Inverse kinematics of a bilateral robotic human upper body model based on motion analysis data

Abstract

Accurately predicting the movements of the human upper body is an obstacle in simulating human movement. This paper describes the optimization and comparison of three inverse kinematic algorithms designed to predict the pose of a 25 degree of freedom robotic human upper body model (RHBM). Motion analysis data of 10 subjects performing 5 activities of daily living were used to evaluate the performance of each method. The first algorithm used a numerically optimized weighted-least-norm (WLN) solution. The second algorithm maximized the joint angle probability density, using the gradient projection method (GP). The third algorithm used a single layer artificial neural network (NN), trained by Levenberg-Marquart backpropagation using the motion analysis data. Error was evaluated using the root mean square of the difference between calculated and recorded joint angles. The robustness was then tested by progressively excluding subject data from the training set, re-training the algorithms, and evaluating the error for all subjects. The numerically optimized WLN solution showed the highest robustness, and the GP and NN solutions had greater accuracy for the data included in training and lower accuracy for the data excluded from training. The gradient projection method showed greater robustness than the artificial neural network, and has potential to be refined and combined with the weighted least norm solution to increase accuracy and robustness. Future work will investigate combined methods and the ability to predict motion of persons using prostheses.