Estimating Map Completeness in Robot Exploration

TL;DR

This work tackles stopping criteria and map completeness in frontier-based indoor exploration by learning from partial grid maps. It introduces two CNN heads on a shared EfficientNet B1 backbone to (i) classify whether a partial map $M_t$ is sufficiently explored and (ii) regress the explored-area fraction $\hat{A}_t$, using $I_t$ derived from $M_t$. Trained on a large synthetic dataset of over $20{,}000$ partial maps with labels based on full environment maps, the method achieves high classification accuracy and accurate area estimation, enabling substantial exploration-time reductions (up to around $40\%$ online) while remaining robust to real-world domain shifts. This approach is practical, strategy-agnostic, and provides interpretable Grad-CAM visualizations to guide further exploration, offering a scalable tool for speeding up mapping and informing higher-level planning.

Abstract

In this paper, we propose a method that, given a partial grid map of an indoor environment built by an autonomous mobile robot, estimates the amount of the explored area represented in the map, as well as whether the uncovered part is still worth being explored or not. Our method is based on a deep convolutional neural network trained on data from partially explored environments with annotations derived from the knowledge of the entire map (which is not available when the network is used for inference). We show how such a network can be used to define a stopping criterion to terminate the exploration process when it is no longer adding relevant details about the environment to the map, saving, on average, 40% of the total exploration time with respect to covering all the area of the environment.

Figures (5)

• Figure 1: The map after $t=90min$ of exploration (a) already represents almost all of the environment (the portion of the mapped area is $A_{t}=0.98$), except a few uninteresting corners and portions of rooms scattered across the whole environment, highlighted in blue in (a). This exploration run ends at $T=150min$ (b) when all frontiers have been visited. During the time interval $T-t$, the robot moves back and forth to explore the remaining frontiers, possibly jeopardizing the whole mapping process. Our method infers that the map in (a) is almost fully explored and stops the exploration, reducing the total exploration time.
• Figure 2: Examples of partial maps and labels from our dataset.
• Figure 3: Incremental examples of our method applied to partial maps obtained after 10, 20, 30, and 40 minutes of exploration of two environments, from batch-wise assessment. In the environment in the first row (a)-(d), the exploration with the baseline stopping criterion ends after $T=70min$; with our method, we can reduce such time by 30 minutes (d). In the environment in the second row (e)-(h), the exploration with the baseline stopping criterion is concluded after $T=90min$; we reduce such time by 50 minutes (h). The areas highlighted in the partial maps are the locations that the neural network of our method identifies as relevant to make its prediction (red and blue denotes the most and the least relevant regions, respectively).
• Figure 4: The error in predicting the percentage of explored area while the percentage of the area increases. The standard deviation is in light blue.
• Figure 5: Examples of real-world maps, all labeled as not-explored. Our method can highlight the parts relevant to explore as well as predict the explored area $\hat{A}$.