FRAME: A Modular Framework for Autonomous Map Merging: Advancements in the Field

TL;DR

FRAME tackles autonomous 3D map merging for egocentric multi-robot exploration by leveraging learned place recognition and yaw-discrepancy descriptors to detect overlaps and generate an initial alignment, which is refined with fast_gicp. Its modular architecture allows swapping descriptors, while adaptive keyframeSampling and adaptive sphere-radius selection minimize manual tuning and computation. Field experiments across multiple subterranean environments demonstrate sub-meter translation and sub-degree rotational errors with sub-second runtimes, outperforming traditional map-merge frameworks in speed and robustness. The framework’s real-time capability and modularity support multi-session exploration in challenging, GNSS-denied underground settings, with clear potential for mining applications and large-scale robotic collaboration.

Abstract

In this article, a novel approach for merging 3D point cloud maps in the context of egocentric multi-robot exploration is presented. Unlike traditional methods, the proposed approach leverages state-of-the-art place recognition and learned descriptors to efficiently detect overlap between maps, eliminating the need for the time-consuming global feature extraction and feature matching process. The estimated overlapping regions are used to calculate a homogeneous rigid transform, which serves as an initial condition for the GICP point cloud registration algorithm to refine the alignment between the maps. The advantages of this approach include faster processing time, improved accuracy, and increased robustness in challenging environments. Furthermore, the effectiveness of the proposed framework is successfully demonstrated through multiple field missions of robot exploration in a variety of different underground environments.

Figures (24)

• Figure 1: (A) One of the custom-built quad rotors that was utilized during the series of experiments in this article. (B) Spot in a construction area during the field trials. (Copyright NCC) (C) The outcome of the proposed framework from a larger scale indoor environment from Luleå University of Technology, where the two robots start together but explore different branches of the building.
• Figure 2: The coordinate frames for the map, denoted as $\mathbb{W}_ 1$ and $\mathbb{W}_2$, along with the robot frames $\mathbb{B}_1$ and $\mathbb{B}_2$, are depicted before and after the map alignment process. The transformations $\mathbf{T}_1$ and $\mathbf{T}_2$ represent known, non-static transforms between the robot and its static coordinate map frame. The spatial coordinate transform ${}^1\mathbf{T}_{2}$, derived from our proposed framework, facilitates the transformation of $\mathbb{W}_2$ to $\mathbb{W}_1$. All other transformations between each robot and map frame can be computed using the previously mentioned information.
• Figure 3: This figure illustrates the overall pipeline of the proposed framework. Both the SLAM process and Descriptor Extraction process are designed to be adaptable to different environments or use cases. White boxes denote the processes during exploration, while gray boxes indicate the merging process once it is initiated.
• Figure 4: The overall map merging pipeline of FRAME. While the robots $\mathbf{r}_1, \mathbf{r}_2$ explore the surroundings, they collect the vector sets $\mathbf{Q}$ and $\mathbf{W}$. A predefined event will trigger the merging process, and as an egocentric approach each robot will create its own merged map $\mathbf{M}$, maintaining its local map frame as the global frame.
• Figure 5: An example of the adaptive keyframe sampling thresholds: On the left, keyframes are illustrated on the map, while on the right, a plot depicts the variations in spaciousness alongside keyframe sampling. The threshold dynamically transitions from a smaller value, as depicted in (A), to a larger value, as demonstrated in (B).