Robot Learning Rules of Games by Extraction of Intrinsic Properties

摘要

Abstract—A major open problem in human-robot interaction remains: how can robots learn from non-technical humans? Such learning requires that the robot can observe behavior and extract the sine qua non conditions for when particular actions can be produced. The observed behavior can be either the robots own explorative behavior, or the behavior of humans that it observes. In either case, the only additional information should be from the human, stating whether the observed behavior is legal or not. Such learning may mimic the way that infants learn, through interaction with their caregivers. In the current research we implement a learning capability based on these principals of extracting rules from observed behavior using ”Human-Robot ” interaction or ”Human-Human ” interaction. We test the system using three games: In the first, the robot must copy a pattern formed by the human; in the second the robot must perform the mirror action of the human. In the third game, the robot must learn the legal moves of Tic Tac Toe. Interestingly, while the robot can learn these rules, it does not necessarily learn the rules of strategy, which likely require additional learning mechanisms. Index Terms—learning machine; robotics; iCub; Reactable; human-robot interaction; human feedback I.