Abstract-The task of autonomous surface discernment by an AIBO robotic dog is addressed. Different surface textures (plywood board, thin foam, short carpet, shag carpet) as well as different inclines (0 and 10 degrees) are considered. Using a genetic algorithm, gaits are designed which allow the robot to traverse each of these surfaces in an (approximately) optimal fashion. Frequency domain analysis of actuator readings from individual leg joints is performed for data collected using each gait on each surface type. It is found that the spectral content of these signals is significantly dependent on the characteristics of both the gait in use and the surface being walked upon. Using tap-delay Adaline neural networks to integrate actuator readings from 15 independent joints into a set of models of different gait/surface experiences, an algorithm is designed which uses these experiences to yield high classification rates across surface transitions and with low latency.

Abstract-Inspired by examples of oscillatory circuits in biological brains, we explore a hypothesis that one role of dynamical neural networks observed in biological sensory systems is to amplify subtle differences in sensory data, which in turn simplifies the task of classifying external stimuli. The authors recently developed a method for classifying the surface walked upon by a quadruped robotic dog [9]. The method developed utilizes time series data from the dog's joint sensors (kinesthetic vector). Employing the same data, an experiment was set up to explore the above hypothesis, comparing the relative accuracy of classifying (discerning) the surface type experienced by the robot, both with and without the inclusion of a system of coupled nonlinear oscillators in the data processing stream. These experiments demonstrated a significant increase in classification rate (on average) when the sensory data was passed through a coupled oscillator system to precondition the signals prior to inputting to a PNN type neural network classifier, in comparison with the result obtained by feeding the data to the PNN without preconditioning. From an implementation point of view, it is significant that these results were obtained via a coupled oscillator whose inter-oscillator weights were randomly instantiated. Some of the results are provided in terms of the Lyapunov Exponent and the spectral radius of the inter-oscillator weight matrix.