Deep neural networks-based real-time optimal navigation for an automatic guided vehicle with static and dynamic obstacles

Abstract

In this paper, a hybrid-intelligent real-time optimal control approach based on deep neural networks (DNNs) is proposed to improve the autonomy and intelligence of automatic guided vehicles (AGVs) navigation control. We first formulate the motion planning problem of an AGV with static and dynamic obstacles as a nonlinear optimal control problem (OCP). Subsequently, a direct method incorporating a smooth transformation technique is utilized to obtain the high-fidelity optimal solutions off-line. The optimal state-action samples are then extracted from the optimal trajectories generated by the OCP from perturbed initial states and stored as a large optimal trajectory data set. Thereafter, the DNNs architecture is designed and trained off-line through the collected data set to learn the optimal state-action relationship. As a result, the well-trained DNNs are used to produce the corresponding to optimal feedback actions on-board. The main advantage of our approach is that the time-consuming computation process is only carried out off-line and thus the traditional repeated off-line planning for the AGV due to changes such as the perturbed initial conditions in the system can be effectively avoided. Numerical experimental results are carried out to demonstrate our designed DNNs-based optimal control approach can generate the optimal control instructions on-board to steer the AGV to the desired location with high robustness to initial conditions as well as satisfying different obstacle constraints.