Reinforcement learning-based optimization approach for dynamic multi robot pickup scheduling

The Dynamic Multi-Robot Pickup Scheduling (DMPS) problem, common in logistics and manufacturing, presents challenges due to its high dynamism and real-time requirements, where traditional methods struggle with responsiveness. This paper proposes Attention–Mask–RND Proximal Policy Optimization (AMRPPO), a reinforcement learning framework integrating attention mechanisms, action masking, and Random Network Distillation (RND) to tackle complex information, strong action constraints, and sparse rewards. A Markov Decision Process (MDP) is formulated to model proactive task discovery and dynamic execution. Attention-based feature extraction and action masking enhance policy generalization and feasibility, while RND addresses reward sparsity. Experiments demonstrate that AMRPPO outperforms baseline PPO with a 12% higher average reward (-1621.08) and improved stability (standard deviation of 111.01). Its per-step computation time is 0.01 s, significantly faster than OR-Tools' 9.7 s, confirming the method’s suitability for real-time dynamic scheduling.