Environmental situation perception and processing serve as critical foundations for multi-agent decision-making systems. Due to the inherent limitations of single-agent perception, recent studies have introduced collaborative perception to enhance perceptual performance by sharing complementary information through communication. However, it is crucial to address mixed real-world factors (e.g., bandwidth, latency) for its applications, which is rarely tackled by most existing works. To break through this bottleneck, we introduce a comprehensive perceptual information collaboration framework RealCP, which empowers agents to handle diverse multi-modal inputs and share sparse yet valuable spatial information in a compressed manner, enabling the adaptability to more complex real-world environments. RealCP has three distinct advantages: i) it can handle varying number of input sensors by utilizing low-coupling encoders and straightforward fusion strategy. ii) it considers more rational feature collection and compression to achieve higher perceptual performance by distinctive perceptual information preservation and conservatively selection; and iii) it improves the efficiency of cross-agent feature aggregation by more explicitly quantifying the individual contribution of each agent. To evaluate it, we consider 3D object detection in both simulated and real-world scenarios. Experimental results show it consistently outperforms other previous methods, which demonstrates its effectiveness and robustness.