Object View Prediction with Aleatoric Uncertainty for Robotic Grasping

In vision-based robotic grasping the prediction of grasps is made on an RGB-D image of the scene from a pregrasp position. Due to perspective occlusions the information about the object to be grasped is incomplete. Therefore the reliable prediction of a successful grasp is difficult. In this study we investigate a convolutional neural network that can predict several depth views of an object from a single depth image and camera movements. We show how the heteroscedastic aleatoric uncertainty can be used to predict the uncertainty of the generated object views. As the classical Negative-Log-Likelihood loss function for the aleatoric uncertainty does not converge to a meaningful object view prediction, we investigate the modified ss-Negative-Log-Likelihood loss function and the Moment Matching loss. For assessing the predicted uncertainty we introduce a new measure, the uncertainty weighted intersection over union value.