Modeling Spatial-Temporal Interactions for Robot Crowd Navigation

Safely and efficiently reaching the target in a pedestrian-rich environment is a demanding task for mobile robots. Robots are susceptible to other human movements in the presence of many pedestrians. Traditional human trajectory prediction methods have problems such as large amount of computation and information loss. In order to support robots to obtain more efficient and accurate navigation, we present a spatio-temporal graph framework based on convolutional neural network (CNN) and recurrent neural network (RNN) for robot navigation decisions in a pedestrian-rich environment. The framework models the spatial relationships between robot and the crowd and captures the spatio-temporal interactions using an improved temporal convolutional network (TCN). We use model-free deep reinforcement learning to train our network to achieve greater efficiency comparable to advanced methods. Experiments show that our model improves the real-time performance and stability compared with other advanced algorithms.