Inferring the correct meaning of natural language commands, as judged by the person who issues commands, is mandatory for natural language commanded robotic systems. There have been some successful research on this; but one of the important and related aspects has not been addressed, i.e. the possibility of learning from natural language commands. Since natural language commands are generated by human users, they contain valuable information. Nevertheless, the learning from such commands, as well as the interpretation of them face many challenges due to the inherent subjectiveness of natural languages. In this paper, we propose a decision making process for natural language commanded robots which is influenced by certain characteristics of human decision making process. The proposed concept is demonstrated with an experiment conducted using a robotic manipulator. First, the robot is controlled with natural language commands to perform some pick and place operations during which the robot builds a knowledge base. After learning, the robot is capable of performing approximately similar tasks by making approximate decisions with the gained knowledge. For the decision making a probabilistic neural network is used.