A new approach for robotics perception, rooted in the sensorimotor paradigm, is proposed in this paper. Making systems able to autonomously adapt themselves to changes in their own body or in their environment is still a challenging question for many different scientific communities. Multiple works propose either sophisticated adaptive model-based or learning-based techniques as a solution. Recent contributions have shown that it is possible for an agent to discover the structure of its interaction with the environment or its own body via the so-called sensorimotor flow. The presented work is based on this idea, and a method for building an internal representation of sensorimotor interaction is proposed, which does not require any a priori knowledge or model. A careful mathematical formalization is outlined, together with simulations, demonstrating the effectiveness of the approach. Several cases are considered allowing for a general discussion. Moreover, plausibility of the internal sensorimotor representation is highlighted by showing that it is possible to consider motion planning directly from it.