High-Efficiency Coding for Multi-Agent Communication via Characteristic-Aware Latent Rectification
Abstract
The rapid advancement of multi-agent systems has catalyzed increasing interest in Multi-Agent Communication (MAC). However, exchanging massive task-oriented multimodal data introduces a critical bandwidth bottleneck. While existing coding standards target human perception or point-to-point machine vision, coding strategies for MAC remain largely unexplored. To address this challenge, we formulate coding for multi-agent communication as a new problem and propose a coding paradigm motivated by the characteristics of MAC systems. Specifically, this paradigm is built upon two key principles. First, we introduce a Transform-before-Coding principle with a Characteristic-Aware Latent Rectification mechanism, which maps irregular multimodal representations into an entropy-friendly latent manifold for efficient coding. Second, we propose a Communication-Guided Coding principle that leverages inter-agent interactions to align the coding process with collaborative objectives. Furthermore, we instantiate the proposed paradigm for multi-agent collaborative perception in autonomous driving by developing an efficient Bird's-Eye-View (BEV) feature compression scheme equipped with Sparse Mask-Guided Attention and Learnable Mask Modulator. Extensive experiments demonstrate that our approach seamlessly integrates with existing collaborative perception systems, reducing communication bandwidth by up to several hundred times while preserving competitive collaborative performance, thereby validating the effectiveness of the proposed coding paradigm.