For humans, gestures are a means of commu- nication. In order to create a more spontaneous interaction between humans and robots, social robots should be able to understand the information we convey with gestures. To this aim, (i) we collected a dataset with 1469 examples of twelve common Italian hand-gestures using a custom-made inertial glove, via experiments organized as human-robot interactions, and (ii) we propose an offline gesture recognition model based on a Long- Short Term Memory (LSTM) Recurrent Neural Network (RNN), which achieved an overall accuracy equal to 87.0 ± 3.7%.