Trends in Dynamic and Embodied Cognition

The dynamical systems approach is becoming increasingly important in cognitive science and robotics. In this perspective an agent and its environment are modeled as two coupled dynamical systems and cognitive activity arises from the history of interactions between the agent’s internal dynamics, its body morphodynamical properties and its environment [1, 2]. Recent advances in this field were presented in a September 2006 workshop on Dynamic Brain Models organized by Stefano Nolfi in Roma. In this short letter, we give a highlight of the approaches presented there and also include additional work by other groups that could not attend the workshop. In their seminal work, Beer and colleagues carried out rigorous analysis of the performance, behavioral strategy, and psychophysics of a model agent capable of minimally cognitive behaviors [3]. Such behaviors include active object discrimination and hexapod walking. The language of dynamical systems significantly contributes to their understanding of the complex, dynamic agentenvironment interaction. The authors have consistently adopted continuoustime recurrent neural networks (CTRNN) as the controller of those robots. The emphasis on continuous dynamics is motivated by the fact that the dynamics of both nervous systems and the macroscopic physical world are continuous in nature [2]. Evolutionary Robotics [4] is an ideal framework to explore rich dynamics between robots and their environment. Researchers at the University of Sussex, EPFL, and CNR Roma investiagted the evolution of various types of dynamic neural networks as well as CTRNN embedded in autonomous robots [5, 6, 7, 8]. These dynamic networks have been widely applied to various behavioral and cognitive tasks including bipedal walking [9], multiple and sequential problem solving [7, 10, 11], obstacle avoidance [12, 13], predator-prey coevolution [14, 15, 16], and T-maze navigation [17, 18] to mention a few. Der and colleagues used a different approach where robots develop sensorymotor coordination by means of self-exploration [19, 20]. Robot congnition is defined as the ability to predict the future consequences of the action undertaken by the robot. As concrete examples, they demonstrated that wheeled and spherical robots can develop dynamic sensory-motor coordination from scratch to purposive actions by self-exploration. Their control systems are often implemented as small fully-connected neural networks with Hebbian learning. Pasemann and colleagues studied structure and function of a extremely small but effective neural controller of autonomous robots [27, 28, 29]. The development of minimal networks is based on their former intensive studies of the rich dynamics in a single neuron and a single loop network [30, 31]. They have shown