A formal approach to the emergence of spatial representations from sensorimotor inputs in robotics

Overview & Synopsis

In this thesis, we shall propose a formalism to develop the notion of sensorimotor spatial perception in a robotic context. Usually in classical approaches to robotics, the perception of space is given to the agent through predefined models of the world such as the agent's forward kinematics and the spatial positions of effectors and sensors. However, the awareness of space does not necessarly require to be a priori provided. As an example, in the sensorimotor contingency theory, developped by J. Kevin O'Regan (2001), it is supposed that knowledge of space can be obtained from the dependencies between sensory inputs and generated actions. In this work, we shall study how an embodied agent, situated in an unknown environment with very little a priori knowledge about its body or its sensors, can build a representation of its interaction with the physical space. First, we shall provide the agent with the minimum a priori knowledge required for interpretation of its sensorimotor flow, such that the approach is general enough and is valid for the majority of robotic agents. Then, it shall be demonstrated that, by following a « refinement process », the agent can exploit basic sensory invariants during successive environments to obtain a representation of its sensors distinguishable spatial configurations in space. However, the state of the environment being unknown to the agent, the sensory invariants can be seen as random variables, so that the formalism shall be extended to stochastic processes. Furthermore, in the probability theory context of the refinement process, the agent obtains an internal representation with a metrical structure based on the sensory invariance probabilities. Hereafter, it shall be demonstrated that under some topological assumptions on the motor space, this metrical internal representation allows planning and representation of sensors' continuous trajectories in space. Finally, by computing similarities between the internal representations obtained from the agent's different sensory streams, it shall be shown that the agent is able to build a representation of its sensors topographical structure, e.g. arrangement of the camera pixels, as well as to know when it interacts with its own body which should lead to the discovery of the self.