Learning Navigation via R-VIN on Road Graphs

Guiding vehicles to their destination is an essential service. Nowadays navigation systems are mainly relying on the traffic conditions of road network, and other influence factors are not taken into account accurately, which is easy to lead to imbalance between the supply and demand of roads, resulting in congestion. In this paper, we introduce an online guiding approach via value iteration network on road graphs, R-VIN for short, which is an end-to-end planning model. In R-VIN, a large-scale real GPS trajectories are mapped via map-matching based road topology, which enables R-VIN to catch the experienced driving knowledge. Then we propose a conversion method from irregular road graphs to regular grid images to formalize the learning model. For a global optimum, ConvLSTM is used to predict the future traffic situation to form "prediction reward" of R-VIN. Combining with "current reward", a double rewarded VIN is used to solve the plan-involved function. Lastly, we train and evaluate R-VIN on planning problem in road networks, showing that R-VIN can achieve segment-based autonomous navigation with high top-k accuracy and less commuting time.