Collision avoidance system of drone based on deep reinforcement learning

Abstract

The collision avoidance function of the hobby drone currently on the market uses several sensors to determine the distance to an object and avoids it, so the collision avoidance performance in a complex environment is low. This study applied deep reinforcement learning to study a collision-avoidance system with higher performance than the current performance. Objects were recognized using a depth camera in front and a 2D LiDAR. The collision avoidance baseline code of a heuristic algorithm was written. The control model was trained by compensating for positive values when it follows a baseline, is fast, and there is no collision. As a result of comparing the learned collision avoidance system with the baseline, the collision avoidance success rate was the same or higher in various situations(include in the complex environment). The average obstacle passing speed was also higher than the baseline. In addition, when the learning model was applied to a real drone and tested in a real environment, the result of avoiding obstacles was obtained. Through this study, we developed a collision-avoidance system for drones that operate in reality as well as in simulation.