Humanoid Motion Control using Auto Resonance Neural Network

Abstract

It has been proven difficult to control robots with many mechanical joints due to a variety of problems, including redundant configurations, non-linear displacement, dynamic user surroundings, etc. Iterative computations have been applied to address inverse kinematic problems of industrial robots with up to six Degrees of Freedom (DoF). With one to three degrees of freedom in each of more than hundred joints used by humans for locomotion, the complexity is unfathomable. Consequently, in humanoid structures, algorithmic and heuristic approaches have failed. Numerous research fields that were earlier thought to be challenging to solve with computers have seen interest piqued by recent advancements in artificial intelligence and machine learning. One such domain is humanoid motion. This paper presents a novel kind of Artificial Neural Network named as Auto Resonance Neural Network (ARN). ARN uses the pull-relax mechanism applied by biological systems to control musculoskeletal motion. A variety of functions can be employed to build the pull-relax model, contingent on variables such as range, necessary coverage, and tunability. When employing ARN for joint control, inverse kinematics or some other kind of repetitive solution is not required. Its application is not influenced by the DoF or joint count. The network can use the learning methods like reinforcement learning or supervised or unsupervised learning.