Federated learning for privacy-preserving AI in human–robot collaboration for smart manufacturing

Purpose The study aims to address privacy and security challenges in AI-driven human–robot collaboration (HRC) by developing a privacy-preserving federated learning framework. Traditional centralized AI models expose sensitive manufacturing data to cybersecurity risks, creating barriers to AI adoption in regulated industries. This research proposes a decentralized learning approach that enables robots to collaboratively train AI models without sharing raw data, ensuring compliance with privacy regulations (e.g. GDPR and CCPA). The study seeks to advance trustworthy AI-driven automation, improving robotic decision-making, scalability and real-time adaptability while safeguarding sensitive industrial information. Design/methodology/approach This study proposes a Multi-Agent Federated Reinforcement Learning (MARL-FL) framework for privacy-preserving AI in human–robot collaboration (HRC) for smart manufacturing. The framework integrates federated learning (FL), reinforcement learning (RL) and differential privacy to enhance robotic decision-making while ensuring data security. A digital twin simulation of a smart factory is used for evaluation, where collaborative robots autonomously learn and optimize tasks using decentralized AI training. Performance is assessed using model accuracy, task success rate, convergence speed and privacy leakage reduction metrics, demonstrating FL’s effectiveness in improving secure AI-driven automation. Findings Experimental results from a digital twin-based smart factory simulation demonstrate that the proposed FL-based framework achieves 91.2% model accuracy, improves task success rates by 7.6% and reduces privacy leakage risks by 41.5% compared to centralized AI models. The federated reinforcement learning approach also accelerates model convergence by 25%, enabling faster adaptation to dynamic manufacturing conditions. The study confirms that FL enhances AI-driven collaboration, operational efficiency and data security, making it a viable solution for privacy-preserving smart manufacturing. Originality/value This research is among the first to integrate federated learning, reinforcement learning and privacy-preserving AI techniques for secure human–robot collaboration in Industry 4.0. Unlike conventional AI models that rely on centralized data processing, the proposed MARL-FL framework enables secure, decentralized learning, reducing cybersecurity risks and regulatory concerns. The study provides new insights into privacy-aware AI governance in industrial automation, making it highly valuable for researchers, policymakers and manufacturers seeking trustworthy AI-driven robotics solutions.