Abstract
During human-robot interaction, errors will occur. Hence, understanding the effects of interaction errors and especially the effect of prior knowledge on robot learning performance is relevant to develop appropriate approaches for learning under natural interaction conditions, since future robots will continue to learn based on what they have already learned. In this study, we investigated interaction errors that occurred under two learning conditions, i.e., in the case that the robot learned without prior knowledge (cold-start learning) and in the case that the robot had prior knowledge (warm-start learning). In our human-robot interaction scenario, the robot learns to assign the correct action to a current human intention (gesture). Gestures were not predefined but the robot had to learn their meaning. We used a contextual-bandit approach to maximize the expected payoff by updating (a) the current human intention (gesture) and (b) the current human intrinsic feedback after each action selection of the robot. As an intrinsic evaluation of the robot behavior we used the error-related potential (ErrP) in the human electroencephalogram as reinforcement signal. Either gesture errors (human intentions) can be misinterpreted by incorrectly captured gestures or errors in the ErrP classification (human feedback) can occur. We investigated these two types of interaction errors and their effects on the learning process. Our results show that learning and its online adaptation was successful under both learning conditions (except for one subject in cold-start learning). Furthermore, warm-start learning achieved faster convergence, while cold-start learning was less affected by online changes in the current context.