USS: Unified Spatial-Semantic Prompts for Embodied Visual Tracking with Latent Dynamics Learning

Researchers have proposed two new frameworks, USS and a self-supervised framework for learning implicit 3D physical dynamics, to improve Embodied Visual Tracking and physical dynamics prediction.

The USS framework uses unified spatial-semantic prompts instead of text-only specifications for Embodied Visual Tracking (EVT), allowing for more precise target indication in cluttered scenes^[1]. EVT requires an agent to continuously follow a target while moving through dynamic environments. Current paradigms rely on language-based target indication, which can be ambiguous. USS achieves state-of-the-art performance among non-MLLM-based methods and competitive results against recent MLLM-based approaches with faster inference speed^[1]. Meanwhile, a separate research effort presents a self-supervised framework for learning implicit 3D physical dynamics directly from video-derived supervisory signals. This framework unprojects semantic features from a Video Joint-Embedding Predictive Architecture (V-JEPA) into a voxelized grid and enables a Volumetric Feature Advection to learn an action-conditioned transition operator^[2]. The model demonstrates good long-term structural stability and physical plausibility on multiple benchmarks^[2].