Balancing safety and efficiency in large-scale HRC: A dynamic scheduling method confronting temporal and spatial uncertainty of human workers.

In large-scale human-robot collaboration (HRC) workshops, humans exhibit dual uncertainty: temporal uncertainty in task completion times and spatial uncertainty in limb movements. These intertwined uncertainties create a fundamental tension between safety and efficiency. We propose a dynamic scheduling framework addressing both dimensions. For spatial uncertainty, a Perception-Prediction-Decision (PPD) framework is constructed, integrating Age-of-Information-based link resource allocation, digital twin elastic deployment, and probabilistic occupancy grid-based collision prediction to proactively maintain safety. For temporal uncertainty, an Improved Deep Q-Network (IDQN) algorithm is proposed, featuring an idle-duration-based executor priority mechanism, constraint-based action masking, and dynamic priority experience replay. Across 15 test scenarios, the PPD framework reduces collision response triggers by 59.9% while decreasing average Makespan by 12.1%; IDQN achieves only 2.5% performance degradation under high-volatility conditions, significantly outperforming DQN (6.0%), GA (9.7%), and rule-based methods (8.4%).