VLA policy reproduction, training, evaluation, and data engineering for bimanual garment manipulation.

• Reproduced and evaluated ACT, Diffusion Policy, SmolVLA, X-VLA, pi0, and KAI0.
• Ran systematic W&B experiments on learning rates, training steps, checkpoints, pretrained weights, and image augmentation.
• Improved the Top-Long success rate from 76.67% to 86.67% under a unified 60-episode evaluation.
• Built an end-to-end data flywheel covering rollout, trajectory filtering, LeRobot Dataset merging, and retraining.
• Curated 291 high-quality successful trajectories.
• Fixed file_index misalignment and missing finalize() issues that caused Parquet and dataset merging failures.
• The team reached a final success rate of 59.25% and ranked 17th globally in the Simulation Track.

Semantic segmentation for smoke with weak boundaries, smoke-cloud confusion, and complex fire-safety scenes.

• Conducted more than ten smoke-burning experiments and collected over one hundred real smoke samples.
• Used SAM 3 for semi-automatic annotation and supplemented the dataset with 90 generated multi-scene smoke images.
• Collected and standardized 9,124 smoke segmentation images.
• Built a train / val / test = 8000 / 833 / 291 split, with 291 high-quality manually annotated samples as the independent test set.
• Introduced CE and Focal Tversky Loss into DeepLabV3+, improving mIoU from 78.4% to 82.4%.

An edge-model-based hazardous-chemical management system for laboratory intake, checkout, return, auditing, and anomaly tracking.

• Delivered and deployed the first runnable version within two days.
• Integrated face recognition, chemical-label OCR, and electronic-scale serial communication.
• Used InsightFace to associate operators with chemical transactions and abnormal events.
• Used RapidOCR and PaddleOCR for chemical-label recognition.
• Combined name normalization and fuzzy matching with an index of 41,977 IECSC chemical names.
• Deployed online at scut.leai.me.

A collaborative execution framework combining a GUI Agent and a Code Agent for computer-use tasks.

• Designed Smart Router, failure fallback, Tool Calling, and Agent Memory mechanisms.
• Supported screenshot understanding, window control, mouse and keyboard interaction, and code execution.
• Used Doubao UI-TARS 1.5 7B for graphical interface understanding.
• Deployed OpenCUA and conducted basic evaluations on OSWorld.
• Developed GUI and CLI interfaces, Docker support, automated tests, and CI workflows.
• Open-sourced the project and released an executable version.