Three-dimensional scene graphs provide robots with structured, queryable representations of their surroundings by encoding objects, their attributes, and spatial relationships in a unified graph. Existing methods for constructing such graphs typically require depth sensors and predefined exploration trajectories, limiting their applicability to platforms equipped with specialised hardware and carefully planned data-collection runs. This thesis presents a complete framework for active, incremental construction of 3D scene graphs from RGB images only. The framework is based on two main contributions. First, an RGB-only 3D scene graph generation pipeline that couples MapAnything, a state-of-the-art feed-forward 3D reconstruction model, with ConceptGraphs, an open-vocabulary semantic mapping system, and replaces the original LLM-based edge inference with a deterministic geometric edge generator.