Can robots understand and remember 3D scenes over time?

Robots need to understand scenes that change over time—objects move, appear, disappear—while mapping semantic meaning to 3D space. DGSG-Mind combines 3D Gaussian splatting with scene graphs to track object instances robustly and build a spatial-semantic memory the robot can reason about. It beats competitors on zero-shot 3D visual grounding and works on real robots, handling dynamic updates without retraining.