Learning agent’s spatial configuration from sensorimotor invariants

Abstract

The design of robotic systems is largely dictated by our purely human intuition about how we perceive the world. This intuition has been proven incorrect with regard to a number of critical issues, such as visual change blindness. In order to develop truly autonomous robots, we must step away from this intuition and let robotic agents develop their own way of perceiving. The robot should start from scratch and gradually develop perceptual notions, under no prior assumptions, exclusively by looking into its sensorimotor experience and identifying repetitive patterns and invariants. One of the most fundamental perceptual notions, space, cannot be an exception to this requirement. In this paper we look into the prerequisites for the emergence of simplified spatial notions on the basis of a robot's sensorimotor flow. We show that the notion of space as environmentindependent cannot be deduced solely from exteroceptive information, which is highly variable and is mainly determined by the contents of the environment. The environment-independent definition of space can be approached by looking into the functions that link the motor commands to changes in exteroceptive inputs. In a sufficiently rich environment, the kernels of these functions correspond uniquely to the spatial configuration of the agent's exteroceptors. We simulate a redundant robotic arm with a retina installed at its end-point and show how this agent can learn the configuration space of its retina. The resulting manifold has the topology of the Cartesian product of a plane and a circle, and corresponds to the planar position and orientation of the retina.

Autonomous grounding of visual field experience through sensorimotor prediction

“…Note that no superscript a is needed as all receptive fields are moved together during saccades. Moreover motors could potentially be redundant regarding the sensor's displacements in the world, in which case the agent would have to discover its actual working space (see [13] for an example of such a sensorimotor structuring). The naive agents we consider have no specific policy to explore the world.…”

Une brève introduction à la perception sensori-motrice en robotique

“…L'utilisation des courbes noyau pour la construction d'une carte interne de l'espace a été évaluée dans (Laflaquière, O'Regan, Argentieri, Gas & Terekov, 2015) lors d'une approche purement expérimentale. Dans ces travaux, un agent redondant équipé d'un bras et d'un capteur analogue à une caméra est utilisé pour découvrir l'ensemble des courbes noyaux et obtenir une carte des configurations atteignables par son capteur.…”

“…L'existence d'un isomorphisme entre l'espace quotient ℳ/≡ A et l'espace de travail < prouve la pertinence de son utilisation dans les travaux précédents (Laflaquière, 2013 ;Laflaquière, O'Regan, Argentieri, Gas & Terekov, 2015 ;Marcel, Garcia, Argentieri & Gas, 2015) pour la découverte de la notion d'espace. Différentes stratégies ont d'ailleurs été proposées pour sa construction dans les travaux précédents, que ce soit pour la construction d'une représentation de l'espace de travail <, ou d'une sous-partie de celui-ci (le corps de l'agent).…”

Where Do I Move My Sensors? Emergence of a Topological Representation of Sensors Poses From the Sensorimotor Flow

“…Working on the agent's body was initially envisaged as a way to put the environment dependency of the representation aside. Indeed, as formalized later in this paper, the fact that the environmental state can possibly evolve along exploration has already proven to be a major theoretical difficulty [17], [21]. This paper proposes to tackle the environment dependency by generalizing the formalism initially proposed in [19].…”

“…This initial formalization led to the demonstration that an agent can, without any a priori, infer the so-called "dimension of space". Further works like [17] have also shown that, beyond the dimension of space, it is possible to build a motor internal representation of the positions occupied by the agent's end-effector without external knowledge about its working space. Despite the use of a curvilinear component analysis (CCA) [18] and the definition of adapted Hausdorff distances in the agent motor space, this work lacks a proper mathematical formalization and they are no clear definitions on the properties or spaces that are actually captured by the agent.…”