Transfer Learning-Based Artificial Neural Network for Forward Kinematic Estimation of 6-DOF Robot

Introduction

Generally, the joints of robots are connected in series and controlled by the actuators. The gripper, generally called the end effector, defines the position of the robotic arm in the three-dimensional coordinate system. Various mathematical functions are formulated to establish the relationship between joint angles and their target positions. The problem of estimation for robotic models can be widely categorized as 1) forward kinematics and 2) inverse kinematics. The kinematic equation used to compute the position of the end-effector with known specified joint angles is called forward kinematics (Saha, 2014). Similarly, calculating the joint angles from the known position of the end-effector is referred to as inverse kinematics.

In order to establish the relation between these angles and position information, transformation matrices are used. The transformation matrices define rotational and translational relations whose generalization cannot be made prior to all types of robots. As each robot has its own joint spaces and constraints, the derivation of the mathematical model is necessary to establish the relation between the joint angles and the position. Fuzzy-based strategies have been proven efficient for such estimation (Jamwal et al., 2010), but cannot be generalized for all robots. While, such relationships can be learned with the help of Machine Learning (ML) models, which may solve the hindered generalization problem (Craig, 2009). Several such ML approaches have been proposed in the past decade, addressing the generalization of robotic manipulators to predict the position of robots.

Applications of Artificial Neural Network for forward kinematics estimation problems are prominent in the literature. Several such approaches are established for various robotic setups such as HEXA Parallel Robot (Dehghani et al., 2008), parallel manipulators (Parikh et al., 2009), 3D cable robot (Ghasemi et al., 2010), 7-DOF Sawyer Robotic Arm (Theofanidis et al., 2018), Delta parallel robot (Liu et al., 2019) model, and so on. Further, improvement was observed with integration of PSO to NN for a 6-DOF parallel robot (Li et al., 2007) to fine-tune backpropagation learning. Also, an iterative approach to the NN strategy proved to be efficient for parallel manipulators (Parikh and Lam, 2009). Another hybrid approach combining neural networks with interval arithmetic to solve forward kinematics (Schmidt et al., 2014) is being experimented for Cable-Driven Parallel Robots. The improved Newton iterative method is validated on a 3-DOF parallel manipulator (Wu and Xie, 2019) for the forward kinematics estimation problem. Other than machine learning, optimization approaches have also been found helpful in estimating position information for 2-DOF using genetic programming (Arellano and Rivera, 2019). Also, the Artificial Neural Network (ANN) for a 6-DOF robot has been experimented with forward kinematics for different pose and joint space representations (Grassmann and Burgner, 2019). Optimum values of ANN model parameters like input data, sample size, training algorithm etc. for inverse kinematics solution of a 3R robotic manipulator are also investigated (Karkalos et al., 2017). Fuzzy control-based path planning of industrial robots is analyzed, and it has been found that Gaussian membership function gives a better result compared to other membership functions (Sahu and Choudhury, 2018).