With the advent of new data collection techniques, there has been a growing interest in studying co-located groups of students using Multimodal Learning Analytics to automatically identify collaborative learning states. In this paper, we analyze a multimodal dataset (N=84) made of eye-tracking, physiological and motion sensing data. We leverage unsupervised machine learning algorithms to find (un)productive collaborative states. We found a three-states solution where different states (and transitions between them) were significantly correlated with task performance, collaboration quality and learning gains. We interpret these findings in light of collaborative learning theories and discuss their implications for studying groups of students using MMLA. [For the full proceedings, see ED599096.]