Internal forward models with efference copies for state estimations in adaptive hexapod locomotion

Abstract

Event Abstract Back to Event Internal forward models with efference copies for state estimations in adaptive hexapod locomotion Poramate Manoonpong1*, Ulrich Parlitz2 and Florentin Wörgötter1 1 Georg-August-Universität Göttingen, Bernstein Center for Computational Neuroscience, Germany 2 Institute for Nonlinear Dynamics, Georg-August-Universität Göttingen, Germany Animal locomotion mechanisms seem to largely rely not only on central pattern generators (CPGs) and sensory feedback but also on internal forward models [1]. These components are used in different degree in different animals. In general, CPGs generate basic rhythmic patterns which can be considered as motor commands. They are usually shaped by sensory feedback (like, proprioceptive signals) to achieve appropriate coordinated movements during walking. Besides, the internal models transform motor commands copied (efference copies) within the central nerve system (CNS) into expected sensory inputs (or sensory prediction). This allows to compare the expected ones to the actual incoming sensory signals for state estimations. As a consequence, the internal models with efference copies together with local leg control [2] allow animals to adapt their locomotion to deal with environmental changes or during traversing over difficult terrain. Based on these biological findings, we extend our existing neural CPG-based locomotion control of the hexapod robot AMOS [3] by introducing six internal forward models here. Each of them serves to transform a motor signal generated by a CPG (efference copy) into an expected sensory signal (i.e., foot contact signal) of each leg. Utilizing the discrete-time dynamical properties of a recurrent neuron (e.g., hysteresis effect) for signal transformation, each forward model is configured as a single recurrent neuron [4] with synaptic plasticity. Gradient descent learning is applied to adapt its presynaptic and recurrent weights. This way it can learn online to correctly transform the motor signal into the expected foot contact signal of each leg during normal walking (e.g., walking on flat surface). After learning the outputs of the learned models (i.e., expected foot contact signals) are used to compare with the actual incoming foot contact signals for the estimation of its walking state. The differences between the expected and actual sensor values are used to adapt leg motion (e.g., extension or elevation) through local leg control mechanisms. As a result, this neural closed-loop controller with the neural forward models enables the robot to perform a variety of locomotion behavior including insect-like walking and climbing. In addition, it allows the robot at the same time to adapt its locomotion to deal with terrain changes, losing of ground contact during a stance phase, or hitting obstacles during a swing phase (see supplementary video at http://www.manoonpong.com/BCCN2012/InternalForwardModels.wmv). We believe that all these biologically inspired components (i.e., CPG, internal forward models, and local leg control mechanisms) are at least important parts for developing robust and adaptable walking robots. Acknowledgements This research was supported by Emmy Noether grant MA4464/3-1 of the Deutsche Forschungsgemeinschaft (DFG), the Bernstein Center for Computational Neuroscience II Göttingen (BCCN grant 01GQ1005A, project D1).