Projects
Agent applications, robotics software, and computational imaging
QQ Agent: Memory & Multiplayer Interaction
Extended an existing QQ plugin and the deepseek-harness runtime with persistent notes, a D&D game-master workflow, and a reusable sticker collection.
- Built Markdown-based notes with BM25 retrieval, search/save tools, and relevant context injection across conversations.
- Connected D&D tools for dice rolls, combat checks, character updates, and JSON game saves. Injected DM rules and current game summaries into the system prompt.
- Improved turn handling with explicit player controls and character, round, and bound-player checks. Skipped actions were no longer observed in subsequent multiplayer testing.
- Added sticker collection with source and size checks, SHA-256 deduplication, emotion labels, and a dynamically updated catalog.
The Agent runtime and QQ transport are provided by the existing framework. Natural-language intent and some exception handling still depend on model behavior.
Robocon: Task Control & Device Integration
Developed upper-computer task logic and a fixed-length serial protocol to coordinate robot movement, ball pickup, and placement.
- Used action types and parameters to communicate with the chassis controller, and position feedback to decide when to stop.
- Combined action timing and camera observations to check ball release before moving to the next task; supported integration, debugging, and competition operation.
The chassis microcontroller handled wheel-level motion conversion. My work focused on upper-computer logic and communication.
Generalized Deep Null-Space Image Reconstruction
Investigated total-variation reconstruction, residual U-Nets, and output correction for data-consistent image reconstruction.
- Implemented finite-iteration PDHG for baseline reconstruction, training-data generation, and network-output correction.
- Evaluated reconstruction quality and measurement residuals on BSDS500 compressed sensing and simulated undersampling of fastMRI magnitude images.
RGB-D Dense Reconstruction with ORB-SLAM3
Built a dense point-cloud reconstruction pipeline using the poses and keyframes provided by ORB-SLAM3.
- Implemented depth back-projection, world-coordinate transforms, keyframe cloud fusion, and voxel filtering.
- Regenerated the point-cloud map from saved keyframes after pose corrections associated with relocalization or loop closure.
