Learning-Based Task Failure Prediction for Selective Dual-Arm Manipulation in Warehouse Stowing

Stowing is one main task of warehouse automation, and manipulation with a vacuum gripper is recently known as a practical method. However, the gripper sticks an object from upper side, which causes task failures such as drop and protrusion by even small disturbance. In this paper, we aim to realize more stable stowing task and propose a stowing system which robot selectively stow an object by two arms in case the task failures may occur. For the selective stowing, we predict task failure occurrence by convolutional neural network (CNN) and select a proper motion from the prediction results. The network predicts probabilities of task failure occurrence for both single-arm and dual-arm stowing motion cases, and we design a motion select algorithm to evaluate the two motions and select optimal one. In experiment, we implemented our system in real stowing task and achieved higher success rate 58.0% than that of single-arm stowing system 49.0% in 100 trials.