Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : June 02, 2018 - June 05, 2018
We propose a deep learning-based method for a manipulator to recognize “pickable” objects from randomly piled objects which are uniformly moved by automated guided vehicle (AGV) delivery. Using our method, the manipulator can achieve a piece picking operation in collaboration with an AGV without stopping it. In our study, we define the “pickability” for each object on the basis of whether it is possible or not to plan the manipulator’s motion to let its hand reach to the target object with avoidance of surrounding obstacles that move with the target. Under the definition, we design our recognition function as a convolutional neural network (CNN) for processing the objects’ image and develop our technique for preparing supervised data. In our simulation, the AGV runs at a velocity of 0.1m/sec, and our CNN can detect 20% more correctly pickable moving objects than the same structured CNN for paused objects.
