Deep Neural Network-Based Real-time Trajectory Planning for an Automatic Guided Vehicle with Obstacles

Abstract

In this paper, a hybrid-intelligent real-time optimal control approach based on deep neural networks (DNNs) for an automatic guided vehicle (AGV) with obstacles is studied. The path planning problem of the AGV is formulated as an optimal control problem (OCP) and the numerical solution of the OCP can be obtained by discretizing the OCP into a nonlinear programming (NLP) problem. However, due to the change of the starting point and the high nonlinearity of the model, the solution process is usually difficult without good initial guesses. Multiple iterations are often needed and real-time performance cannot be easily achieved. To this end, we first use the Gauss pseudospectral method incorporated a smoothing technique to solve the AGV optimal control problem with high computational efficiency offline. Subsequently, a DNNs is developed to learn the functional relationship between the optimal state and action pairs obtained by the approximate indirect method, and thus the resulting DNNs can generate the optimal control instructions in real time for the AGV navigation. Numerical results are illustrated to verify the real-time performance, control optimality and robustness of the developed DNN-based controller.