Input-Contract-Aware Conditioning for Reproducible Volumetric Mapping Across Heterogeneous Robotic Platforms

Abstract

Volumetric mapping pipelines rely on assumptions regarding pose validity, frame interpretation, cloud-pose timing, and observation admissibility before integration. In practice, these assumptions are rarely made explicit, making mapping failures difficult to attribute and replay conditions difficult to reproduce across heterogeneous robotic platforms and data sources. This paper presents an explicit, auditable, and reproducible input-contract formulation together with a backend-preserving wrapper that enforces the contract while leaving the underlying mapping pipeline unchanged. The approach is evaluated through paired replay experiments on Spot, ANYmal, and a retained Dynablox data source using archived PCD-derived proxy metrics and a runtime timing-admissibility audit. Results show that the wrapper is non-invasive when input assumptions already hold. However, when they do not, it enables clear attribution of mapping differences to upstream assumption violations. These findings support input-contract-aware conditioning as a practical mechanism for making mapping assumptions visible, checkable, and reproducible.