Abstract: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 real-time 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.

Real-Time Robot End-Effector Pose Estimation with Deep Network

Dual Regression for Efficient Hand Pose Estimation

Real-time Holistic Robot Pose Estimation with Unknown States

An Attention-Based Deep Learning Approach for Inertial Motion Recognition and Estimation in Human-Robot Collaboration

CtRNet-X: Camera-to-Robot Pose Estimation in Real-world Conditions Using a Single Camera

Accurate Robot Arm Attitude Estimation Based on Multi-View Images and Super-Resolution Keypoint Detection Networks

Object Detection and Pose Estimation from RGB and Depth Data for Real-time, Adaptive Robotic Grasping

Learning Pose Estimation for High-Precision Robotic Assembly Using Simulated Depth Images

Pose Estimation of a Cobot Implemented on a Small AI-Powered Computing System and a Stereo Camera for Precision Evaluation

A vision-based end pose estimation method for excavator manipulator

Grasping Pose Estimation for Robots Based on Convolutional Neural Networks

Sparse Convolution Based 6D Pose Estimation for Robotic Bin-Picking with Point Clouds

Deep Vision Networks for Real-Time Robotic Grasp Detection

A Dexterous Hand-Arm Teleoperation System Based on Hand Pose Estimation and Active Vision

DeepPCO: End-to-End Point Cloud Odometry through Deep Parallel Neural Network

End-Effector Force and Joint Torque Estimation of a 7-DoF Robotic Manipulator Using Deep Learning

Online Kinematic Calibration of Robot Manipulator Based on Neural Network

Visual Localization of Robotic End Effector Via Fusion of 3D Gaussian Splatting and Heuristic Optimization Algorithm

Real-time Pose Estimation from Images for Multiple Humanoid Robots

Grasping pose estimation for SCARA robot based on deep learning of point cloud

A CNN Model for Real Time Hand Pose Estimation.