In spoken communications, correction utterances, which are utterances correcting other participants’ utterances and behaviors, play crucial roles, and detecting them is one of the key issues. Previously, much work has been done on automatic detection of correction utterances in human-human and human-computer dialogs, but they mostly dealt with the correction of erroneous utterances. However, in many real situations, especially in communications between humans and mobile robots, the misunderstandings manifest themselves not only through utterances but also through physical actions performed by the participants. In this paper, we focus on action corrections and propose a classification of such utterances into Omission, Commission, and Degree corrections. We present the results of our analysis of correction utterances in dialogs between two humans who were engaging in a kind of on-line computer game, where one participant plays the role of the remote manager of a convenience store, and the other plays the role of a robot store clerk. We analyze the linguistic content, prosody as well as the timing of correction utterances and found that all features were significantly correlated with action corrections.