Evaluating the Effect of Local Map Completion on Navigation

Abstract

Map completion has been shown to improve mapdependent robotic tasks such as navigation by predicting unobserved parts of the environment. Most existing approaches either perform global map completion or predict local regions around the robot, and evaluate performance using high-level metrics such as coverage or path length. In contrast, this paper investigates how local map completion around frontiers affects the internal behavior of a standard ROS navigation stack. Our goal is not to introduce a new completion model, but to provide a system-level evaluation of how frontier-centric completion affects navigation stability under SLAM noise. To isolate this effect from model-specific factors, we use an aligned ground-truth occupancy map as an oracle that emulates a near-perfect completion module. Local patches are extracted around selected frontiers and fused into the active navigation map. We study increasingly stronger completion strategies, including global wall closing and invalid frontier filtering, within a frontier-based exploration pipeline. We evaluate their impact using navigation-specific metrics such as global planning cycles and navigation duration. By treating map completion as a modular capability rather than a specific model, the proposed methodology provides a model-agnostic diagnostic framework and an upper bound on the navigation benefits achievable through local completion in classical ROS pipelines.