物体検出画像と深度画像を用いたCNNによる移動ロボットのEnd-to-End動作計画

抄録

For autonomous navigation of mobile robots, obstacle avoidance in consideration of the destination is an essential capability. In this paper, we focus on a mobile robot equipped with RGB-D camera and LiDAR sensors, and propose an end-to-end motion planner based on a convolutional neural network, CNN, through imitation learning. In order for the robot to avoid various obstacles, we generate novel object detection images from the original RGB images. The object detection and depth images are then fed as inputs to the CNN. In a fully connected layer, moreover, a direction angle to the destination is inputted. In the navigation experiments, we show that the robot based on the proposed motion planner is able to move toward the goal destination while avoiding collisions with various obstacles.