Abstract: Aiming at the problems of multi-AGV (automated guided vehicle) path planning in intelligent warehouse, such as poor real-time performance, weak obstacle recognition ability, multi-AGV collision, deadlock and congestion, a collaborative control method for intelligent warehouse traffic signals control and multi-AGV path planning is proposed. Traffic signals and multi-AGV path planning are regarded as a whole in this method. A collaborative control framework for traffic signals and multi-AGV path planning is designed. The LS-A3C (long short-asynchronous advantage actor-critic) algorithm and Bi-LSTM-CBAM (bi-long short-term memory-convolutional block attention module) algorithm are proposed as the core algorithm of the framework. The long-term and short-term information of traffic signals are encoded in LS-A3C algorithm, which uses a long short-term encoder and attention mechanism. They are represented by learning cell features. The A3C framework is used to calculate the Q value of the cell and the control strategy. The traffic signals time adapting to AGV flow is adjusted to solve the problems of multi-AGV collision, deadlock and congestion. The output result is spliced by calculating the state characteristics of the present moment and the leading moment in Bi-LSTM-CBAM algorithm to solve the problem of gradient disappearance and explosion in neural network effectively and improve real-time AGV path planning. The attention mechanism module CBAM is introduced. Weights are assigned based on how important the input is in order to strengthen the AGV’s ability to identify obstacles. Finally, simulation experiments are carried out on Sumo and Gazebo joint simulation platform. The experimental results show that the collaborative control method significantly reduces the AGV collision, deadlock and congestion, significantly improves the obstacle recognition ability, and greatly enhances the real-time path planning. The purpose of improving the AGV operation efficiency is achieved.

Key words: intelligent warehouse, deep reinforcement learning, path planning, bi-long short-term memory (Bi-LSTM), asynchronous advantage actor-critic (A3C), convolutional block attention module (CBAM)

摘要： 针对智能仓储多AGV（automated guided vehicle）路径规划实时性差，障碍物识别能力弱，多AGV碰撞、死锁和拥堵等问题，提出了一种智能仓储交通信号控制与多AGV路径规划协同控制方法，将交通信号与多AGV路径规划视为一个整体，设计一种交通信号与多AGV路径规划协同控制框架，并提出LS-A3C（long short-asynchronous advantage actor-critic）算法和Bi-LSTM-CBAM（bi-long short-term memory-convolutional block attention module）算法作为框架的核心算法。LS-A3C算法使用长短时编码器和注意力机制分别对交通信号的长期信息和短期信息进行编码，以学习元特征表示，并使用A3C框架计算元Q值和控制策略，实现交通信号时间自适应AGV流量，解决多AGV碰撞、死锁和拥堵等问题。Bi-LSTM-CBAM算法通过计算本时刻和前置时刻状态特征，对输出结果进行拼接处理，可以有效解决神经网络梯度消失和爆炸的问题，提高AGV路径规划实时性；引入注意力机制模块CBAM，根据输入信息重要程度分配权重，加强AGV对障碍物识别能力。在Sumo和Gazebo联合仿真平台进行仿真实验，实验结果表明，该协同控制方法使AGV碰撞、死锁及拥堵情况明显降低，障碍物识别能力显著提高，路径规划实时性大幅增强，达到提升AGV作业效率的目的。

关键词: 智能仓储, 深度强化学习, 路径规划, Bi-LSTM, A3C, CBAM