In this chapter we develop methods to extract knowledge from data, in what we call “bottom-up” knowledge extraction. When we think about robot movements we usually think about optimized movements, that try to minimize the energy consumption, the time required for a task, or the stress (and consequent damage) on the motors. These optimizations can be achieved because the dynamical model of the robot allows to have a mathematical and physical description of these quantities to minimize. While this reasoning is true in industrial robotics, it becomes flawed when talking about surgical robotics. Indeed, in this scenario the goal is to perform a delicate procedure while maximizing the safety of the patient. The problem is that, while we are able to provide a mathematical description of the energy consumption and of the motor fatigue, there is not a way to define the ‘patient safety’ or the ‘after-intervention patient health’. Additionally, while there is an exhaustive literature describing all the phases of an intervention, surgeons learn to move purely by practicing, building dexterity from experience. For these reasons, we decided to use a learning approach to describe the movements, called ‘primitives’ or ‘surgemes’. In this way, we can learn the low-level movements that the robot should perform by imitating expert surgeons, trying to mimicking their movements. In particular, in Sect. 1 we present the framework we adopted to describe the kinematics, while in Sect. 2 we discuss the model used to extract these movements from data.