Abstract
Developing autonomous agents capable of sustaining goal-oriented behaviour in unconstrained dynamic environments requires the acquisition of skills sufficient for performing both perceptual conceptualisation and behavioural functionality. Such systems are expected to gain the cognitive categories hierarchically in order to make learning an incrementally open-ended process. Acquiring representations in the form of internal motor parameters enables the mechanism of learning to construct a set of hypotheses that directly refer to physical properties of the world. A collection of a priori categories that are known to be 'correct' must thus be innate to the system in order to constitute the initial representation of the environment. Learning then proceeds by the development of novel cognitive capabilities validated in terms of the a priori categories, thus, the system becomes capable of bootstrapping itself to further levels of grounded representation. We hence propose a bootstrapping approach that utilises a direct percept-action coupling that avoids the description process as an intermediate stage. Cognitive capabilities are represented as explicit models (i.e. having an explicit semantic meaning in contrast to implicit distributed knowledge at the sensorimotor level) that parametrically link elements of the percept-action domain. This leads to the development of compact symbolic representations of cognitive knowledge. The driving machinery of bootstrapping is thus attaining perceptual goals by action. Once a goal is detected as a salient perceptual state, the system has a 'desire' to learn a motor solution to satisfy the goal by means of its own action capabilities. Such a system configuration generates an incremental update of the perceptual domain in terms of the detected salient states and activates exploratory mechanisms in order to generalise the corresponding motor solutions. The hierarchical bootstrapping of the grounded cognitive concepts exhibits similarities to certain aspects of other learning techniques, such as Reinforcement Learning or SLAM Robotics. However, the proposed approach demonstrates a significant improvements of learning performance (namely, being characterised by a linear increase in computational requirements when learning in a typical unconstrained environment, as compared with the near-exponential increase for conventional percept-action learners).