Abstract: In addressing the issues of low data utilization efficiency and challenging effective data collection in the context of using reinforcement learning algorithms for Automated Guided Vehicle (AGV) path planning in Robotic Mobile Fulfilment Systems (RMFS),a novel reinforcement learning training framework that combined the behavior cloning method and reward reconstruction method was proposed to enhance the training effectiveness of neural networks.The behavior cloning method accelerated the neural network's decision-making abilities by allowing it to directly learn from expert experience through supervised learning.Meanwhile,the reward reconstruction method improved the training effectiveness of reinforcement learning by designing a more precise reward function.Experimental results demonstrated that the combined use of behavior cloning and reward reconstruction methods in the reinforcement learning process significantly outperformed standard reinforcement learning algorithms that did not employ either of these methods.

Key words: robotic mobile fulfilment system, automated guided vehicle, path finding problem, policy gradient, behavioral cloning, reward reconstruction

摘要： 针对使用强化学习算法解决移动机器人拣选系统(RMFS)中AGV路径规划所存在的数据利用效率低、有效数据采集困难的问题,提出一种结合行为克隆方法和奖励重构方法的新的强化学习训练框架,来提升神经网络的训练效果。行为克隆方法通过监督学习的方式,让神经网络直接学习专家经验,来迅速提升神经网络的决策能力;奖励重构方法通过更加精细的奖励值函数设计,来提升强化学习的训练效果。实验表明,同时使用行为克隆方法与奖励重构方法的强化学习过程,其训练效果远优于标准的强化学习算法(既不使用行为克隆方法也不使用奖励重构方法)。

关键词: 移动机器人拣选系统, 自动导引小车, 路径规划, 策略梯度算法, 行为克隆, 奖励重构