Neural Network-Based Joint Velocity Estimation Method for Improving Robot Control Performance

Abstract

Joint velocity estimation is one of the essential properties that implement for accurate robot motion control. Although conventional approaches such as numerical differentiation of position measurements and model-based observers exhibit feasible performance for velocity estimation, instability can be occurred because of phase lag or model inaccuracy. This study proposes a model-free approach that can estimate the velocity with less phase lag by batch training of a neural network with pre-collected encoder measurements. By learning a weighted moving average, the proposed method successfully estimates the velocity with less latency imposed by the noise attenuation compared to the conventional methods. Practical experiments with two robot platforms with high degrees of freedom are conducted to validate the effectiveness of the proposed method.