Robot execution failure prediction using incomplete data

Robust execution of robotic tasks is a difficult learning problem. Whereas correctly functioning sensors' statements are consistent, partially corrupted or otherwise incomplete measurements will lead to inconsistencies within the robot's learning model of the environment. So, methods of prediction (classification) of robot failure detection with erroneous or incomplete data deserve more attention. A probabilistic approach for the classification of incomplete data (which has three versions) is developed and evaluated using five robot execution failures datasets. We show that by improving the estimation of probabilities, our approach offers considerable computational savings and outperforms the other methods.