AGV Path Planning Using Curiosity-Driven Deep Reinforcement Learning

In the field of intelligent manufacturing, automated guided vehicles (AGV s) are required to plan an optimal path to a destination quickly and safely in complex environments. Several research papers have shown that Deep Reinforcement Learning (DRL) can plan a reasonable path for AGVs. However, almost all studies do not consider the influence of sparse external rewards. In addition, most research still uses an unrealistic 2D grid for their path planning environments. To alleviate these problems, we propose a novel AGV path planning method and provide the virtual environments of AGV s that approximate the real physical world. Moreover, the curiosity mechanism helps to enhance the Proximal Policy Optimization (PPO) method to provide additional intrinsic rewards to AGV agents in the sparse reward scenario. As for the environments, we no longer divide them into a 2D grid. Thus, the AGV has a continuous position in our environments. In addition, dynamic obstacles are added to make our environments more realistic. Based on our experiments, we make qualitative and quantitative analyses of the performance of the proposed method and the baseline DRL method. The empirical results show that our proposed method can make AGVs plan a reasonable and safe path with considerable speed during the training process in our virtual environments.