Scholarlink

Multi-robot systems deployed in diverse, unstructured environments often face significant domain shifts, leading to poor generalization when models are solely trained on local data. Federated learning (FL) offers a promising solution by enabling collaborative model training without centralized data aggregation, but conventional FL methods struggle with non-IID data distributions across robots. To address this challenge, FedGDA (Federated Global Domain Alignment) is proposed, which is a novel FL framework that explicitly mitigates domain shift through a combination of kernel-based maximum mean discrepancy (MMD) loss for feature alignment and domain adversarial training for invariant representation learning. Meanwhile, an adaptive mechanism that dynamically adjusts alignment strength based on observed local-global discrepancies is introduced to enhance robustness in heterogeneous settings. A number of experiments on multi-robot maze exploration tasks demonstrate that FedGDA significantly outperforms traditional FL methods (FedAvg, FedProx, PerAvg) in classification accuracy and generalization to unseen environments, establishing its efficacy in real-world robotic applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FedGDA: A Federated Domain Alignment Approach for Multi-robot Maze Exploration

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chenhao Ye,
  • Hongxue Huang

摘要

Multi-robot systems deployed in diverse, unstructured environments often face significant domain shifts, leading to poor generalization when models are solely trained on local data. Federated learning (FL) offers a promising solution by enabling collaborative model training without centralized data aggregation, but conventional FL methods struggle with non-IID data distributions across robots. To address this challenge, FedGDA (Federated Global Domain Alignment) is proposed, which is a novel FL framework that explicitly mitigates domain shift through a combination of kernel-based maximum mean discrepancy (MMD) loss for feature alignment and domain adversarial training for invariant representation learning. Meanwhile, an adaptive mechanism that dynamically adjusts alignment strength based on observed local-global discrepancies is introduced to enhance robustness in heterogeneous settings. A number of experiments on multi-robot maze exploration tasks demonstrate that FedGDA significantly outperforms traditional FL methods (FedAvg, FedProx, PerAvg) in classification accuracy and generalization to unseen environments, establishing its efficacy in real-world robotic applications.