Towards Robust Multi-UAV Collaboration: MARL with Noise-Resilient Communication and Attention Mechanisms

TL;DR

The paper tackles robust, information-rich path planning for multiple UAVs in full 3D under noisy communications. It introduces a MARL framework based on Counterfactual Multi-Agent Policy Gradients (COMA), augmented with the SenDFuse Network for sensor denoising and fusion and an attention-enhanced CBAM-based Actor-Critic. Key contributions include extending NestFusion to multi-UAV fusion, offline pretraining of SenDFuse, and centralized training with a counterfactual baseline for credit assignment, achieving improved information gain in 3D IPP tasks. Experiments on synthetic and real-world data demonstrate superior path planning efficiency and noise robustness, including up to a 78% reduction in entropy, highlighting practical impact for scalable, robust UAV information collection.

Abstract

Efficient path planning for unmanned aerial vehicles (UAVs) is crucial in remote sensing and information collection. As task scales expand, the cooperative deployment of multiple UAVs significantly improves information collection efficiency. However, collaborative communication and decision-making for multiple UAVs remain major challenges in path planning, especially in noisy environments. To efficiently accomplish complex information collection tasks in 3D space and address robust communication issues, we propose a multi-agent reinforcement learning (MARL) framework for UAV path planning based on the Counterfactual Multi-Agent Policy Gradients (COMA) algorithm. The framework incorporates attention mechanism-based UAV communication protocol and training-deployment system, significantly improving communication robustness and individual decision-making capabilities in noisy conditions. Experiments conducted on both synthetic and real-world datasets demonstrate that our method outperforms existing algorithms in terms of path planning efficiency and robustness, especially in noisy environments, achieving a 78\% improvement in entropy reduction.

Figures (7)

• Figure 1: Multiple drones collaborate to scan an area of interest. They communicate in a noisy environment, exchange data, perform sensor fusion with denoising, and plan paths independently. By working together, they enhance the understanding of the region.
• Figure 2: The overall structure of our proposed framework primarily illustrates the data flow and processing modules. Multiple drones sample an area of interest, and each drone fuses the visual data from multiple agents using the SenDFuse Network. The fused visual perspective is then combined with other local information and stacked as input to the Actor network, which makes decisions. The input to the Critic network includes both the input to the Actor network and global information, serving to evaluate the performance of the Actor network.
• Figure 3: The SenDFuse Network operates in two phases: training and deployment. During training, artificial noise is added to the input data, and the network is trained to reconstruct noise-free images. In the deployment phase, an attention-based fusion strategy is applied to process deep-level features.
• Figure 4: Our experimental environments includes three types: Env1: synthetic data, Env2: thermal imaging data, and Env3: visible-light data. Note that in Env3 we stipulate that drones are not allowed to fly into building areas, and the sampling category for building areas is designated as valueless.
• Figure 5: Env3: Reward variations for each method over 1,200 training episodes. Among them, Base represents method with the absence of SenDFuse and CBAM. The method incorporating both major modules performs the best in terms of both convergence speed and final convergence value.