Jinyang Wang (王金阳)

Autonomous Driving Algorithm Intern · Wan2.1 DiT · 6-Camera Video Generation · RGB-LiDAR Streaming

• Multi-view video generation. Owned the six-camera branch of a pose-controllable driving world model based on Wan2.1 DiT, extending the data pipeline, unified latent layout, and synchronized surround-view generation; designed ring-style local Cross-View Attention, reducing allowed cross-view Q-K connections by 50%.
• Geometric consistency. Designed a front-view Plücker Ray Map and Epipolar-Masked Attention for cross-view and RGB-LiDAR token interaction, improving VSC from 0.771 to 0.830 and reducing DAS from 2.41 m to 2.19 m.
• Long-context training. Adapted 24-way Attention Sequence Parallel, view-sharded VAE encoding, latent-level aggregation, and pre-sharded Plücker conditioning, increasing the maximum training length from 113 to 381 frames.
• Real-time world generation. Contributed to the DMD Distillation, Self-Forcing, and Temporal-Decay KV Cache pipeline, helping the system reach approximately 10 Hz RGB-LiDAR streaming generation with long-horizon rollouts beyond 240 frames.

Vision-Language Navigation · 3D Reconstruction · Robot Deployment

• Built a generative world-model-driven navigation pipeline from natural-language instruction → target-scene generation → trajectory recovery → real-observation refinement, combining LLM and world-model priors for robot goal localization and navigation decision making.
• Reproduced and integrated DUSt3R and VGGT for multi-view reconstruction, including camera-pose estimation, dense point-cloud reconstruction, and scale calibration; completed 500+ image-processing annotations and participated in embedded deployment and unmanned-robot system integration.