Redundancy Reduction of Musculoskeletal Model for Robots with Group Sparse Neural Network

Abstract

Creating robots with similar behavior and appearance to humans has been a long-term dream for roboticists. However, the highly redundant characteristic of the muscle system brings tremendous difficulty for the control and manufacture of the musculoskeletal robot. In this paper, we propose a heuristic pruning-based recurrent neural network method with group sparse regularization to reduce redundancy of muscles in musculoskeletal model, which can maintain the motion accuracy of musculoskeletal robots without activating the redundant muscles. Experiments on an upper extremity musculoskeletal model validate the effectiveness of the proposed method.