CNN-Based Motion Planning for Object Storage by Dual-Arm Robot

In this paper, we focus on object storage by a dual-arm robot. The robot is required place objects grasped by its hands onto a shelf. The task is treated in terms of motion planning. For image inputs captured by a camera, the object storage motion outputs are determined by the robot itself based on a motion planner. For this purpose, we propose a motion planner based on Convolutional Neural Network, CNN. In order for the robot to plan the storage motion toward the shelf composed of multiple spaces, goal directions for both hands, indicating the center coordinates of target spaces, are used as inputs in addition to images and current hand positions. For these multimodal inputs into the motion planner, shortcut connections are further applied to convolutional layers in the CNN to adjust learning bias and extract image features. Through the experiments, we discuss the effectiveness of the motion planner with the goal directions and shortcut connections for object storage by the robot.