Resolving Cognitive Dissonance for Agentic AI via Self-Distilled Collaborative Perception

Abstract

With the progressive evolution of physical entities at the network edge into embodied intelligent agents, the paradigm of network intelligence is shifting from isolated processing to Agentic AI collaboration. As a cornerstone for transcending the physical sensing limits of individual agents and enabling swarm intelligence, collaborative perception (CP) still faces a severe challenge: cognitive dissonance when integrating shared perceptual information from heterogeneous viewpoints. In particular, viewpoint-induced discrepancies in raw feature transmission are prone to cause semantic misalignment, which in turn substantially degrades collaborative accuracy. To address these challenges, we propose SDCooper, a semantically invariant Self-Distilled Collaborative Perception framework. Specifically, we introduce a disparity-guided contrastive learning mechanism that utilizes physical disparity as a supervisory signal to filter viewpoint-specific disturbances. Furthermore, we propose a novel cross-view self-knowledge distillation strategy to align local perception with global consensus, enabling agents to internalize collective environmental cognition without incurring additional inference overhead. Experiments on OPV2V and DAIR-V2X datasets demonstrate that SDCooper significantly enhances collaborative robustness and cognitive consistency, outperforming state-of-the-art methods by 2.30% and 5.08% in terms of AP@0.7 respectively. Meanwhile, its plug-and-play nature validates the scalability of the framework in constructing resilient Agentic AI networks.