取以弥合虚实鸿沟。（注：该工作处于 T-ASE 双盲审阶段，仅展示通用描述。）
• Hierarchical Federated Learning
  • Explored graph-based edge server clustering to improve communication efficiency in distributed AI systems.
  • Resulted in a co-first author paper accepted at ICC 2026 WS.
  • 研究基于图论的边缘服务器聚类，以提升分布式 AI 系统的通信效率；以共同第一作者身份产出论文 1 篇（ICC 2026 WS 已接收）。
• Embodied Intelligence Literature
  • Contributed to the "Embodied Intelligence" chapter of the book "Foundations of Artificial Intelligence".
  • 参与编撰《人工智能基础》，负责撰写“具身智能”章节。

• Weakly Supervised Multiple Instance Learning (WS-MIL): Aggregating local instances for global decision-making under sparse labels.
• Deep Reinforcement Learning (DRL): Solving multi-objective combinatorial optimization via Masked PPO and complex reward shaping.
• Visual Foundation Models (VFM): Leveraging large-scale pre-trained weights (e.g., UNI / ViT-Large) for general feature extraction.
• Decoupled Knowledge Distillation (DKD): Refining feature transfer by decoupling target and non-target class knowledge.
• Hardware-Aware Model Compression (HAMC): Deep optimization (Pruning, SVD, INT8) tailored for hardware (NPU/FPGA) operator characteristics.
• Gated Feature Refinement: Utilizing Gated Linear Units (GLU) for domain-specific feature enhancement.
• Hybrid Spatiotemporal Modeling: Integrating Monte Carlo simulations with Copula-based dependency analysis for complex systems.
• Temporal Correlation Analysis: Extracting delayed responses in high-dimensional sensor data via lag-correlation and hybrid CNN-LSTM.

• Programming & Hardware: Python, C++, MATLAB, Verilog.
• Research & Modeling Tools: SOFA Framework, Blender, 3D Slicer, Inkscape, Git, LaTeX.

• Email: fanyuxiang@tongji.edu.cn
• Focus: Embodied AI | Dexterous Manipulation | General Robotics

“Exploring the intelligence that emerges through physical interaction.” “探索在物理交互中涌现的智能。”