USS: Unifying Spatial-Semantic Prompting for End to End Embodied Visual Tracking
This work introduces USS, an end-to-end framework for embodied visual tracking that uses unified spatial-semantic prompting. It found that explicit spatial prompts (like bounding boxes) outperform text-only prompts in difficult tracking scenarios, especially when targets are similar or the tracking period is long. The system achieves this by using dedicated encoders for different prompt types and a latent world model to improve temporal stability.