Yu-Cheng Chou

VFlash is a System-1 VLM for fast, accurate visual planning in Physical AI. It predicts complete geometric primitives in parallel, from bounding boxes to 3D trajectories and motion. Across robotics and autonomous driving, VFlash improves spatial accuracy and accelerates inference, achieving 42.6% lower average 3D robot trajectory error and up to 5.2× faster robot inference.

Captain Safari is a world engine for exploring 3D scenes through camera-controlled video generation. It retrieves pose-aligned 3D memory to preserve scene geometry while following challenging camera paths, including sharp turns and large viewpoint changes. On complex outdoor scenes, Captain Safari improves 3D consistency and camera control, earning 67.3% of 1,500 votes in a five-way human preference study.

SpatialReasoner is a VLM for understanding and reasoning about the 3D world from images. It explicitly represents object locations and orientations, using them to connect visual perception with geometric computation and multi-step reasoning. SpatialReasoner generalizes to unseen spatial questions and achieves 60.3% accuracy on 3DSRBench, outperforming Gemini 2.0 by 9.2%.