I am a Mathematics undergraduate at Xi'an Jiaotong-Liverpool University (XJTLU), working across operations research, optimization, robotics and intelligent control, and applied AI. I am interested in how mathematical models and carefully designed experiments can support reliable decision-making in autonomous and intelligent systems.
My projects emphasize reproducible workflows, interpretable methods, honest scope boundaries, and evidence that can be inspected rather than presentation alone. I am currently developing research-oriented work in low-friction mobile-robot control, LLM self-correction evaluation, and adaptive dynamic programming.
ROS 2 · Gazebo · Python · Control · Robotics
A simulation-based study of adaptive safety control for a differential-drive AGV operating on low-friction surfaces. The project investigates startup false activation, executed command-motion mismatch, and the trade-off between stronger intervention and task efficiency.
The later ROS 2 / Gazebo workspace uses TurtleBot3 Burger and includes a formal 2 surfaces × 4 controllers × 5 trials experiment, for 40 completed runs. The study compares Baseline, Fixed, Base Adaptive (adaptive_v2), and Acceleration-Aware Adaptive (adaptive_v3) control. It also includes phase-based analysis, a separate controller ablation, supplementary friction-robustness experiments, statistical analysis, paper figures, and a manuscript in preparation.
The public repository currently contains an earlier deterministic two-dimensional prototype, checked-in outputs, and migration notes. The full ROS 2 workspace and formal-experiment raw data are not yet public, and the reported mismatch metric is not presented as a direct physical wheel-slip measurement.
LLM Agents · Evaluation · Python · GSM8K
A reproducible evaluation framework for asking a more careful question than “does self-correction improve accuracy?”: when does a revision fix a wrong answer, and when does it break a correct one?
The project compares Baseline, Self-Refine, Reflexion, and Tool Critic strategies on a fixed 100-question GSM8K subset. In the recorded evaluation, Baseline reached 98%; two-round Self-Refine reached 99% by fixing one answer and breaking none; Reflexion fixed two and broke two; and Tool Critic broke one without fixing another. The framework therefore reports answer transitions, net corrections, and final accuracy together. It includes cached experiment runs, offline tests, validation reports, and item-level result artifacts for reproducibility.
Adaptive Dynamic Programming · Control Theory · Scientific Communication
A paper-study and presentation-engineering project focused on value-iteration adaptive dynamic programming for nonlinear optimal control. It examines the distinction between asymptotic convergence and whether a finite-iteration controller is admissible and usable, together with the interpretation of critic and action networks.
The repository contains a ten-slide Chinese technical presentation, a matching speaker script, and reproducible document-generation code. Its scope is scientific reading and communication rather than a full numerical reproduction of the original paper.
- Operations Research and Optimization
- Applied Mathematics and Decision-Making under Uncertainty
- Robotics and Autonomous Systems
- Adaptive and Intelligent Control
- AI Agents and LLM Evaluation
- Reproducible Experimental Design
I am especially interested in problems where optimization, feedback, uncertainty, and empirical evaluation meet: choosing interventions that improve safety or reliability while making their costs and limitations explicit.
Programming: Python · C++ · MATLAB
Robotics: ROS 2 · Gazebo · TurtleBot3 · Nav2
Optimization: Mathematical Programming · Gurobi
Research: Experimental Design · Statistical Analysis · Scientific Writing · Reproducible Reporting
I am refining the AGV study into a research manuscript while preserving a clear boundary between simulation evidence and real-robot claims. In parallel, I am exploring evaluation methods for AI agents that reveal failure transitions rather than relying on a single aggregate score.
My longer-term direction is to connect mathematical optimization and decision-making with robotics, control, and dependable applied AI. I am particularly interested in graduate research environments where theoretical reasoning, implementation, and careful experimental validation are treated as parts of the same workflow.