Scholarlink

<p>To address the problem of path conflicts and group deadlocks caused by limited local observations when logistics robots perform collaborative picking tasks in warehouse supply chains, this paper proposes an optimization decision model (Graph Attention Transformer-Multi-Agent Soft Actor-Critic, GAT-MASAC) that integrates Graph Attention Transformer and Multi-Agent Soft Actor-Critic (SAC) algorithms. First, a locally observable Markov decision process is constructed, and a dense reward function that balances goal orientation and collision avoidance is designed. Second, dynamic local spatiotemporal features are extracted using gated recurrent units and scaled dot-product graph attention networks. By introducing a feature dimension scaling factor to smooth the attention variance, high-dimensional feature redundancy and extreme weight skewness are effectively overcome, thereby identifying potential conflict nodes. Finally, maximum entropy regularization is introduced into the heterogeneous Actor-Critic architecture, giving the agents a wider action exploration space to autonomously plan suboptimal detour paths, thus effectively resolving group deadlocks. Meanwhile, the progressive learning approach, incorporating phase-based experience replay, effectively prevents the model from forgetting early basic obstacle avoidance skills in long-sequence, high-concurrency tasks. Experimental results in MovingAI benchmark scenarios demonstrate the model maintains a median pathfinding success rate of 0.95 to 0.97 and limits the maximum completion time degradation rate to 21% under 20% dynamic obstacle injection with 200 robots. Under long local communication radius constraints, deadlock frequency drops to 2 per 100 rounds with a 12.5% flow time offset rate, maintaining an average decision latency of 9 milliseconds. This research effectively resolves local congestion deadlocks, providing an expansion solution for large-scale dynamic logistics scheduling that balances high collaborative efficiency with low computational burden.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing collaborative picking path conflicts in warehouse supply chains by integrating graph attention transformer and multi-agent soft actor-critic (SAC) algorithms

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月17日

查看全文
  • Binbin Tan,
  • Gengxin Li,
  • Hongjiang Shang

摘要

To address the problem of path conflicts and group deadlocks caused by limited local observations when logistics robots perform collaborative picking tasks in warehouse supply chains, this paper proposes an optimization decision model (Graph Attention Transformer-Multi-Agent Soft Actor-Critic, GAT-MASAC) that integrates Graph Attention Transformer and Multi-Agent Soft Actor-Critic (SAC) algorithms. First, a locally observable Markov decision process is constructed, and a dense reward function that balances goal orientation and collision avoidance is designed. Second, dynamic local spatiotemporal features are extracted using gated recurrent units and scaled dot-product graph attention networks. By introducing a feature dimension scaling factor to smooth the attention variance, high-dimensional feature redundancy and extreme weight skewness are effectively overcome, thereby identifying potential conflict nodes. Finally, maximum entropy regularization is introduced into the heterogeneous Actor-Critic architecture, giving the agents a wider action exploration space to autonomously plan suboptimal detour paths, thus effectively resolving group deadlocks. Meanwhile, the progressive learning approach, incorporating phase-based experience replay, effectively prevents the model from forgetting early basic obstacle avoidance skills in long-sequence, high-concurrency tasks. Experimental results in MovingAI benchmark scenarios demonstrate the model maintains a median pathfinding success rate of 0.95 to 0.97 and limits the maximum completion time degradation rate to 21% under 20% dynamic obstacle injection with 200 robots. Under long local communication radius constraints, deadlock frequency drops to 2 per 100 rounds with a 12.5% flow time offset rate, maintaining an average decision latency of 9 milliseconds. This research effectively resolves local congestion deadlocks, providing an expansion solution for large-scale dynamic logistics scheduling that balances high collaborative efficiency with low computational burden.