Research Appointments
Zayed University
Co-developed MulRobBench, an offline benchmark for safe and security-policy-compliant multimodal UAV agents.
- Co-defined protocol-conditioned UAV decision tasks and a 3,024-sample benchmark spanning 17 task-taxonomy nodes under visual degradation and noisy operator instructions.
- Built data generation, inference, and evaluation pipelines for 17 multimodal models, with semantic and structural diagnostics for safety-policy compliance.
- Co-first-authored the manuscript, submitted to Applied Soft Computing. Preprint.
Huawei MindSpore Open-Source Community
Contributed to the MindNLP open-source project within the MindSpore ecosystem, implementing and fine-tuning Autoformer, BEiT, and ALBERT models.
- Worked on model integration, training stability, and cross-task evaluation across NLP and vision-language tasks.
- Improved the reproducibility and robustness of training workflows through open-source issue, review, and documentation processes.
Automatic Inspection and Rescue Based on Drone Nest
Developed UAV-based inspection services for open-pit mine scenarios, including water counting, conveyor-belt deviation detection, and truck tracking.
- Used DeepSORT and YOLOv8 for practical UAV inspection and rescue pipelines.
- Optimized PyTorch detection models and exported them through ONNX for real-time inference on edge devices.