Model-based and model-free collision detection and identification for a parallel Delta robot with uncertainties

Abstract

This paper presents two methods for human collision detection and identification for a parallel Delta robot considering uncertainties. In the model-based, a generalized momentum observer based on the explicit model of the Delta robot is used to estimate a proposed uncertainty model without collision. A linear regression algorithm is approached to recognize actuator hysteresis friction models and modeling error parameters, constructing the uncertainty model. The boundary and convergence of the identification parameters are guaranteed based on a least square method. Then, the estimated uncertainty is compared with the generalized momentum observer to detect and identify collision impacts as external torques. In the model-free, the neural network model for actuated torque prediction controlling the Delta robot is developed based on a model-driven and data-driven approach. The training procedure of the proposed network only uses no collision datasets. The model-free method uses a neural network to provide detection signals and external torques identified by comparing its conservative predictions to the measured torques when a collision occurs. In this work, only low-cost current sensors and low-resolution encoders are attached to the Delta robot system for generating the datasets. Finally, the identification capacity of the proposed methods is also compared with a force sensor to verify the effectiveness in real-time implementation.