Contract-guarded orchestration of the UAV photogrammetry-to-engine handoff: artifact contracts, guarded language-model repair, and a fault-injection benchmark

Abstract

This paper presents a reliable AI-agent framework for automating the delivery of UAV photogrammetric reconstructions into real-time engines for visualization and digital-twin applications. Conventional workflows often rely on process exit status, allowing semantic errors—such as incorrect units, coordinate systems, material bindings, textures, or file dependencies—to remain undetected until engine import. To address this problem, the study introduces contract-guarded orchestration, in which every pipeline boundary is governed by explicit, machine-checkable artifact contracts. Known violations are repaired using deterministic rules, while unresolved cases are passed to a constrained language model. The model may inspect and repair the workspace but cannot perform the final export or declare success. The framework was demonstrated using a 79-image urban UAV survey, producing a georeferenced reconstruction with 15 million dense points and an engine-compatible textured asset. In a 144-trial fault-injection benchmark, the proposed controller recovered 96.3% of recoverable faults with no false completion. It outperformed both a fixed-rule controller and an unconstrained language-model agent while using only 36% of the latter’s prompt tokens and 35% of its execution time. The paper also evaluates reconstruction repeatability, showing that run-to-run variation can exceed apparent differences between reconstruction configurations. Overall, artifact contracts provide workflow safety, while constrained language-model repair improves adaptability and fault recovery.