Blog / Docs • Email • ORCID
I am currently a Master's student in Artificial Intelligence at Shanghai Normal University. My research and engineering efforts are heavily focused on Vision-Language-Action (VLA) models and Reinforcement Learning for physical robotic manipulation. I bridge the gap between open-world semantic understanding and contact-rich precision control.
- VLA Deployment: Fine-tuning and deploying state-of-the-art VLA models (
SmolVLA,π0.5) on the SO-ARM101 bimanual robot. - RL for Robotics: Implementing residual online Reinforcement Learning (e.g., RL Token) to decouple high-level reasoning from low-level muscle reflexes.
- Hardware-Software Co-design: Optimizing 50Hz control loops via Client-Server async inference architectures and addressing Sim-to-Real gaps using Isaac Lab / MuJoCo.
- Embodied AI & Robotics: LeRobot | SO-ARM101 | Action Chunking | Flow Matching | Sim2Real
- Reinforcement Learning: TD3 | PPO | Behavior Cloning (BC) | DAgger | Residual RL
- Deep Learning: PyTorch | HuggingFace | CUDA | 3D Gaussian Splatting
- Simulation: MuJoCo | Isaac Lab / Isaac Sim
- DevOps & Backend: Linux | Docker | FastAPI | Git | Claude Agentic Workflow
- 📄 GSplatLoc: Ultra-Precise Camera Localization via 3D Gaussian Splatting (1st Author)
- Formulated camera pose estimation as a gradient optimization problem to minimize translation and rotation errors in high-resolution 3DGS environments.
- 🦾 LeRobot (VLA Pipeline SOPs)
- Maintaining comprehensive Chinese engineering docs for
SO-ARM101 × VLA. Covering data collection, 3D-2D mask projection filtering, LoRA fine-tuning, and asynchronous inference loops.
- Maintaining comprehensive Chinese engineering docs for
- ⚡ fastapi-supabase-template
- A highly concurrent async backend scaffolding boilerplate, deeply integrated with Docker and built via agentic workflows.
"Replacing pure imitation with physical reflexes."