Rapid development of social robots stimulates active research in human motion modeling, interpretation and prediction, proactive collision avoidance, human-robot interaction and co-habitation in shared spaces. Modern approaches to this end require high quality datasets for training and evaluation. However, the majority of available datasets suffers from either inaccurate tracking data or unnatural, scripted behavior of the tracked people. This paper attempts to fill this gap by providing high quality tracking information from motion capture, eye-gaze trackers and on-board robot sensors in a semantically-rich environment. To induce natural behavior of the recorded participants, we utilise loosely scripted task assignment, which induces the participants navigate through the dynamic laboratory environment in a natural and purposeful way. The motion dataset, presented in this paper, sets a high quality standard, as the realistic and accurate data is enhanced with semantic information, enabling development of new algorithms which rely not only on the tracking information but also on contextual cues of the moving agents, static and dynamic environment.