Federated Multi-Agent Mapping for Planetary Exploration

TL;DR

This work tackles data-efficient collaboration for multi-agent planetary exploration under bandwidth constraints by introducing a federated, implicit 2D mapping framework. The method combines offline Earth-based meta-training with online, one-shot federated parameter sharing to build a robust global traversability map while minimizing data transmission. Key contributions include a Gaussian Fourier-feature encoding for high-fidelity spatial mapping, Reptile-based meta-initialization for rapid adaptation, and a FedAvg-based federated mapping pipeline with a lightweight 2D implicit mapper; experiments on Athabasca Glacier and Martian-like terrains show data-transfer reductions up to 93.8% and strong downstream path-planning performance (F1 up to ~0.95). The results demonstrate practical viability for autonomous, bandwidth-limited space missions and point toward extensions to 3D mapping and heterogeneous agent collaborations.

Abstract

Multi-agent robotic exploration stands to play an important role in space exploration as the next generation of robotic systems ventures to far-flung environments. A key challenge in this new paradigm will be to effectively share and utilize the vast amount of data generated onboard while operating in bandwidth-constrained regimes typical of space missions. Federated learning (FL) is a promising tool for bridging this gap. Drawing inspiration from the upcoming CADRE Lunar rover mission, we propose a federated multi-agent mapping approach that jointly trains a global map model across agents without transmitting raw data. Our method leverages implicit neural mapping to generate parsimonious, adaptable representations, reducing data transmission by up to 93.8% compared to raw maps. Furthermore, we enhance this approach with meta-initialization on Earth-based traversability datasets to significantly accelerate map convergence; reducing iterations required to reach target performance by 80% compared to random initialization. We demonstrate the efficacy of our approach on Martian terrains and glacier datasets, achieving downstream path planning F1 scores as high as 0.95 while outperforming on map reconstruction losses.

Figures (6)

• Figure 1: In future multi-agent space exploration, balancing communication efficiency with effective collaboration is paramount. Traditional approaches that rely on sharing all raw data back to a base station quickly become infeasible due to bandwidth constraints. Federated learning addresses this by allowing rovers to learn maps and adapt their skills locally. They then share only the trained models with a base station, minimizing communication overhead. This enables efficient creation of a shared, global map representation while empowering individual rover autonomy.
• Figure 2: CADRE rovers in a Clean Room cadre-rovers.
• Figure 3: Our proposed federated mapping approach begins with an offline preparation stage where a neural network is trained on an empty (unknown) grid map and then meta-trained with Reptile on a map dataset for quick adaptation (this is done only once, offline). Next, multiple agents explore within a global reference frame, generating local area maps. Each agent utilizes a neural network to learn its local map. These learned network parameters are then shared with a central server where model aggregation occurs. Finally, the updated joint model parameters, now containing global map knowledge, are distributed back to each robot. Locally, robots can further refine (e.g. remove noise) the global map generated from the updated model.
• Figure 4: Meta-initialization with Reptile enables faster network convergence for map reconstruction from KITTI data (far right). This is demonstrated by comparing a randomly initialized model (top) and its meta-initialized counterpart (bottom) from steps 0-4. Each step represents two optimization iterations.
• Figure 5: Athabasca Glacier: from landscape picture to map representation.