An MPPI-based Navigation Algorithm Using Trajectory Prediction and Pedestrian States with Hybrid Warning Policy

Pedestrians can be classified as cooperators and non-cooperators based on their awareness states. If a pedestrian is a cooperator, it has the ability to adjust its velocity to avoid collisions with robots. In contrast, non-cooperators are unable to do so. In crowded environments with large numbers of pedestrians, pedestrians are mostly cooperators, but some pedestrians are non-cooperators due to distractions such as using mobile phones. The movements of the robot encountering cooperators and non-cooperators are different, and collisions with non-cooperators are sometimes inevitable in crowded environments. To improve navigation efficiency and safety, we propose MPPIPS-HWP, model predictive path integral control (MPPI) using trajectory prediction and pedestrian states with the hybrid warning policy. MPPIPS-HWP includes a navigation policy, MPPIPS, and a warning policy, HWP. We propose STGN-S, a spatio-temporal graph network using pedestrian states, to predict pedestrian trajectories. Then, we create separate cost maps for each prediction time step, and the velocity is finally calculated by MPPI. The collision risk of the future trajectory and the distance to the nearest pedestrian are used to evaluate the necessity of performing a warning action. Experiments show that STGN-S has state-of-the-art performance in prediction accuracy, and MPPIPS-HWP has good performance in average speed, collision rate, and warning action frequency.