In human robot interaction any input that might help the robot to understand the human behaviour is valuable, and the eyes and their movement undoubtedly hold valuable information. In this paper we propose a novel algorithm for intention recognition using eye tracking in human robot collaboration. We first explore how the Cascade Effect hypothesis and a LSTM-based machine learning model perform to classify intent from gaze. Second, an algorithm is proposed, which can be used in a real time interaction to infer intention from the human user with a small uncertainty. A data collection with 30 participants was conducted in virtual reality to train and test the algorithm. The algorithm allows to detect the user intention up to two seconds before any user action with a success rate of up to 75%. These results open the possibility to study human robot interaction, where the robot can take the initiative based on the intention recognition.