In multi-agent robotic exploration, managing and effectively utilizing the vast, heterogeneous data generated from dynamic environments poses a significant challenge. Federated learning (FL) is a promising approach for distributed mapping, addressing the challenges of decentralized data in collaborative learning. FL enables joint model training across multiple agents without requiring the centralization or sharing of raw data, overcoming bandwidth and storage constraints. Our approach leverages implicit neural mapping, representing maps as continuous functions learned by neural networks, for compact and adaptable representations. We further enhance this approach with meta-initialization on Earth datasets, pre-training the network to quickly learn new map structures. This combination demonstrates strong generalization to diverse domains like Martian terrain and glaciers. We rigorously evaluate this approach, demonstrating its effectiveness for real-world deployment in multi-agent exploration scenarios.

翻译：在多智能体机器人探索中，管理和有效利用动态环境中产生的大量异构数据是一项重大挑战。联邦学习（FL）是一种有前景的分布式地图构建方法，能应对协作学习中数据分散的挑战。FL使得多个智能体能够在不集中或共享原始数据的情况下进行联合模型训练，从而克服带宽与存储限制。我们的方法利用隐式神经映射技术，通过神经网络学习连续函数来表征地图，从而获得紧凑且自适应的表示。我们进一步采用基于地球数据集的元初始化策略，对网络进行预训练以快速学习新地图结构。这种组合展现出对火星地形、冰川等不同领域的强泛化能力。我们对该方法进行了严格评估，证明了其在多智能体探索场景中实际部署的有效性。