ToolMind: Embodied Agent Pipeline
An end-to-end agent pipeline that lets a robotic arm perceive, reason about, and manipulate objects using RGB-D sensing and VLA-style planning.
- Embodied AI
- VLA
- RGB-D
Yuanyi Yan
Undergraduate student, School of Computer Science and Technology,
Huazhong University of Science and Technology (HUST)
I work on generative models,flow matching,embodied AI, andAI systems — recently focusing on training-free test-time compute for flow-matching generators and robotic-arm agent pipelines.
About
I am an undergraduate student at the School of Computer Science and Technology, Huazhong University of Science and Technology. My research centers on generative modeling — in particular flow matching, rectification-enhanced autoencoders (RAE), and Diffusion Transformers — and on training-free methods that make inference faster and better without retraining.
On the applied side, I build embodied AI systems: robotic-arm agents that combine RGB-D visual perception with LLM-driven planning, and the inference-serving infrastructure (e.g. SGLang-based deployments) needed to run them. I am broadly interested in HPC and efficient large-model inference, and I enjoy working across the stack from algorithms to systems.
Research Highlights
Looped Flow Matching
A training-free inference framework that repeatedly refines intermediate hidden states in frozen flow-matching image generators.
- Flow Matching
- DiT
- Test-Time Compute
News
- Started building this personal website.
- Started the ToolMind embodied-agent project.
- Exploring training-free test-time compute for flow matching models.
Honors & Awards
- 2025National Scholarship — Ministry of Education, China
- 2024First-Class Academic Excellence Scholarship — Huazhong University of Science and Technology
- 2024Meritorious Winner, Mathematical Contest in Modeling (MCM/ICM) — COMAP