Online kinematic calibration of robot manipulator based on neural network

The precision of a robot's positioning is paramount for its utility in industrial contexts. Enhancing this accuracy commonly involves robot calibration, primarily focused on pinpointing discrepancies in the robot's kinematic parameters. Nevertheless, given that a robot's kinematic model constitutes a highly nonlinear system characterized by non-Gaussian noise, the traditional calibration methods cannot meet the accuracy requirements of robots performing work under high load and high speed conditions for large load robots. To address this issue, this paper innovatively proposes a neural network-based approach. This method uses online learning and realtime update of DH parameters to estimate the most suitable Denavit-Hartenberg (DH) parameters according to the motion state of the robot arm. This method increases the flexibility of prediction parameters and eliminates the limitations and constraints of fixed parameters on the accuracy of the robot arm. Compared with the existing methods, the reliability and accuracy of positioning of manipulator can be improved by using the state of recent motion to adjust parameters in real time. This method uses the algorithm Long Short-Term Memory- Extended Kalman Filter (LSTM-EKF), combined with EKF and LSTM, to estimate the attitude based on laser tracker. Compared with traditional algorithms like EKF and Kalman Filter (KF), LSTM-EKF uses recurrent neural network to simulate noise, which can grasp the distribution of noise more accurately and estimate the attitude more accurately. In parameter estimation, we propose two networks to predict the compensation values of static parameters and dynamic parameters respectively. On the one hand, the use of neural network to predict parameters is more suitable for the constant update of DH parameters. On the other hand, neural network can more accurately fit the distribution of parameters and the internal relations before parameters. An empirical study conducted on the JR6210 industrial robot from Huashu Robotics demonstrates the superior performance of the proposed algorithm in solving the calibration problem of a large-load, high-velocity robot compared to conventional calibration algorithms.