The analysis of embodied and situated agents form a dynamical system perspective is often limited to a geometrical and qualitative description. However, a quantitative analysis is necessary to achieve a deep understanding of cognitive facts. The field of embodied cognition is multifaceted, and the first part of this thesis is devoted to exploring the diverse meanings proposed in the existing literature. This is a preliminary fundamental step as the creation of synthetic models requires well-founded theoretical and foundational boundaries for operationalising the concept of embodied and situated cognition in a concrete neuro-robotic model. By accepting the dynamical system view the agent is conceived as highly integrated and strictly coupled with the surrounding environment. Therefore the antireductionist framework is followed during the analysis of such systems, using chaos theory to unveil global properties and information theory to describe the complex network of interactions among the heterogeneous sub-components. In the experimental section, several evolutionary robotics experiments are discussed. This class of adaptive systems is consistent with the proposed definition of embodied and situated cognition. In fact, such neuro-robotics platforms autonomously develop a solution to a problem exploiting the continuous sensorimotor interaction with the environment. The first experiment is a stress test for chaos theory, a mathematical framework that studies erratic behaviour in low-dimensional and deterministic dynamical systems. The recorded dataset consists of the robots’ position in the environment during the execution of the task. Subsequently, the time series is projected onto a multidimensional phase space in order to study the underlying dynamic using chaotic numerical descriptors. Finally, such measures are correlated and confronted with the robots’ behavioural strategy and the performance in novel and unpredictable environments. The second experiment explores the possible applications of information-theoretic measures for the analysis of embodied and situated systems. Data is recorded from perceptual and motor neurons while robots are executing a wall-following task and pairwise estimations of the mutual information and the transfer entropy are calculated in order to create an exhaustive map of the nonlinear interactions among variables. Results show that the set of information-theoretic employed in this study unveils characteristics of the agent-environemnt interaction and the functional neural structure. This work aims at testing the explanatory power and impotence of nonlinear time series analysis applied to observables recorded from neuro-robotics embodied and situated systems.

Document Type


Publication Date