Simulator of the navigation equipped with LIDAR of the mobile robot based on the neural network

Abstract

The paper considers the issue of using the robot simulators equipped with LIDAR (Light Detection and Ranging). The simulator application allows decreasing the experiment costs, their time and risks or negative aftereffects of accidents, but, at the same time, the simulated environment should identical to the real one. In practice, the simulated environment can have differences and there comes the issue of estimating the robot ability to adapt to new conditions. Neural networks are used to teach the robots, due to which the robots reach the target in the shortest or the fastest way, at the same time, there should be no accident/collision. The initial series represents the multiple time series. To teach the neural networks to control the robots, the paper proposes to introduce the additional index in the form of the minimal distance to the nearest object in the actual time moment. The approach novelty in using the new penalty function for teaching the neural network is the speed index of approaching the nearest object/obstacle. The neural networks were taught on the maps without dynamic objects, but tested in the conditions when several mobile robots were used on the map, thus, they became the dynamic obstacles for each other. The accident risk for robots without the proposed penalty function during testing was 4.5% against 0.3% - with the penalty function.