[1G3] ロボットと実世界-人間拡張・動きの学習

[1G3-03] Learning-based Selective Dual-arm Grasping for Warehouse Picking

〇Shingo Kitagawa¹, Kentaro Wada¹, Kei Okada¹, Masayuki Inaba¹ (1. the University of Tokyo)

キーワード：Robot manipulation, Grasp Learning, Deep Learning

We propose a learning-based system of selective dual-arm grasping and use Convolutional Neural Networks (CNN) for grasping point prediction and semantic segmentation. First, the network learns grasping points with the automatic annotation. and the grasping points are automatically calculated based on the shape of an object and annotated for both single-arm and dual-arm grasping. The robot then samples various grasping points with both grasping ways and learns optimal grasping points and grasping way. As a result of multi-stage learning, the robot learns to select and execute optimal grasping way depending on the object status. In the experiments with the real robot, we demonstrated that our system worked well in warehouse picking task.