AstraLOD3: Zero-shot multimodal agentic reconstruction of LOD3 building models

AstraLOD3 addresses the challenge of generating structured LOD3 building models—which include detailed architectural components like windows, doors, and roof features—from heterogeneous reality-capture data. Unlike traditional methods that rely on fixed, purpose-built geometric or learning-based pipelines, this study introduces an agentic framework. The system uses a general-purpose multimodal foundation model (Astra GPT) that operates under bounded autonomy. It takes multi-view images, calibrated camera parameters, and a sparse Structure-from-Motion (SfM) point cloud as input, alongside a natural-language reconstruction specification. The agent then dynamically plans and executes computational tasks using Python and Blender to generate the 3D model.

The framework was evaluated across 24 diverse building cases. AstraLOD3 achieved a mean Facade Re-projection Dice Score (FRDS) of 0.9647, demonstrating high consistency between the generated models and the source visual evidence. Quantitative comparisons against previous purpose-built reconstruction methods showed that AstraLOD3 generally provides superior image-space agreement and recovers a richer set of architectural components while maintaining comparable geometric fidelity. Controlled experiments revealed that the agent can adapt its computational trajectory based on the available evidence modalities and the level of guidance provided in the reconstruction specification, proving that structured reconstruction can be effectively formulated as a dynamic, agent-directed process.

This research shifts the paradigm of automated building reconstruction from rigid, algorithm-heavy pipelines to flexible, evidence-driven agentic workflows. By demonstrating that a general-purpose foundation model can perform complex 3D reconstruction tasks zero-shot, the study opens new possibilities for scalable digital twin generation. This approach reduces the need for task-specific training data and manual intervention, allowing for more adaptive and application-specific modeling in fields ranging from infrastructure management to disaster damage assessment.