wendellz524@gmail.com · (+86) 15392924328
- Algorithms & Models: 3DGS, Diffusion Models (FLUX-fill, WAN2.1), VLA (Simlingo), Reinforcement Learning, BEVFusion, PersFormer, LaMa
- Simulation & Infra: Unreal Engine 5, Carla, Carla-RL, Ray, Nunchaku (SVD Quant), Docker, Ceph (DFS), Ansible, Prometheus, Grafana
- Programming: Python, C++, PyTorch
2024.9 – Present
Closed-Loop Simulation & Reinforcement Learning
- 3DGS × VLA Co-Simulation: Designed a co-simulation pipeline coupling Simlingo (VLA model) with 3DGS-rendered environments, enabling photorealistic visual feedback for embodied-AI closed-loop training.
- Distributed RL Training: Orchestrated distributed closed-loop RL experiments inside 3DGS AEB tracks using Ray for multi-node scheduling and Carla-RL as the training backbone; validated policy convergence under NCAP-equivalent scenarios.
- AEB Closed-Loop Reconstruction: Aligned 3DGS reconstructions to UE5 coordinate frames and re-engineered the Carla simulation engine, enabling deterministic end-to-end AEB evaluation without physical track dependency.
- Edge-Case Data Generation: Physical AEB test track rental costs ¥20,000/day with ~20 clips collectible per day due to setup, collision, and reset overhead. Delivered equivalent simulation data T+1 with unlimited volume and full trajectory compliance, reducing data acquisition cost by >99% (~1,000 clips generated).
Cross-Domain Data Synthesis & Generative Augmentation
- Traffic Sign Domain Transfer: Engineered a pipeline embedding UE5 photorealistic traffic sign assets into 3DGS-reconstructed domestic road scenes, generating geographically diverse perception datasets.
- Image Harmonization: Applied the Harmonizer framework for lighting-consistent compositing, achieving pixel-level color and luminance coherence between inserted assets and background scenes.
- Lane-Line Inpainting at Scale: Deployed LaMa to systematically erase false-positive lane predictions in complex rural scenarios, producing 83+ hrs (10,000 clips) of clean ground-truth data for model fine-tuning.
- Long-Tail Augmentation via FLUX-fill: Built a fully automated retouching pipeline using FLUX-fill to directionally over-sample rare traffic sign categories, reducing class imbalance without manual collection.
- Inference Optimization: Integrated Nunchaku SVD quantization to compress large generative models to run on a single RTX 4090, significantly reducing the cost of generative data synthesis.
2022.2 – 2024.9
Perception Research & Ground Truth
- Monocular 3D Lane Detection: Developed a lane detection model fusing PersFormer's perspective transformer with BEVFusion, incorporating LiDAR point clouds as both geometric priors (improving anchor L2 precision) and positional embeddings (improving recall). Both branches were retained in the final architecture after ablation. Deployed on proprietary datasets spanning highway and urban scenarios, achieving F1 98% on highway (single-class) and 72% recall on urban scenes across a fine-grained taxonomy of 10+ categories including road edges, dashed-to-solid transitions, colored double lines, and snow-occluded markings.
- Active Learning & Hard-Sample Mining: Applied perceptual hashing and IMU-derived motion signatures to identify divergent, annotation-worthy frames, reducing labeling cost while increasing model coverage on edge cases.
- Annotation Quality Assurance: Deployed CLIP-IQA for image quality filtering and Grounded-SAM for semantic verification of lane-line labels, reducing annotation error rate across large-scale datasets.
AI Infrastructure & DevOps
- Cluster Deployment: Stood up the department's on-premise GPU training cluster from bare-metal, including a self-hosted Docker registry for image version management.
- Automation & Monitoring: Automated node configuration and provisioning via Ansible playbooks; built a Grafana + Prometheus observability stack covering GPU utilization, job queues, and storage health.
- Distributed File System: Architected a high-throughput Ceph DFS tuned for random-read CV workloads, supporting concurrent access from the full training cluster without I/O bottlenecks.
2019 – 2021
- Digital Twin for Connected Vehicles: Built a Unity-based digital twin for Toyota Infotech integrating SUMO traffic simulation, physical vehicle telemetry, and LGSVL autonomous driving simulator via a unified middleware bridge.
- VR Rendering Optimization: Profiled and optimized PBR material pipelines to sustain a stable 45 Hz frame rate in Oculus Rift 2 on RTX 2080 hardware, meeting comfort thresholds for extended VR use.
| Degree | Institution | Period |
|---|---|---|
| M.S. Information Technology | The University of Melbourne | 2021 – 2022 |
| Exchange Student | University of California, Riverside | 2019 – 2020 |
| B.S. Electronic & Computer Engineering | Wuhan University of Technology | 2017 – 2021 |
- Source-Free Semi-Supervised Domain Adaptation (2022): Adapted a GTA5-pretrained DeepLabV3 segmentation model to the Cityscapes domain using pseudo-labels derived from semantic-kitti, without access to source-domain data during adaptation.
- Feature-Based Indoor Localization (2020): Constructed a multi-scale VGG feature index for image-retrieval-based localization; integrated OANet for robust homography estimation to recover 6-DoF query pose.