Multi-Layer Auto Resonance Network for Robotic Motion Control

Abstract

Recent developments in machine learning and artificial intelligence have evoked interest in many research areas considered as NP Hard. Humanoid motion is one such area. Controlling robots with large number of mechanical joints poses challenges due to non-linearity of displacement, redundant configurations, dynamic user environments, etc. Analytical and approximate solutions to Inverse kinematic problems of typical industrial robots with up to 6 Degrees of Freedom (DoF) have been presented in literature. Humans use more than a hundred joints for locomotion, each with one to three degrees of freedom increasing the complexity beyond comprehension. Therefore algorithmic and even heuristic approaches have not been successful in humanoid structures. Recently, machine learning and artificial intelligence are being used for robotic applications. This paper reports a new type of Artificial Neural Network (ANN) called Auto-Resonance Network (ARN) derived from synergistic control of biological joints. The network can be tuned to any real valued input without any degradation in learning ability. Due to the approximating nature of ARN, neuronal density of the network is low and grows at a linear or low order polynomial rate with input classification. Input coverage of the neuron can be tuned dynamically to match properties of input data. ARN can be used as a part of hierarchical structures to support deep learning applications. As a case study, the network has been used to generate optimal path for locomotion avoiding obstacles. The network can optimize the length of traverse and resolution during training much like a biological system. The network is able to learn without supervision taking cues from the environment.