Collaborative Exploration with a Marsupial Ground-Aerial Robot Team through Task-Driven Map Compression
Angelos Zacharia, Mihir Dharmadhikari, Kostas Alexis · 2025
Read Collaborative Exploration with a Marsupial Ground-Aerial Robot Team through Task-Driven Map Compression by Angelos Zacharia, Mihir Dharmadhikari, Kostas Alexis (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-07655.txt.
Introduction
Angelos Zacharia and 2 co-authors design collaborative exploration for a marsupial robot team in a 2025 paper. A marsupial team pairs a ground robot that carries an aerial robot, releasing it when flight helps. Exploration of unknown, confined, or large spaces is crucial for such robots, yet communication between them is limited, so shared maps must stay small.
A graph-based path planner guides exploration and deploys the drone where its expected gain clearly beats the ground robot's, such as open halls or spots the ground unit cannot reach. A task-driven map compression strategy shares volumetric maps at chosen resolutions while preserving exploration-critical detail, even at high compression rates. Simulation and real-world experiments validate the approach. Efficient teamwork like this suits search, inspection, and disaster response where bandwidth is scarce.
Catalog record
| CREATOR | Angelos Zacharia, Mihir Dharmadhikari, Kostas Alexis |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-07655
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Collaborative Exploration with a Marsupial Ground-Aerial Robot Team through Task-Driven Map Compression Angelos Zacharia, Mihir Dharmadhikari, Kostas Alexis Abstract Efficient exploration of unknown environments is crucial for autonomous robots, especially in confined and large-scale scenarios with limited communication. To address this challenge, we propose a collaborative exploration framework for a marsupial ground-aerial robot team that leverages the complementary capabilities of both platforms. The framework employs a graph-based path planning algorithm to guide exploration and deploy the aerial robot in areas where its expected gain significantly exceeds that of the ground robot, such as large open spaces or regions inaccessible to the ground platform, thereby maximizing coverage and efficiency. To facilitate large-scale spatial information sharing, we introduce a bandwidth-efficient, task-driven map compression strategy. This method enables each robot to reconstruct resolution-specific volumetric maps while preserving exploration-critical details, even at high compression rates. By selectively compressing and sharing key data, communication overhead is minimized, ensuring effective map integration for collaborative path planning. Simulation and real-world experiments validate the proposed approach, demonstrating its effectiveness in improving exploration efficiency while significantly reducing data transmission. I Introduction Advancements in robotic systems have facilitated their deployment across a wide range of autonomous missions. Both aerial and ground robots are now extensively utilized for various applications, including search and rescue [ 1 ] , surveillance [ 2 ] , inspection [ 3 ] , and exploration [ 4 ] . Efficiently exploring unknown environments remains a significant challenge, whether navigating confined indoor spaces or traversing expansive outdoor landscapes. Single-robot systems often encounter limitations in terms of speed, sensing range, and their ability to navigate complex terrains. These challenges can be effectively addressed through the deployment of heterogeneous robot teams, where the strengths of one robot complement the limitations of the others. Marsupial ground-aerial systems exemplify a collaborative approach, combining the robustness and load capacity of ground robots with the agility and three-dimensional mobility of aerial robots. This synergy enables the strategic deployment of the aerial robot to efficiently explore large open spaces and high-ceiling environments, providing rapid coverage and accessing areas beyond the ground robot’s reach for seamless mapping. Fig. 1: Combined map from a real-world experiment, where the marsupial ground–aerial robot team collaborated in exploration. The aerial robot took over after a physical obstacle (obs) prevented the ground robot’s progress. Collaborative exploration relies on efficient data sharing between robots. Real-time exchange of sensor data, maps, and positions enables coordination and prevents redundant work. However, differences in processing power, sensor types, and communication constraints—such as limited bandwidth, high latency, and intermittent links—pose significant challenges, especially in complex environments. To address these challenges, this paper presents a comprehensive framework that integrates planning and deployment strategies for a marsupial ground-aerial robot team. Unlike prior work that either focuses on communication-efficient compression of generic point cloud data or on heterogeneous robot coordination without scalable data sharing, our contributions are twofold. First, we propose a bandwidth-efficient, task-driven point cloud compression method tailored for volumetric map reconstruction at mission-relevant resolutions. By emphasizing occupancy-relevant structure over raw point cloud fidelity, our approach achieves high compression rates while retaining the information essential for planning. This method is open-sourced at https://github.com/ntnu-arl/pcl-vae/ . Second, we introduce a decentralized collaborative exploration framework that leverages the complementary capabilities of a marsupial robot team. It features an aerial robot deployment strategy, keyframe-based map sharing for coordinated planning, and an energy-aware regrouping strategy. The framework is validated in large-scale simulations and real-world experiments, demonstrating improvements in both exploration efficiency and bandwidth usage. More experimental results from the real-world trials can be found at https://ntnu-arl.github.io/marsupial-collaborative-exploration/ The remainder of this paper is organized as follows. Section II reviews related work on marsupial robot teams and map compression and sharing techniques. Section III details the proposed framework, followed by evaluation studies in Section IV . Finally, Section V concludes the paper. II Related Work II-A Marsupial Systems-of-Systems Research in marsupial robotics, particularly ground/aerial systems, has historically been sparse. However, recent advancements have emerged, notably through the efforts of several teams involved in the DARPA Subterranean Challenge [ 5 ] . The ground platforms primarily consisted of tracked vehicles [ 6 ] , rovers [ 7 ] , or legged robots [ 8 ] , while the aerial platforms were predominantly multirotor systems. Focusing on combined docking-and-recharging, the work in [ 9 ] integrated a VTOL drone with a quadruped robot. Arguably the most well-known example, the Ingenuity helicopter, has completed multiple missions on Mars after being ferried and launched by the Perseverance rover [ 10 ] . Beyond system design, a set of works relates to the problem of planning for marsupial systems. These include works on path and trajectory planning of tethered aerial-ground systems [ 11 , 12 ] , and stochastic assignment for the deployment of multiple marsupial robots [ 13 ] . The deployment of marsupial robots for multi-agent exploration is studied in [ 14 , 8 , 15 ] . In this framework, the issue of communication constraints has repetitively attracted attention [ 14 , 5 ] . II-B Point Cloud Compression and Map Sharing A set of methods have been developed to enable efficient point cloud compression, including both conventional and neural strategies. Exploiting predictive deep learning models and leveraging the image representation of LiDAR data, RIDDLE [ 16 ] achieves high-degree of compression. The contribution in [ 17 ] exploits recurrent neural networks for efficient compression, while the work in [ 18 ] leverages a convolutional autoencoder learning compact feature descriptors from point clouds. A survey on deep learning-based point cloud compression is presented in [ 19 ] . Targeting autonomous driving, [ 20 ] utilizes range image-based segmentation and clustering to reduce spatial redundancy, with video coding enhancing compression. [ 21 ] exploits spatial and temporal redundancies in point clouds for real-time, high-efficiency compression. Beyond the general application of compression, a niche body of work exists focusing on compression for map sharing in multi-robot operations [ 22 ] . RecNet [ 23 ] transforms 3D point clouds into compact range image embeddings for efficient encoding and sharing while it serves both the goal of place recognition tasks and collaborative mapping in resource-constrained settings. The work in [ 24 ] first maps 3\textrm{D} point clouds into panoramas, uses event-triggered updates, and applies frequency-domain point cloud compression for efficient multi-robot systems. Departing from the current state-of-the-art, this work prioritizes a high degree of point cloud compression by encoding into a latent representation that explicitly focuses on the information necessary to reconstruct the occupancy map for planning and collision avoidance. While prior methods focus on compression ratios in the order of 10\times - 80\times [ 23 , 24 , 25 ] , our approach targets and achieves 300\times compression, enabling efficient map sharing for collaborative exploration in communication-constrained environments. III Proposed Approach Fig. 2: Overview of the proposed collaborative exploration framework with bandwidth-efficient map sharing, employed by a marsupial ground-aerial robot team. As the ground robot explores while carrying the aerial robot, it continuously evaluates whether to deploy the aerial robot by comparing their respective exploration gains. At the same time, it compresses sparse point clouds, generating and storing keyframes. Upon deployment, a subset of keyframes is shared with