A “Look-Backward-and-Forward” Adaptation Strategy of NN Model Parameters for Prediction of Motion Trajectory

Abstract

Prediction of human motion trajectory is crucial for safe human-robot collaboration (HRC). The existing prediction method based on the adaptive neural network (NN) model couples the parameter estimation error with the priori estimation error of trajectory. The increments of the parameter vector over time steps is unavailable. This causes an inaccurate assessment of the motion trajectory mean-square estimation error (MSEE) and the associated estimated value, which is a potential threat to safe HRC. In this work, we seek a “look-backward-and-forward” approach. That is, the estimation error (EE) of the parameter vector at a certain time step ago is firstly calculated by reversely using the offline trained NN model. Later, the estimated parameter vector at more recent time steps are computed recursively till the present time step. By doing this, the coupling of the MSEE of the parameter vector with the MSEE of the trajectory is cut off. And the effect of EE of the parameter vector’s increments to the EE of the motion trajectory’s diminishes in the finite time steps. Thus, more accurate predictions of motion trajectory and associated MSEE are achieved, which is important for the upcoming robot controller design. The experimental results on predicting a 3-D motion trajectory show the practical appeal of the proposed method.

Supported by National Key R&D Plan of China (2017YFB1301204), National Natural Science Foundation of China (51875554, 51705510), Zhejiang Key R&D Plan (2018C01086), Zhejiang Provincial Key Laboratory of Robotics and Intelligent Manufacturing Equipment Technology (2015E10011), Equipment R&D Fund (6140923010102), and Ningbo S&T Innovation Key Project (2018D10010).