Machine Learning-Driven Edge Caching for Industry 5.0: A Survey of Collaborative and Inference-Enabled Techniques

Industry 5.0 is transforming edge caching from a passive traffic offloading mechanism into an active enabler of distributed intelligence and real-time analytics. However, conventional caching strategies lack the collaborative adaptability, inference-awareness, and domain-specific customization required for dynamic Industry 5.0 workloads. This survey provides a comprehensive and systematic review of machine learning (ML)-driven edge caching organized around two fundamental pillars:collaborative caching(distributed, federated, and multi-agent reinforcement learning paradigms) andinference-enabled caching(content and model placement accelerating real-time edge analytics). We systematically examine how traditional ML, deep learning (DL), deep reinforcement learning (DRL), and multi-agent RL (MARL) techniques address domain-specific challenges across five Industry 5.0 applications: robotics, manufacturing, logistics, immersive training, and healthcare. For each domain, we provide comparative assessments based on latency requirements, scalability targets, and collaboration needs. Our analysis reveals fundamental tradeoffs: MARL excels for distributed coordination with promising efficiency gains but requires higher computational resources, while DRL balances adaptation and complexity, and federated learning enables privacy-preserving collaboration across multiple sites. Each application section concludes with Key Insights synthesizing technique suitability, domain requirements, performance gains, and unresolved challenges. We identify five critical research directions: Bayesian optimization for sampleefficient learning, federated workflow-aware caching with sub-second staleness, resilient caching under intermittent connectivity, scalable multi-modal caching with low latency, and mission-critical prioritization meeting industry standards. This work provides a roadmap for designing adaptive, safe, and scalable ML-driven edge caching systems for intelligent industrial operations.