Perception balance with uncertainty-guided fusion and proposal-wise mixture-of-experts for robust multi-agent 3D object detection

LiDAR-based multi-agent 3D object detection is central to autonomous driving perception systems. However, most studies optimize collaborative performance peak while overlooking non-collaborative mode, where collaboration-tuned systems running in isolation fall below independently trained single-agent baselines. In this work, we revisit this system imbalance and present the Independence-Collaboration Perception Balance (ICPB) framework, which explicitly models epistemic uncertainty to adaptively trade off single-agent perception against collaborative gains. Unlike prior collaborative methods that implicitly mitigate domain shifts in the collaborative regime, ICPB employs an uncertainty-aware, loosely coupled fusion mechanism. It sequentially estimates uncertainty at both feature and proposal levels via a Hyper-dimensional Uncertainty-aware Self-Attention Fusion module (HuSaF) and a Proposal-wise Uncertainty-aware Mixture-of-Experts module (PuMoE). To further facilitate stabilize learning, a progressive dual-teacher distillation aligns the unified student with both individual and collaborative teachers, preserving independent competence with a Label-guided Knowledge Distillation (LGKD) while adapting to collaboration by alternate supervision. Extensive experiments under communication degradation, agent dropout, and asynchrony show that ICPB consistently reduces performance drop and surpasses collaboration-tuned and single-agent baselines, providing system-level robustness for safety-critical practical downstream applications.