In autonomous robotics the question of the representation of the world is of crucial importance for the realization of complex tasks. However, building such a representation is often rooted on human-crafted a priori about the world. But as complexity increases this idea is not adapted anymore: fully autonomous agents in the real world require generalized representations. These must be built from experience and possibly with minimal external assumptions. This context is perfectly suited to the approach of sensorimotor perception, where the agent has to interpret the effects of naive actions on its inputs that come from exteroceptive and proprioceptive sensors. By exploiting basic sensory invariants, we show that it is possible to project the highly dimensional motor configurations into an internal representation of the sensors’ configuration space initially unknown to the agent. This allows the agent to build an internal model of the sensitive configurations.