This paper deals with the qualitative and robust modelling of the relative positioning of on-line handwritten strokes. We exploit the fuzzy approach to take the imprecision of such relations into account. We first transpose a well-formalized method which proved itself in the domain of image analysis to the on-line case; it aims at evaluating the relation “to be in a given direction” relatively to a reference. Our first contribution is a solution to deal with the particular nature of on-line strokes, which are constituted of non-connected points. Our second and main contribution is a method to learn automatically fuzzy relative position relationships. It aims at evaluating the relation “to be in a given position” relatively to a reference using jointly the direction and the distance. We test the impact of this new fuzzy positioning approach on one possible application: the recognition of handwritten graphic gestures, which requires spatial context information to be discriminated. Whereas the recognition rate is 52.95% without any spatial information, we obtain a maximum of 95.75% when we use learnt relative position relationships.