SPW-Nav: A Streaming Panoramic World Model for Language-Guided Navigation
SPW-Nav turns a single panorama and changing language commands into a real-time, one-minute 2K 360-degree video stream.
The paper frames the model as a navigation-oriented panoramic world model, not just a fixed-path video generator. It treats each instruction as camera motion in the previously generated panorama, with methods for spherical rotation, bounded translation conditioning, and memory across changing instructions. The authors also introduce SPW-NavSet, a dataset of panoramic videos with trajectories and verified instructions. They report better camera-following accuracy and video quality than prior panoramic generators, including support for switching instructions mid-stream. HF Daily Papers' note
The paper frames the model as a navigation-oriented panoramic world model, not just a fixed-path video generator. It treats each instruction as camera motion in the previously generated panorama, with methods for spherical rotation, bounded translation conditioning, and memory across changing instructions. The authors also introduce SPW-NavSet, a dataset of panoramic videos with trajectories and verified instructions. They report better camera-following accuracy and video quality than prior panoramic generators, including support for switching instructions mid-stream. HF Daily Papers' note
score 5