Selected Projects

We propose a highly efficient GPU-accelerated 4D path planning framework for Autonomous Underwater Gliders (AUGs) operating in complex, time-varying ocean currents. To overcome the curse of dimensionality inherent in long-range missions, we introduce a fixed-dimensional B-spline trajectory encoding. Dynamic feasibility is strictly verified through RK4 integration, while continuous-domain collision safety over complex seabed topography is guaranteed using an ESDF-based Sphere Tracing strategy, eliminating the tunneling effect seen in traditional discrete sampling.

To tackle the highly multimodal search landscape induced by ocean currents and terrain, we develop a novel optimizer named MadDE-NDA (Niching Dual-Archive). It features dual cooperative archives to balance quality retention and diversity preservation, coupled with a stagnation-triggered niching mechanism. Accelerated by massive GPU parallelism for population-level evaluation, the framework demonstrates superior safety, efficiency, and robustness in realistic simulations utilizing high-resolution GEBCO bathymetry and CMEMS spatiotemporal ocean currents.

In collaboration with the Shandong Port Yantai Port Science and Innovation Center, this project targets rapid mesh asset generation, MB-level model archiving, non-contact cargo acquisition, and real-scale restoration for port breakbulk cargo digital twins. Based on Tencent Hunyuan3D-2.1, the system builds an integrated workflow covering image acquisition, generative 3D reconstruction, scale calibration, and digital-twin-ready asset export.

For fast white-mesh generation after crane-triggered image capture, I integrated HiCache++ into the Shape DiT sampling loop by reusing historical velocity fields and skipping selected DiT forward passes. On an NVIDIA RTX 4080, the 30-step shape sampling time was reduced from 40 s to 7 s, and the shape end-to-end latency decreased from 67.3 s to 24.0 s, achieving about 2.8x acceleration while preserving the downstream mesh pipeline.

To address the 16GB VRAM limit of a single RTX 4080, I designed a FastAPI + Redis + PostgreSQL asynchronous serving system with separate shape and texture queues. GPU0 runs a resident white-mesh worker, while GPU1 runs a resident PBR texture worker. The service supports priority scheduling, backpressure, retry/recovery, job auditing, artifact tracking, and resource monitoring. The measured Shape/Texture P95 latencies are 45.4 s / 321.3 s, corresponding to about 269 textured GLB assets per day on one dual-4080 node.

To improve Hunyuan3D-2.1 on irregular port cargo, I used Hunyuan3D-3.1 as a teacher model to construct 100 real yard-image / high-quality GLB pseudo-label pairs. After mesh cleaning, watertight conversion, SDF sampling, and multi-view conditional rendering, a Rank-16 LoRA domain adaptation was trained for cargo geometry. On the validation set, Chamfer-L1 decreased from 0.0511 to 0.0327 and F-score@0.05 improved from 0.670 to 0.793.