時系列3軸触覚情報を用いた物体把持結果の予測

抄録

Most of robotic manipulation tasks are time-related process so it is important to handle time-series information for better manipulation performance. In this paper, we propose a LSTM based prediction system which predicts grasping outcome using time-series 3-axis tactile information. Tactile data was collected from grasping and early post-lifting stage using 25 objects. For evaluation, K-fold cross validation was used considering relatively small size of datasets. Also, ROC curve and PCA were used for analysis. As a result, the proposed model achieved 77.4% prediction accuracy and it performs better than a comparison model which does not utilize time-series information.