Performance Evaluation of the Various Training Algorithms and Network Topologies in a Neural-Network-Based Inverse Kinematics Solution for Robots

Abstract

Recently, artificial neural networks have been used to solve the inverse kinematics problem of redundant robotic manipulators, where traditional solutions are inadequate. The training algorithm and network topology affect the performance of the neural network. There are several training algorithms used in the training of neural networks. In this study, the effect of various learning algorithms on the learning performance of the neural networks on the inverse kinematics model learning of a seven-joint redundant robotic manipulator is investigated. After the implementation of various training algorithms, the Levenberg-Marquardth (LM) algorithm is found to be significantly more efficient compared to other training algorithms. The effect of the various network types, activation functions and number of neurons in the hidden layer on the learning performance of the neural network is then investigated using the LM algorithm. Among different network topologies, the best results are obtained for the feedforward network model with logistic sigmoid-activation function (logsig) and 41 neurons in the hidden layer. The results are presented with graphics and tables.

Keywords

Robotics Neural Networks Training Algorithms Machine Learning Inverse Kinematics Solution

1. Introduction

The inverse kinematics problem is one of the most important problems in robotics. Fundamentally, it consists in finding the set of joint variables to reach a desired configuration of the tool frame. Computer based-robots are usually acting in the joint space, whereas objects are usually expressed in the Cartesian coordinate system. In order to control the position of the end-effector of a robotic manipulator, an inverse kinematics solution should be established to make the necessary conversions. This problem usually involves a set of nonlinear coupled, algebraic equations. The inverse kinematics problem is especially complex and time-consuming for redundant types of robotic manipulators [1,2].

In the literature, inverse kinematics solutions for robotic manipulators are drawn from various traditional methods such as algebraic methods, geometric methods and numerical methods. Neural networks have also become popular [3 –6].

Intelligent techniques have been one popular subject in recent years in robotics as one way to make control systems able to act more intelligently and with a high degree of autonomy. Artificial neural networks have been widely applied in robotics for the extreme flexibility that comes from their learning ability and function-approximation capability in nonlinear systems [2]. Many papers have been published about the neural-network-based inverse kinematics solution for robotic manipulators [7 –14]. Tejomurtula and Kak presented a study based on the solution of the inverse kinematics problem for a three-joint robotic manipulator using a structural neural network which can be trained quickly to reduce training time and to increase accuracy [7]. Xia et al. presented a paper about formulating the inverse kinematics problem as a time-varying quadratic optimization problem. For this purpose they suggested a new recurrent neural network. According to their studies, their suggested network structure is capable of asymptotic tracking for the motion control of redundant robotic manipulators [8]. Zhang et al. used Radial Bases Function networks (RBF) for the inverse kinematics solution of a MOTOMAN six-joint robotic manipulator. They used the solution to avoid complicated traditional procedures and programming to derive equations and programming [9]. An adaptive learning strategy based on using neural networks to control the motion of a six-joint robot was presented by Hasan et al. Their study was implemented without explicitly specifying the kinematic configuration or the working-space configuration [10]. Rezzoug and Gorce studied the prediction of finger posture by using artificial neural networks based on an error backpropagation algorithm. They obtained lower prediction errors compared to the other studies in the literature [11]. Chiddarwar et al. published a paper based on the comparison of radial-base functions and multilayer neural networks for the solution of the inverse kinematics problem for a six-joint serial robot model, using a fusion approach [12]. Zhang et al. presented a paper about the kinematic analysis of a novel 3-DOF actuation-redundant parallel manipulator using neural networks. They applied different intelligent techniques such as multilayer-perception neural network, Radial Bases Function neural network and Support Vector Machine to investigate the forward kinematic problem of the robot. They found that SVM gave better forward kinematic results than other applied methods [13]. Köker et al. presented a study based on the inverse kinematics solution of a Hitachi M6100 robot based on committee-machine neural networks. They showed that using a committee-machine neural network instead of a unique neural network increased the performance of the solution; in other words, the error was decreased [14].

A neural network's working principle is based on learning from previously obtained data known as a learning or training set, and then checking the system's success using test data. The learning algorithm also affects the success of the neural network implementation significantly. In this paper, the effect of various learning algorithms, network types, activation functions and numbers of neurons in the hidden layer have been examined for the inverse kinematics solution of a seven-joint robotic manipulator. The results show that the Levenberg-Marquardt training algorithm, feedforward networks, logsig activation function and 41 hidden-layer sizes gave the best results for the inverse kinematics solution of the used robot model.

In the paper, section 2 provides a kinematics analysis of a Schunk LWA3 robot [15], section 3 outlines the neural-network-based inverse kinematics solution, section 4 describes training and testing, section 5 gives results and a discussion, and section 6 provides conclusions.

2. Kinematic analysis of Schunk LWA3 robot

A manipulator is composed of serial links that are connected to each other with revolute or prismatic joints from the base frame through the end-effector. The calculation of the position and orientation of the end-effector in terms of the joint variables is known as forward kinematics. In order to obtain forward kinematics of a robot mechanism in a systematic manner, one should use a convenient kinematics model. The Denavit-Hartenberg method, which uses four parameters, is the most common method to describe robot kinematics [16 –18]. These four parameters can be described as a_i−1, α_i−1, d_i, and θ_i, which are the link length, link twist, link offset and joint angle, respectively. A coordinate frame is attached to each joint for the determination of DH parameters, and the z_i axis of the coordinate frame points along the rotary or sliding direction of the joints [19,20].

A seven-DOF Schunk LWA3 robot model has been used in this study, as shown in Figure 1. A 3-D view of the Schunk LWA3 robot is given in Figure 2. By using the above notations and Figure 1, the D-H parameters are obtained for this robot model as given in Table 1.

Figure 1.

The kinematic structure of Schunk LWA3 robot [15,21]

Figure 2.

3-D view of Schunk LWA3 robot [15,21]

Table 1.

D-H parameters of the robot

i	α_i−1(°)	d_i (mm)	θ_i (°)
1	−90	300	θ₁
2	90	0	θ₂
3	−90	328	θ₃
4	90	0	θ₄
5	−90	276,5	θ₅
6	90	0	θ₆
7	−90	171,7	θ₇

The general transformation matrix ⁱ⁻¹_iT for a single link can be obtained as follows:

\begin{matrix} {}_{i}^{i - 1}T = (\propto_{i - 1}) D_{x} (a_{i - 1}) R_{z} (θ_{i}) Q_{i} (d_{i}) \\ = [\begin{matrix} 1 & 0 & 0 & 0 \\ 0 & {c α}_{i - 1} & {- s α}_{i - 1} & 0 \\ 0 & {s α}_{i - 1} & {c α}_{i - 1} & 0 \\ 0 & 0 & 0 & 1 \end{matrix}] [\begin{matrix} 1 & 0 & 0 & a_{i - 1} \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{matrix}] \end{matrix}

(1)

x [\begin{matrix} c_{i} & - s_{i} & 0 & 0 \\ s_{i} & c_{i} & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{matrix}] [\begin{matrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & d_{i} \\ 0 & 0 & 0 & 1 \end{matrix}]

(2)

= [\begin{matrix} c_{i} & - s_{i} & 0 & a_{i - 1} \\ s_{i} {c α}_{i - 1} & c_{i} {c α}_{i - 1} & {- s α}_{i - 1} & {- s α}_{i - 1} d_{i} \\ s_{i} {s α}_{i - 1} & c_{i} {s α}_{i - 1} & {c α}_{i - 1} & {c α}_{i - 1} d_{i} \\ 0 & 0 & 0 & 1 \end{matrix}]

(3)

In (1), (2) and (3), R_x and R_z present rotation, D_x and Q_i denote translation, and c_i, s_i, cα_i−1, and sα_i−1 are the short-hands for cosθ_i, sinθ_i, cosα_i−1 and sinα_i−1, respectively.

For the Schunk TWA3 robot, it is straightforward to compute each of the link transformation matrices using (1), as follows.

{}_{1}^{0}{T =} [\begin{matrix} c_{1} & {- s}_{1} & 0 & 0 \\ 0 & 0 & 1 & 300 \\ {- s}_{1} & - c_{1} & 0 & 0 \\ 0 & 0 & 0 & 1 \end{matrix}],

(4)

{}_{2}^{1}{T =} [\begin{matrix} c_{2} & {- s}_{2} & 0 & 0 \\ 0 & 0 & - 1 & 0 \\ s_{2} & c_{2} & 0 & 0 \\ 0 & 0 & 0 & 1 \end{matrix}],

(5)

{}_{3}^{2}{T =} [\begin{matrix} c_{3} & - s_{3} & 0 & 0 \\ 0 & 0 & 1 & 328 \\ {- s}_{3} & - c_{3} & 0 & 0 \\ 0 & 0 & 0 & 1 \end{matrix}],

(6)

{}_{4}^{3}{T =} [\begin{matrix} c_{4} & - s_{4} & 0 & 0 \\ 0 & 0 & - 1 & 0 \\ s_{4} & c_{4} & 0 & 0 \\ 0 & 0 & 0 & 1 \end{matrix}],

(7)

{}_{5}^{4}{T =} [\begin{matrix} c_{5} & {- s}_{5} & 0 & 0 \\ 0 & 0 & 1 & 276,5 \\ {- s}_{5} & {- c}_{5} & 0 & 0 \\ 0 & 0 & 0 & 1 \end{matrix}],

(8)

{}_{6}^{5}{T =} [\begin{matrix} c_{6} & - s_{6} & 0 & 0 \\ 0 & 0 & - 1 & 0 \\ s_{6} & c_{6} & 0 & 0 \\ 0 & 0 & 0 & 1 \end{matrix}],

(9)

{}_{7}^{6}{T =} [\begin{matrix} c_{7} & - s_{7} & 0 & 0 \\ 0 & 0 & 1 & 171,7 \\ {- s}_{7} & {- c}_{7} & 0 & 0 \\ 0 & 0 & 0 & 1 \end{matrix}] .

(10)

The forward kinematics of the end-effector with respect to the base frame are determined by multiplying all of the ⁱ⁻¹_iT matrices as follows:

{}_{e}^{b}{T =} {}_{1}^{0}T {}_{2}^{1}T {}_{3}^{2}T \dots {}_{i - 1}^{i - 2}T {}_{i}^{i - 1}T

(11)

An alternative representation of _e^bT can be written thus:

{}_{e}^{b}{T =} [\begin{matrix} n_{x} & s_{x} & a_{x} & p_{x} \\ n_{y} & s_{y} & a_{y} & p_{y} \\ n_{z} & s_{z} & a_{z} & p_{z} \\ 0 & 0 & 0 & 1 \end{matrix}] .

(12)

In (12), n_x, n_y, n_z, s_x, s_y, s_z, a_x, a_y, a_z refer to the rotational elements of the transformation matrix. p_x, p_y, and p_z show the elements of the position vector.

For a seven-jointed manipulator, the position and orientation of the end-effector with respect to the base is given in (13).

{}_{7}^{0}{T =} {}_{1}^{0}T {}_{2}^{1}T {}_{3}^{2}{T {}_{4}^{3}T {}_{5}^{4}T {}_{6}^{5}T {}_{7}^{6}T} .

(13)

All obtained notations after the calculations based on (13), are given in the Appendix.

By using (15)–(26) (see Appendix), the training, validation and test sets are prepared for the inverse kinematics model learning. All of these algorithms are implemented in MATLAB.

3. Neural-network-based inverse kinematics solution

Neural networks are generally used in the modelling of nonlinear processes. An artificial neural network is a parallel-distributed information processing system. It stores the samples with distributed coding thus forming a trainable nonlinear system. Training of a neural network can be expressed as a mapping between any given input and output data set. Neural networks have some advantages, such as adoption, learning and generalization. Implementation of a neural-network model requires us to decide the structure of the model, the type of activation function and the learning algorithm [22, 23].

In Figure 3, the schematic representation of a neural-network-based inverse kinematics solution is given. The solution system is based on training a neural network to solve an inverse kinematics problem based on the prepared training data set using direct kinematics equations. In Figure 3, “e” refers to error – the neural network results will be an approximation, and there will be an acceptable error in the solution.

Figure 3.

Neural-network-based inverse kinematics solution system [14]

The designed neural-network topology is given in Figure 4. A feed-forward multilayer neural-network structure was designed including 12 inputs and seven outputs. Only one hidden layer was used during the studies.

Figure 4.

The neural network topology used in this study

The trained neural network can give the inverse kinematics solution quickly for any given Cartesian coordinate in a closed system such as a vision-based robot control system.

4. Training and testing

In this study, a fifth-order polynomial trajectory planning algorithm has been used to prepare data for training testing and validation sets of the neural network. The equation for fifth-order polynomial trajectory planning is given in (14):

\begin{matrix} θ_{i} (t) = θ_{i 0} + \frac{10}{t_{f}^{3}} (θ_{i f} - θ_{i 0}) t^{3} + \frac{15}{t_{f}^{4}} (θ_{i 0} - θ_{i f}) t^{4} + \frac{6}{t_{f}^{5}} (θ_{i f} - θ_{i 0}) t^{5} \\ i = 1,2, \dots, m \end{matrix}

(14)

where θ_i(t) denotes the angular position at time t, θ_if is the final position of the i^th joint, θ_i0 is the initial position of the i^th joint, m is the number of joints, and t_f is the arrival time from initial position to the target [24 –28].

Some starting and final angular positions are defined to produce data in the work volume of the robotic manipulator. A sample trajectory is given in Figure 5 between 0 and 90 degrees for a joint.

Figure 5.

A sample fifth-order polynomial trajectory

Twelve different training algorithms have been used in this study. The fundamentals of the 12 training algorithms are summarized in Table 2.

Table 2.

Different training algorithms used in this study

Algorithm	Purpose	Description
Levenberg–Marquardt
Trainlm (LM)	Levenberg-Marquardt backpropagation	trainlm is a network-training function that updates weight and bias values according to Levenberg-Marquardt optimization.
Conjugated gradient descent
Traincgb (CGB)	Conjugate gradient backpropagation with Powell-Beale restarts	traincgb is a network-training function that updates weight and bias values according to the conjugate gradient backpropagation with Powell-Beale restarts.
Traincgf (CGF)	Conjugate gradient backpropagation with Fletcher-Reeves updates	traincgf is a network-training function that updates weight and bias values according to conjugate gradient backpropagation with Fletcher-Reeves updates.
Trainscg (SCG)	Scaled conjugate gradient backpropagation	trainscg is a network-training function that updates weight and bias values according to the scaled conjugate gradient method.
Traincgp (CGP)	Conjugate gradient backpropagation with Polak-Ribiére updates	traincgp is a network-training function that updates weight and bias values according to conjugate gradient backpropagation with Polak-Ribiére updates.
Quasi-Newton algorithm
Trainoss (OSS)	One-step secant backpropagation	trainoss is a network-training function that updates weight and bias values according to the one-step secant method.
Trainbfg (BFG)	BFGS quasi-Newton backpropagation	trainbfg is a network-training function that updates weight and bias values according to the BFGS quasi-Newton method.
Resilient backpropagation
Trainrp (RP)	Resilient backpropagation	trainrp is a network-training function that updates weight and bias values according to the resilient backpropagation algorithm (Rprop).
Gradient descent with variable learning rate
Traingd (GD)	Gradient descent backpropagation	traingd is a network-training function that updates weight and bias values according to gradient descent.
Traingda (GDA)	Gradient descent with adaptive learning rate backpropagation	traingda is a network-training function that updates weight and bias values according to gradient descent with adaptive learning rate.
Traingdm (GDM)	Gradient descent with momentum backpropagation	traingdm is a network-training function that updates weight and bias values according to gradient descent with momentum.
Traingdx (GDX)	Gradient descent with momentum and adaptive learning rate backpropagation	traingdx is a network-training function that updates weight and bias values according to gradient descent momentum and an adaptive learning rate.

For the training 7000 data values corresponding to the (θ₁,θ₂,θ₃,θ₄,θ₅,θ₆,θ₇) joint angles according to the different (n_x, n_y, n_z, s_x, s_y, s_z, a_x, a_y, a_z, p_x, p_y, p_z) Cartesian coordinate parameters were generated by using the fifth-order polynomial trajectory planning given in (14) based on kinematic equations (15) to (26), given in the Appendix. A sample data set produced for the training of neural networks is given in Table 3. Here, the trajectory-planning algorithm has been used just to produce data between any given starting angular position and the final position, as mentioned above. Data preparation was done in a predefined area (robot working area). We tried to obtain well-structured learning sets to make the learning process successful and easy. These values were recorded in the files to form the training, validation and testing sets of the networks. Each 2380 of these data values were used in the training of neural networks, and 2310 were used in the validation of each neural network to see their success for the same data set. 2310 data values were used for the testing process as well. Training of the neural networks was completed when the error reached the possible minimum. The training and test results are given with details in the following section.

Table 3.

A sample data set produced for the training of neural networks

Inputs												Outputs
n _x	n _y	n _z	s _x	s _y	s _z	a _x	a _y	a _z	p_x(mm)	p_y(mm)	p_z (mm)	θ₁^o	θ₂^o	θ₃^o	θ₄^o	θ₅^o	θ₆^o	θ₇^o
−0.987	−0.001	0.161	−0.160	−0.086	−0.983	0.014	−0.996	0.085	−583.312	154.689	117.887	10	78.4	0	20	5	82	0
−0.872	−0.001	0.489	−0.487	−0.086	−0.869	0.043	−0.996	0.075	−507.814	154.689	310.282	30	78.4	0	20	5	82	0
−0.991	−0.004	0.133	−0.130	−0.169	−0.977	0.026	−0.986	0.167	−570.188	156.511	129.943	10	73.8	0	30	10	76	0
−0.958	−0.005	0.288	−0.283	−0.164	−0.945	0.052	−0.986	0.156	−524.624	156.377	220.915	20	69.2	0	40	10	71	0
−0.900	−0.006	0.436	−0.430	−0.158	−0.889	0.074	−0.987	0.140	−461.612	156.207	297.819	30	64.6	0	50	10	65	0
−0.950	−0.012	0.312	−0.279	−0.417	−0.865	0.141	−0.909	0.393	−531.876	170.088	278.655	20	78.4	5	20	20	82	0
−0.955	−0.050	0.293	−0.264	−0.308	−0.914	0.136	−0.950	0.281	−451.005	162.612	322.051	30	64.6	0	50	20	65	5
−0.991	−0.053	0.123	−0.080	−0.505	−0.859	0.108	−0.861	0.496	−516.006	178.544	246.109	15	66.9	5	45	28	68	0
−0.998	−0.050	0.033	−0.010	−0.415	−0.910	0.059	−0.908	0.414	−575.608	169.795	174.391	10	78.4	0	20	25	82	5
−0.917	−0.061	0.393	−0.316	−0.491	−0.812	0.242	−0.869	0.431	−469.332	176.847	378.486	30	78.4	5	20	25	82	5

5. Results and discussions

In this study, 12 training algorithms were used in the inverse kinematics model learning. The results of these algorithms are presented in Table 4. According to the table, the last five training methods – BFG, GDA, GDM, GD and GDX – are unsatisfactory. These are therefore neglected and not used in the graphical comparisons. The remaining seven training algorithms are compared.

Table 4.
Best performances for training, validation, and testing training times and number of epochs for different training algorithms

Algorithm Training time (sec.) Number of epochs Training Validation Test

Best-performance MSE (deg²) Best-performance MSE (deg²) Best-performance MSE (deg²)

LM 5433.6 1000 0.0016 0.0030 0.0026

CGB 217.723 967 0.1816 0.1855 0.1894

CGF 215.354 1000 0.1981 0.2243 0.2207

SCG 139.891 756 0.2296 0.2537 0.2551

CGP 182.368 827 0.2514 0.2587 0.2692

OSS 147.725 538 0.9826 1.0354 1.0529

RP 95.707 1000 1.444 1.5025 1.5352

BFG 351.21 1000 13.1157 14.1383 13.2108

GDA 14.65 150 34.7802 35.1373 36.7285

GDM 0.755 6 1265 1317.9 1294.8

GD 0.76 6 1569.7 1591.3 1584.8

GDX 1.288 6 6002.4 5983.3 6008.4

Algorithm	Training time (sec.)	Number of epochs	Training	Validation	Test
LM	5433.6	1000	0.0016	0.0030	0.0026
CGB	217.723	967	0.1816	0.1855	0.1894
CGF	215.354	1000	0.1981	0.2243	0.2207
SCG	139.891	756	0.2296	0.2537	0.2551
CGP	182.368	827	0.2514	0.2587	0.2692
OSS	147.725	538	0.9826	1.0354	1.0529
RP	95.707	1000	1.444	1.5025	1.5352
BFG	351.21	1000	13.1157	14.1383	13.2108
GDA	14.65	150	34.7802	35.1373	36.7285
GDM	0.755	6	1265	1317.9	1294.8
GD	0.76	6	1569.7	1591.3	1584.8
GDX	1.288	6	6002.4	5983.3	6008.4

Firstly, the comparisons are made by the number of epochs versus the mean squared error (MSE) values of the training, validation and testing. The graphical representations are given in Figures 6a, 6b and 6c, respectively. It is obvious from Figures 6a –c that the LM algorithm is the best one. The comparisons are then carried out by training time versus MSE values. These graphical comparisons are given for training, validation and test values in Figures 6d, 6e and 6f, respectively. The meanings of the lines used in these graphics are given in Figure 6g. Although 5433.6 seconds elapses for 1000 epochs during the training using LM, it is clearly seen in the three comparisons that the MSE value for the LM training algorithm is less than the MSE values of other algorithms around 65 seconds. According to these graphical representations, it is clear that the LM algorithm is a good and robust training algorithm for the inverse kinematics model learning of a robot. The MSE values of training, validation and testing according to the number of epochs for the LM algorithm is been given in the same graphic in Figure 7.

Figure 6a.

Comparison of training performances versus number of epochs for different training algorithms

Figure 6b.

Comparison of validation performances versus number of epochs for different training algorithms

Figure 6c.

Comparison of testing performances versus number of epochs for different training algorithms

Figure 6d.

Comparison of training performances versus time for different training algorithms

Figure 6e.

Comparison of validation performances versus time for different training algorithms

Figure 6f.

Comparison of testing performances versus time for different training algorithms

Figure 6g.

The meanings of the line styles in Figures 6a–f

Figure 7.

Training, validation and testing values for LM algorithm

Since the LM algorithm is found to be significantly efficient, it is chosen for the analysis of the effect of various network types, activation functions and number of neurons in the hidden layer.

Six different network types, including 40 neurons in the hidden layer, were designed and examined by using the LM training algorithm and hyperbolic tangent sigmoid-activation function (tansig). The best results were obtained from the feedforward network model, as seen in Table 5. After this, 13 different activation functions were examined in the feedforward neural network model, including 40 neurons in the hidden layer, using the LM training algorithm. According to the results presented in Table 6, the logsig activation function was better than the others. Finally, the effect of different number of neurons in the hidden layer was analysed by using the feedforward neural network model with logsig activation function using the LM training algorithm. As is evident in Table 7, 41 neurons in the hidden layer gave the best result.

Table 5.

Best performances for training, validation, and testing, training times and number of epochs for different network types

Network Type	Training time (sec.)	Number of epochs	Training	Validation	Test
Network Type	Training time (sec.)	Number of epochs	Best-performance MSE (deg²)	Best-performance MSE (deg²)	Best-performance MSE (deg²)
Feedforward	6112.4	1000	0.00090432	0.0013	0.0013
Fitnet	5433.6	1000	0.0016	0.003	0.0026
Cascade	6423.8	1000	0.0027	0.004	0.0039
Elman	2194.3	1000	0.0127	0.0134	0.0132
LVQnet	104.685	2	162.8645	171.5717	166.5339
Patternnet	1538	270	248.0215	245.6518	252.9521

Table 6.

Best performances for training, validation, and testing, training times and number of epochs for different activation functions

Activation Function	Training time (sec.)	Number of epochs	Training	Validation	Test
Activation Function	Training time (sec.)	Number of epochs	Best-performance MSE (deg²)	Best-performance MSE (deg²)	Best-performance MSE (deg²)
Logsig	4074.5	733	0.00084905	0.0012	0.0013
Tansig	6112.4	1000	0.00090432	0.0013	0.0013
Softmax	6897.2	801	0.0011	0.0015	0.0064
Radbas	2584.4	442	0.0061	0.0085	0.0087
Poslin	317.406	57	0.0228	0.0339	0.0337
Satlin	192.848	33	0.0324	0.0523	0.0445
Satlins	265.246	46	0.0425	0.0564	0.0567
Tribas	402.489	73	0.0735	0.1229	0.1097
Purelin	28.154	4	1.6975	1.6901	1.7922
Compet	10.513	1	42.4409	41.6729	42.0576
Hardlim	10.386	1	169.4213	163.6717	167.204
Hardlims	11.272	1	161.8098	168.5041	170.5411
Netinv	95.723	17	144.2543	301.7994	4093.2

Table 7.

Best performances for training, validation, and testing, training times and number of epochs for different number of neurons in the hidden layer

Hidden Layer Size	Training time (sec.)	Number of epochs	Training	Validation	Test
Hidden Layer Size	Training time (sec.)	Number of epochs	Best-performance MSE (deg²)	Best-performance MSE (deg²)	Best-performance MSE (deg²)
10	1200.6	1000	0.0209	0.0212	0.0217
15	1363	756	0.0051	0.0063	0.0064
20	2583.6	1000	0.0028	0.0031	0.0033
25	3300.9	1000	0.0029	0.0036	0.004
30	1146	290	0.0104	0.0134	0.0132
35	5038.6	1000	0.0013	0.0018	0.0019
36	4850.9	1000	0.0011	0.0015	0.0016
37	3053.3	617	0.0026	0.0041	0.0039
38	5268.5	1000	0.0011	0.0014	0.0015
39	5472.7	1000	0.001	0.0014	0.0027
40	4074.5	733	0.00084905	0.0012	0.0013
41	5907.7	1000	0.00067996	0.0008654	0.001
42	6138.3	1000	0.0011	0.0016	0.0016
43	6372	1000	0.00090911	0.0014	0.0106
44	6381.7	1000	0.0012	0.0017	0.0017
45	6317.8	1000	0.0014	0.0023	0.007
50	7102.1	1000	0.001	0.0015	0.0016

Singularities and uncertainties in the robotic arm configurations are one of the essential problems in kinematics robot control resulting from applying a robot model. A solution based on using artificial neural networks was proposed by Hasan et al. [29]. The main idea is based on using neural networks to learn the characteristics of the robot system rather than having to specify an accurate robotic system model. This means the inputs of the neural network will also have linear velocity with the Cartesian position and orientation information to overcome singularities and uncertainties. Since the main focus of this paper is to investigate the performance of the various training algorithms in the neural-network-based inverse kinematics solution, however, this solution is not applied here.

6. Conclusions

This study has presented a performance evaluation of the various training algorithms and network topologies in a neural-network-based inverse kinematics solution for a seven-DOF robot. Twelve different training algorithms were analysed for their performances in the inverse kinematics solution for robotic manipulators. The LM training algorithm was found to be significantly the most efficient. It was evident that the LM algorithm was the fastest-converging algorithm and performed with high accuracy compared to the other examined training algorithms. Additionally, 13 different activation functions were examined during the study. The LM training algorithm was the fastest-converging one since it reached the lowest MSE value, around 65 seconds. In conclusion, the feedforward neural network model consisting of 41 neurons in the hidden layer with logsig activation function using the LM algorithm is successful for the inverse kinematics solution for a seven-DOF robot. As a future study, these training algorithms with a neural network could be examined on a real robotic manipulator, including hardware and software implementation of the solution scheme. Linear velocity could also be used as input of the neural network together with Cartesian position and orientation information, to overcome singularities and uncertainties.

Footnotes

Appendix

(15)

\begin{array}{c} n_{x} {=c}_{1} c_{2} c_{3} c_{4} c_{5} c_{6} c_{7} {‐s}_{1} s_{3} c_{4} c_{5} c_{6} c_{7} {‐c}_{1} s_{2} s_{4} c_{5} c_{6} c_{7} {‐c}_{1} c_{2} s_{3} s_{5} c_{6} c_{7} {‐s}_{1} c_{3} s_{5} c_{6} c_{7} {‐c}_{1} c_{2} c_{3} s_{4} s_{6} c_{7} n_{x} \\ {=c}_{1} c_{2} c_{3} c_{4} c_{5} c_{6} c_{7} {‐s}_{1} s_{3} c_{4} c_{5} c_{6} c_{7} {‐c}_{1} s_{2} s_{4} c_{5} c_{6} c_{7} {‐c}_{1} c_{2} s_{3} s_{5} c_{6} c_{7} {‐s}_{1} c_{3} s_{5} c_{6} c_{7} {‐c}_{1} c_{2} c_{3} s_{4} s_{6} c_{7} \\ s_{1} s_{3} s_{4} s_{6} c_{7} {‐c}_{1} s_{2} c_{4} s_{6} c_{7} {‐c}_{1} c_{2} c_{3} c_{4} s_{5} s_{7} {+s}_{1} s_{3} c_{4} s_{5} s_{7} {+c}_{1} s_{2} s_{4} s_{5} s_{7} {‐c}_{1} c_{2} s_{3} c_{5} s_{7} {‐s}_{1} c_{3} c_{5} s_{7} \end{array}

(16)

n_{y} {=s}_{2} c_{3} c_{4} c_{5} c_{6} c_{7} {+c}_{2} s_{4} c_{5} c_{6} c_{7} {‐s}_{2} s_{3} s_{5} c_{6} c_{7} {‐s}_{2} c_{3} s_{4} s_{6} c_{7} {+c}_{2} c_{4} s_{6} c_{7} {‐s}_{2} c_{3} c_{4} s_{5} s_{7} {‐c}_{2} s_{4} s_{5} s_{7} {‐s}_{2} s_{3} c_{5} s_{7}

(17)

\begin{array}{c} n_{z} {=−s}_{1} c_{2} c_{3} c_{4} c_{5} c_{6} c_{7} {‐c}_{1} s_{3} c_{4} c_{5} c_{6} c_{7} {+s}_{1} s_{2} s_{4} c_{5} c_{6} c_{7} {+s}_{1} c_{2} s_{3} s_{5} c_{6} c_{7} {‐c}_{1} c_{3} s_{5} c_{6} c_{7} \\ s_{1} c_{2} c_{3} s_{4} s_{6} c_{7} {+c}_{1} s_{3} s_{4} s_{6} c_{7} {+s}_{1} s_{2} c_{4} s_{6} c_{7} {+s}_{1} c_{2} c_{3} c_{4} s_{5} s_{7} + \\ c_{1} s_{3} c_{4} s_{5} s_{7} {‐s}_{1} s_{2} s_{4} s_{5} s_{7} {+s}_{1} c_{2} s_{3} c_{5} s_{7} {‐c}_{1} c_{3} c_{5} s_{7} \end{array}

(18)

\begin{array}{c} s_{x} {=−c}_{1} c_{2} c_{3} c_{4} c_{5} c_{6} s_{7} {+s}_{1} s_{3} c_{4} c_{5} c_{6} s_{7} {+c}_{1} s_{2} s_{4} c_{5} c_{6} s_{7} {+c}_{1} c_{2} s_{3} s_{5} c_{6} s_{7} + \\ s_{1} c_{3} s_{5} c_{6} s_{7} {+c}_{1} c_{2} c_{3} s_{4} s_{6} s_{7} {‐s}_{1} s_{3} s_{4} s_{6} s_{7} {+c}_{1} s_{2} c_{4} s_{6} s_{7} {‐c}_{1} c_{2} c_{3} c_{4} s_{5} c_{7} + \\ s_{1} s_{3} c_{4} s_{5} c_{7} {+c}_{1} s_{2} s_{4} s_{5} c_{7} {‐c}_{1} c_{2} s_{3} c_{5} c_{7} {‐s}_{1} c_{3} c_{5} c_{7} \end{array}

(19)

s_{y} {=−s}_{2} c_{3} c_{4} c_{5} c_{6} s_{7} {‐c}_{2} s_{4} c_{5} c_{6} s_{7} {+s}_{2} s_{3} s_{5} c_{6} s_{7} {+s}_{2} c_{3} s_{4} s_{6} s_{7} {‐c}_{2} c_{4} s_{6} s_{7} {‐s}_{2} c_{3} c_{4} s_{5} c_{7} {‐c}_{2} s_{4} s_{5} c_{7} {‐s}_{2} s_{3} c_{5} c_{7}

(20)

\begin{array}{c} s_{z} {=s}_{1} c_{2} c_{3} c_{4} c_{5} c_{6} s_{7} {+c}_{1} s_{3} c_{4} c_{5} c_{6} s_{7} {‐s}_{1} s_{2} s_{4} c_{5} c_{6} s_{7} {‐s}_{1} c_{2} s_{3} s_{5} c_{6} s_{7} + \\ c_{1} c_{3} s_{5} c_{6} s_{7} {‐s}_{1} c_{2} c_{3} s_{4} s_{6} s_{7} {‐c}_{1} s_{3} s_{4} s_{6} s_{7} {‐s}_{1} s_{2} c_{4} s_{6} s_{7} {+s}_{1} c_{2} c_{3} c_{4} s_{5} c_{7} + \\ c_{1} s_{3} c_{4} s_{5} c_{7} {‐s}_{1} s_{2} s_{4} s_{5} c_{7} {+s}_{1} c_{2} s_{3} c_{5} c_{7} {‐c}_{1} c_{3} c_{5} c_{7} \end{array}

(21)

a_{x} {=−c}_{1} c_{2} c_{3} c_{4} c_{5} s_{6} {+s}_{1} s_{3} c_{4} c_{5} s_{6} {+c}_{1} s_{2} s_{4} c_{5} s_{6} {+c}_{1} c_{2} s_{3} s_{5} s_{6} {+s}_{1} c_{3} s_{5} s_{6} {‐c}_{1} c_{2} c_{3} s_{4} c_{6} {+s}_{1} s_{3} s_{4} c_{6} {‐c}_{1} s_{2} c_{4} c_{6}

(22)

a_{y} {=−s}_{2} c_{3} c_{4} c_{5} s_{6} {‐c}_{2} s_{4} c_{5} s_{6} {+s}_{2} s_{3} s_{5} s_{6} {‐s}_{2} c_{3} s_{4} c_{6} {+c}_{2} c_{4} c_{6}

(23)

a_{z} {=s}_{1} c_{2} c_{3} c_{4} c_{5} s_{6} {+c}_{1} s_{3} c_{4} c_{5} s_{6} {‐s}_{1} s_{2} s_{4} c_{5} s_{6} {‐s}_{1} c_{2} s_{3} s_{5} s_{6} {+c}_{1} c_{3} s_{5} s_{6} {+  s}_{1} c_{2} c_{3} s_{4} c_{6} {+c}_{1} s_{3} s_{4} c_{6} {+s}_{1} s_{2} c_{4} c_{6}

(24)

\begin{array}{c} p_{x} {=−171,7c}_{1} c_{2} c_{3} c_{4} c_{5} s_{6} {+171,7s}_{1} s_{3} c_{4} c_{5} s_{6} {+171,7c}_{1} s_{2} s_{4} c_{5} s_{6} {+171,7c}_{1} c_{2} s_{3} s_{5} s_{6} + \\ {171,7s}_{1} c_{3} s_{5} s_{6} {‐171,7c}_{1} c_{2} c_{3} s_{4} c_{6} {+171,7s}_{1} s_{3} s_{4} c_{6} {‐171,7c}_{1} s_{2} c_{4} c_{6} {‐276,5c}_{1} c_{2} c_{3} s_{4} \\ {+276,5s}_{1} s_{3} s_{4} {‐276,5c}_{1} s_{2} c_{4} {‐328c}_{1} s_{2} \end{array}

(25)

p_{y} {=−171,7s}_{2} c_{3} c_{4} c_{5} s_{6} {‐171,7c}_{2} s_{4} c_{5} s_{6} {+171,7s}_{2} s_{3} s_{5} s_{6} {‐171,7s}_{2} c_{3} s_{4} c_{6} {+171,7c}_{2} c_{4} c_{6} {‐276,5s}_{2} c_{3} s_{4} {+276,5c}_{2} c_{4} {+328c}_{2} +300

(26)

\begin{array}{c} p_{z} {=171,7s}_{1} c_{2} c_{3} c_{4} c_{5} s_{6} {+171,7c}_{1} s_{3} c_{4} c_{5} s_{6} {‐171,7s}_{1} s_{2} s_{4} c_{5} s_{6} {‐171,7s}_{1} c_{2} s_{3} s_{5} s_{6} {+171,7c}_{1} c_{3} s_{5} s_{6} + \\ {171,7s}_{1} c_{2} c_{3} s_{4} c_{6} {+171,7c}_{1} s_{3} s_{4} c_{6} {+171,7s}_{1} s_{2} c_{4} c_{6} {+276,5s}_{1} c_{2} c_{3} s_{4} {+276,5c}_{1} s_{3} s_{4} {+276,5s}_{1} s_{2} c_{4} {+328s}_{1} s_{2} \end{array}

References

Chapelle

Bidaud

(2004) Closed form solutions for inverse kinematics approximation of general 6R manipulators. Mechanism and Machine Theory 39, pp. 323–338.

Hasan

A.T.

Al-Assadi

H.M.A.A.

Ahmad

A.M.I.

(2011) Neural Networks' Based Inverse Kinematics Solution for Serial Robot Manipulators Passing Through Singularities, Artificial Neural Networks. Industrial and Control Engineering Applications, Prof. Suzuki

Kenji

(Ed.), ISBN: 978-953-307-220-3, InTech, DOI: 10.5772/14977.

Bingul

Ertunc

H.M.

Oysu

(2005) Applying neural network to inverse kinematics problem for 6R robot manipulator with offset wrist. In: Proceedings, 7th International Conference on Adaptive and Natural Computing Algorithms, Coimbra, Portugal, pp. 112–115.

Daunicht

(1991) Approximation of the inverse kinematics of an industrial robot by DEFAnet. In: Proceedings, IEEE International Joint Conference on Neural Networks, Singapore, pp. 531–538.

Mayorga

R.V.

Sanongboon

(2005) Inverse kinematics and geometrically bounded singularities prevention of redundant manipulators: an artificial neural network approach. Robotics and Autonomous Systems 53, pp. 164–176.

Martin

J.A.

Lope

J.D.

Santos

(2009) A method to learn the inverse kinematics of multi-link robots by evolving neuro-controllers. Neurocomputing 72, pp. 2806–2814.

Tejomurtula

Kak

(1999) Inverse kinematics in roboticsusingneuralnetworks. Inf Sci 116, pp. 147–164.

Xia

Wang

(2001) Neural Networks for Kinematic Control of Redundant Robot Manipulators. IEEE Trans. Syst. Man Cybern., Part B, Cybern. 31 (1), pp. 147–154.

Zhang

P.Y.

Lü

T.S.

Song

L.B.

(2005) RBF Networks-Based Inverse Kinematics of 6R Manipulator. Int. J. Adv. Manuf. Technol. 26: pp. 144–147.

10.

Hasan

A.T.

Hamouda

A.M.S.

Ismail

Al-Assadi

H.M.A.A.

(2006) An adaptive-learning algorithm to solve the inverse kinematics problem of a 6 D.O.F. serial robot manipulator. Advanced in Engineering Software 37, pp. 432–438.

11.

Rezzoug

Gorce

(2008) Prediction of Fingers Posture Using Artificial Neural Networks. Journal of Biomechanics 41, pp. 2743–2749.

12.

Chiddarwar

S.S.

Babu

N.R.

(2010) Comparison of RBF and MLP Neural Networks to Solve Inverse Kinematic Problem for 6R Serial Robot by a Fusion Approach. Engineering Applications of Artificial Intelligence 23, pp. 1083–1092.

13.

Zhang

Lei

(2011) Kinematic Analysis of a Novel 3-DOF Actuation Redundant Parallel Manipulator Using Artificial Intelligence Approach. Robotics and Computer-Integrated Manufacturing 27, pp. 157–163.

14.

Köker

Çakar

Sari

(2013) A neural-network committee machine approach to the inverse kinematics problem solution of robotic manipulators. Engineering with Computers, Springer-Verlag London, DOI 10.1007/s00366-013-0313-2.

15.

Schunk GmbH. [Online] http://www.schunk.com.

16.

Kelly

(1994) Essential Kinematics for Autonomous Vehicles. Carnegie Mellon University, The Robotics Institute.

17.

K.S.

Gonzalez

R.C.

Lee

C.S.G.

(1987) Robotics: Control, Sensing, Vision, and Intelligence. McGraw-Hill, New York.

18.

Lee

G.C.S.

(1982) Robot arm kinematics, dynamics, and control. IEEE Computer 15 (12), pp. 62–79.

19.

Kucuk

Bingul

(2006) Industrial Robotics: Theory, Modelling and Control. Cubero

Sam

(Ed.), ISBN: 3-86611-285-8, InTech.

20.

Craig

J.J.

(1989) Introduction to Robotics Mechanics and Control. Addison-Wesley Publishing Company, USA.

21.

Pluzhnikov

(2012) Motion Planning and Control of Robot Manipulators. Master's thesis, Norwegian University of Science and Technology, Department of Engineering Cybernetics.

22.

Haykin

(2009) Neural Networks and Learning Machines. Third ed., Pearson, New Jersey, USA.

23.

Kheirkhah

Azadeh

Saberi

Azaron

Shakouri

(2012) Improved estimation of electricity demand function by using of artificial neural network, principal component analysis and data envelopment analysis. Computers & Industrial Engineering 64, pp. 425–441.

24.

Gupta

Kamal

(1998) Practical Motion Planning in Robotics: Current Approaches and Future, John Wiley & Sons, Inc. New York, NY, USA.

25.

Ata

A.A.

(2007) Optimal Trajectory Planning of Manipulators: A Review. Journal of Engineering Science and Technology2 (1), pp. 32–54.

26.

Spong

M.S.

Vidyasagar

(1989) Robot Dynamics and Control. John Wiley and Sons, Inc, New York.

27.

Biagiotti

Melchiorri

(2008) Trajectory Planning for Automatic Machines and Robots, Springer, Berlin.

28.

Tonbul

L.S.

Saritas

(2003) Inverse Kinematics Calculations and Trajectory Planning of a 5 DOF Edubot Robot Manipulator. Journal of Faculty of Engineering and Architecture, Gazi University, 18 (1), pp. 145–167.

29.

Hasan

A.L.

Ismail

Hamouda

A.M.S.

Aris

Marhaban

M.H.

Al-Assadi

H.M.A.A.

(2010) Artificial Neural-Network-Based Kinematics Jacobian Solution For Serial Manipulator Passing Through Singular Configurations, Advanced in Engineering Software 41, pp. 359–367.