Learning from Approximate Human Decisions by a Robot

Robot systems operating under natural-language commands must be able to infer the meaning intended by the issuer. Despite some successful research, however, an important related aspect not yet addressed has been the possibility of learning from natural-language commands. Such commands, generated by human users, contain valuable information. The inherent subjectivity of natural language, however, complicates potential learning from such commands and their interpretation. We propose decision making for robots operating under natural-language commands influenced by human aspects of decision making. Under our proposed concept, demonstrated in experiments conducted using a robotic manipulator, the robot is controlled using natural-language commands to conduct pick-and-place operations, during which the robot builds a knowledge base. After this learning, which uses a probabilistic neural network, the robot conducts similar tasks based on approximate decisions from the knowledge gained.