At WeRide, I work on autonomous driving simulation, sensor modeling, and 3D scene reconstruction, and currently lead sensor simulation algorithms.

My recent work includes:

• closed-loop simulation for multi-agent interaction, RL post-training, and large-scale testing

• high-fidelity LiDAR/camera simulation across weather, lighting, and dynamic conditions

• 3D Gaussian Splatting based driving-scene reconstruction and dynamic asset extraction

• feed-forward generative 3D scene systems with text/image/HD map conditioning for rendering and LiDAR simulation

• Git-like HD map maintenance and data pipelines for continuous map update and training-data production

Proposed a loss function by fusing a classification loss with a weighted target-relevance loss for online fine-tuning.

By applying this method to the MDNet, we observed a performance improvement with an increase in accuracy from

53.2% to 53.9% and in robustness from 31.5 to 27.6 on the VOT dataset.

Positional Tracking Group

Merged multi-session maps from the same location.

Improved features matching by selecting distinct descriptors based on TF-IDF scores and gravity information.

Built a feature selector to pick up stable points for better map building.

Used a CNN-PCA method to detect needles in MRI with an accuracy improvement from 75.0% to 93.1%.

Deployed the model in a real-time surgery system and co-developed the website: http://needlefinder.org.