The 5th Annual IEEE International Conference on

Cyber Technology in Automation, Control and Intelligent Systems

June 8-12, 2015, Shenyang, China

RBF Neural Network Compensation Based Trajectory

Tracking Control for Rehabilitation Training Robot

Gui Yin,Xiaodong Zhang*,Jiangcheng Chen, Qiangyong Shi

(Key Laboratory of Education Ministry for Modern Design and Rotor-Bearing System, Xi’an Jiaotong University, Xi’an, China)
*[email protected]

Abstract—For rehabilitation robot, how to effectively control impedance, the ability of LOPES rigid demeanor changed.
the movement of training, which depends on the performance of Because of the above control methods mostly ignored robot
the robot control system, is very important to improve the quality dynamics model uncertainty and the environment disturbance,
of rehabilitation. Lower limb rehabilitation robot system is a which affects the desired joint angular displacement and torque
nonlinear time-varying system, so the real-time calculation and tracking capabilities.
compensation for the nonlinear coupling term is always
neccessary, and linear control method can be used to achieve There are many typical methods are used: feed forward
trajectory tracking with a high precision. For this target, a radial compensation control; feedback control based on computed
basis function (RBF) neural network compensation control torque model, sliding control based on computed torque model,
method based on computed-torque is put forward. First, the as well as control method combining adaptive control and
controlled object and movement features of the rehabilitation intelligent control. Computed torque control method is based
robot system are briefly introduced.Then, computed torque on a real-time dynamics of compensation system model, the
control method is analyzed,and for the uncertainty part of computed torque control method is better than the line feed
computed torque as well as the environment disturbance,the RBF forward compensation method for non-linear feedback control
neural network compensator is designed .Finally, the simulation of compensation[7]. Sliding mode control using switch
for the proposed algorithm is conducted and analyzed. The estimated reference model, it’s hard to decide the model
results show that the computed torque controller with RBF
parameters to determine the upper and lower bounds and it’s
neural network compensator has smaller tracking error than the
also hard to ensure the existence of a sliding model, and the
PD feedback controller based on computed torque.
sliding mode control in the form of switch easily prone to
Keywords—lower limb rehabilitation robot; RBF neural shake under load changes [8].
network compensation; gait trajectory tracking control; PD Robot control method is closely related to the system
computed torque control dynamics model. For an accurate model of the robot system, in
I. INTRODUCTION theory, computed torque control can guarantee asymptotic
convergence of tracking. In fact, model of robotic system exists
Robot assisted patients with gait training has become the error. In order to compensate modeling error to improve the
main rehabilitation treatment for those who suffered from trajectory tracking ability, compensation algorithm to
hemiplegia, spinal cord injury , stroke and other neurological calculated torque is another focus of research.
diseases. The Robot assisted with gait training can increases
the intensity of training, extends the training time and reduces From the literature [9-11], we know that, for complex,
the workload of therapist. But how to effectively control the nonlinear, uncertain robotic systems, the neural network has
movement of training to improve the quality of rehabilitation the following advantages: Ԙ ability to approximate any
will be the main research work on rehabilitation robot. nonlinear system; ԙ the control system does not require
primary knowledge of controlled object ; Ԛself-learning and
At present, domestic and foreign scholars study on how to
effectively control the robot which assisted gait training, adaptive capacity and other characteristics. So the neural
besides, they carry out exploring research on the walking network control has been widely used to the robot controlling.
motion control strategy, and gain some preliminary conclusions. Roselito [12] proposed a neural network computed torque
Jezernik and Rienern [1,2] design Proportional - Differential control to accomplish position control on PUMA560 and
(proportion-differential, PD) feedback position control method explained the definition of the model learning. The method is
for Lokmat robot in the active training mode. Based on human- proved superior to the classical PD controller through
machine cooperation of active training ideas, force/position simulation. S.Hu [13] and other scholars designed neural
hybrid adaptive control methods has been applied. Agrawal [3, 4] network controller to control the robot with constrained
use PD feedback tracking control for the BGO and proposed problems in the operating space. The proposed control scheme
control algorithm based on impedance control for active can adapt to the variation of operating conditions and kinetic
training. Edwin and Veneman [5,6] and other scholars who parameters of the robot. Khemaissia [14] and other scholars
studies impedance control proposed collaboration (assist as introduced the control identification and adaptive control
needed) control algorithm based on adaptive. By adjusting the scheme based on the nonlinear functions approximation
performance of neural network to the robot control. Gang Feng

978-1-4799-8730-6/15/$31.00 © 2015 IEEE 359

[15]
study on neural network adaptive control method to mostly ignored robot dynamics model uncertainty and the
compensate dynamics model uncertainty for robot, and focuses environment disturbance, which affects the desired joint
on the neural network loop system stability and convergence angular displacement and torque tracking capabilities.So PD
condition of system errors. feedback controller based on computed torque model was
applied in hip’s and knee’s flexion/extension DOF controlling.
In order to improve the trajectory tracking performance For the uncertainty part of computed torque and the
and enhance the robustness of the system without the precise environment disturbance, the radial basis function (RBF)
dynamic structure and parameters. In addition, considering the neural network compensation method can be used to eliminate
coupling problem of rehabilitation robot system, the radial the uncertainty and the disturbance in controlling.
basis function (RBF) neural network based compensation
method to the uncertainty part of computed torque is proposed B. Exoskeleton manipulator dynamics of rehabilitation robot
in this paper. The validity of the method is certified by For rehabilitation robot, there is only 2 DOF of exoskeleton
simulation. in the sagittal plane. The rehabilitation robot manipulator can
II. COMPUTED TORQUE CONTROL be modeled as a set of 2 rigid bodies connected in a serial chain
with torques/forces acting at the hip and knee joints. By
A. Controlled object and movement features of the applying the Lagrange equation of motion , the rehabilitation
exoskeleton system robot dynamic equation is derived as follows:
The design of rehabilitation robot’s powered exoskeleton
specific structure composition was shown in Figure 1, the
M (θ )θ + C (θ , θ)θ + G (θ ) = T (1)
powered exoskeleton is fixed on the pelvic motion that is fixed
where θ is the 2x 1 vector of joint positions, T is the 2x 1
on the new BWS system. Each leg has three degrees of
freedom. they are hip’s and knee’s flexion / extension DOF in vector of applied joint forces/torques, M (θ ) is the 2x 2
the sagittal plane, and hip’s abduction /adduction DOF in the symmetric Positive definite manipulator inertia matrix,
coronal plane . As described above, the number of degrees of C(θ , θ)θ is the 2x 2 vector of centripetal and Coriolis torques,
freedom that the powered exoskeleton owns meets all basis
conditions when body is doing a natural gait . The main active G (θ ) is the 2x 1 vector of gravitational torques.
movement joint is flexion / extension movement of hip joint
C. PD feedback controller based on Computed torque
and knee joint in the sagittal plane. The driven program is
electric motor pass through crank block structure to make The computed-torque control approach involves
joints doing rotations. Hip’s abduction / adduction movement incorporating the robot dynamic model into the robot controller
is passive movement joint. In the early rehabilitation, the to transform the highly nonlinear robot dynamics equation (1)
patients’ lower extremity muscle strength is on low level that it into equivalent linear systems. Linear control theory can then
cannot control abduction or adduction of thigh. As a result, the be applied to synthesize controllers to specify the desired
powered exoskeleton’s abduction / adduction DOF will be closed-loop response. In the case of knowing joint position
locked to cause exoskeleton only working in the sagittal plane. trajectory θ and the angular velocity θ ; θd is substituted into
Thus, only hip’s and knee’s flexion / extension DOF need
the Formula (1) as a control system input variable , we have:
controlling.
M (θ )θd + C (θ , θ)θ + G (θ ) = T (2)

From the formula (1) and (2), eliminated the nonlinear term,
we can get decoupled linear systems based on joint variables :

θ = θd ҏҞ (3)

For above linear time-invariant systems, using PD control
servo compensation to reduce the error, servo law:
u = θd + K D e + K P e
(4)
e = θ d − θ ; e = θd − θ (5)

θd is the ideal trajectory acceleration; θ d is ideal trajectory

angle; K D is positive definite derivative matrix and K P is
proportional gain matrix. Thus, the control law of PD feedback
Fig. 1. Schematic view of powered exoskeleton controller based on Computed torque is:
Since two legs of walker robot can be regarded as two
independent control body, so each leg of the walker robot can T = M (θ )(θd + K D e + K p e) + C (θ , θ)θ + G (θ ) (6)
consider as the control unit except that they are 180 degrees
different in phase.The introduced control methods above As is frequently encountered in practice, the performance
of this controller is highly sensitive to the accuracy of the robot

360
model.However, the actual system modeling exists errors, as The figure 2 shows that the control input of robot system
well as the varying load and uncertainties of the disturbance can be written as the following:
outside factors, the robot system dynamic equation can be
modified as follows: T = TCT + TNNC (9)

Mˆ (θ )θ + Cˆ (θ , θ)θ + Gˆ (θ ) = T + f (7)

Where, TCT is output of computed torque controller,
TNNT is the RBF neural network compensation control torque,
where is Mˆ (θ ) the ideal symmetric Positive definite and it’s defined as:
manipulator inertia matrix, Cˆ (θ , θ) is the ideal centripetal and
ˆ (θ ) is the ideal gravitational torques.T is TNNC = − M (θ ) lf (10)
Coriolis torques, G
∧
the 2x 1 vector of applied joint forces/torques, f is the
Among them, f is an estimate of uncertainty for the
modeling errors and uncertainties of the disturbance outside.
walker leg system, and it is defined as
From formula (6)and (7), the following can be obtained:
lf = W
lϕ ( X ) (11)
e + K D e + K p e
ϕ ( x) = exp( X − ci / σ i2 )
2
= [ M (θ )]−1 × [ΔM (θ )θ + (8) Where, i = 1, 2" , n

ΔC (θ , θ)θ + ΔG (θ ) − f ] ϕ ( x) = [ϕ1 , ϕ2 "ϕn ] is the output of the Gaussian function,

∧
W = [W1 ,W2 "Wn ] is the estimated value of W * which is
T
Where , ΔC (θ , θ) = Cˆ (θ , θ) − C (θ , θ)
optimum power Value for RBF neural network , ci is the
ΔM (θ ) = Mˆ (θ ) − M (θ ) , ΔG (θ ) = Gˆ (θ ) − G (θ )
center of the vector for Gaussian basis functions , σ i is width
Formula (8) shows that the system model is affected by the shape for Gaussian basis functions DŽ
uncertainty factor, which degrades the performance of this
control method. The motivation for this work is to design a B. Design of RBF neural network compensation controller
RBF neural network controller to compensate for the model From PD feedback controller based on computed-
mismatch. torque , TCT can be written as the following:
III. RBF NEURAL NETWORK COMPENSATOR DESIGN
TCT = M (θ )(θd + K D e + K P e) + C (θ , θ)θ + G (θ ) (12)
A. Structure of RBF neural network compensation control
scheme From formula (2), (9), (10), (11) and (12) the following
RBF neural network compensation control scheme based can be obtained:
on the calculation of the torque was shown in Figure 2.The
controller consists of a normal PD feed forward computed- e + K D e + K p e

torque controller, which makes full use of all known dynamics
of the robot, as well as an additional RBF neural network = [ M (θ )]−1 × [ΔM (θ )θ + (13)
compensator, which is embedded in the forward path of the
system to account for the remaining unknown parameters by
ΔC (θ ,θ)θ + ΔG (θ ) + M (θ ) lf ]
suitable weight adaptation. The above formula represents the relationship between the
robot inverse dynamics modeling error and the tracking error, it
can be expressed in the form of the state space as follows:

TNNC
X = AX + B[−W ϕ ( x) + η ] (14)

θd μ TCT
θ Where, η − RBF neural network modeling errors,
θ
l −W ∗
i =W
W
KD KP
θ ªe º ª 0 I º ª0º
d e X = « »˗A= « »˗B = « »
θd e ¬e ¼ ¬−K p −KD ¼ ¬I ¼
When the neural network adaptive weights is selected
Fig. 2. the structure of computed torque controller with RBF neural network appropriate to update, the system tracking error vector is
compensator

361
asymptotically stable. Lyapunov equation is defined as 2
follows: − K e W Wmax + K e W
F F
(26)
θ max ·
2 2
V (t ) = X T PX + W (15) §
= K e ¨ W −
Ke
¸ − Wmax
2

© F 2 ¹ 4
( )
2
i
W i
= T WW
Let r
,where Tr( ) is the trace of a From formula (24),(25)and(26), the following can be
matrix ,For formula(15), the derivative can be obtained: obtained:

V (t ) = X T PX + X T PX + 2Tr (W T W ) (16) V = − X T QX + 2ke X ( W W − W

2
)+
F F F
From formula (14) and (16) the following can be obtained:
2η T BT PX
V = X P[ AX + B(−W ϕ ( X ) + η )] +
T
2
≤ −λmin (Q) X + ke X ( W − W
2
W )+
i (17)
[ AX + B(−W ϕ ( X ) + η )]T PX + 2Tr WW ( ) F F F

2λmax ( P) η0 X (27)

The Lyapunov equation PA + A P = −Q is substituted θ max

T

into the above equation , the following can be obtained: = − X [λmin (Q) X + ke ( W − )2 −
F 2
V = − X T QX − ϕ (X)WB
 T PX + ke 2
(18) Wmax − λmax ( P) η0 ]
η T BT PX + 2Tr (W TW ) 4
If V ≤ 0 ,the following conditions are needed:
From the equation
ke 2
i BT PX = T ª BT PX ϕ T ( X )W
ϕ ( X )W r ¬
iº
¼ (19) λmin (Q) X ≥ Wmax + λmax ( P ) η0 (28)
4
It can be obtained: That is to say, convergence condition as follows:
V = − X T QX + 2Tr[− BT PX ϕ T (X)W ] + η T BT PX (20) 2 k
X ≥ 2
[ e Wmax + λmax ( P) η0 ] (29)
Due to the presence of error ,set λmin (Q) 4

− BT PX ϕ T ( X )W iT W
i +W i ≠0 (21)
The above method can ensure the control system state
estimation and neural network weights uniformly ultimately
Meanwhile, in order to improve the robustness of the bounded. system modeling errors Domain η0 due to
system, adding the robustness items V = K e X and K e > 0 Uncertain factors decides the tracking error convergence
that we can define as followings: domain X .
i + W T W
− ª¬ BT PX ϕ T ( X ) − k e X º¼ W i =0 (22) IV. SIMULATION ANALYSIS
Proposed by the previous control method for lower
Adaptive weights of neural network adjusted for the extremity gait rehabilitation robot trajectory tracking,
following: computed torque controller with RBF neural network
compensator can be designed. Then, Simulink simulation
W = ª¬ϕ ( X ) X T PB + K e X º¼ W (23) diagram can be draw for the lower limb gait rehabilitation
robot trajectory tracking control. S-Function module in
From formula (20) and (23) we can obtain the following: Simulink simulation can get data input and torque calculation.
In addition, it also can realize the preparation of the neural
V = − X T QX + 2 K e X Tr (W TW ) + 2η T BT PX (24) network compensation control algorithm.
In RBF neural network compensation control method, let
According to the nature of the norm, the following can
be launched
ª900 0 º ª60 0 º
kp = « » ; kd = « »
Tr (W T Wˆ ) ≤ W
2
Wˆ − W (25) ¬ 0 400 ¼ ¬ 0 40 ¼
F F F
Since two legs of rehabilitation robot can be regarded as
Besides, due to two independent control body, so each leg of the rehabilitation
robot can consider as the control unit. the input layer is the

362
error value e ] created by a two-leg joint desired [e 60

Angular displacement of knee joint/deg

Simulation
trajectory value [θ d θd ] and the actual measured 50 Theory

values [θ θ ] of the track joint; output layer is compensated 40

torque of the corresponding joint; transfer function is a
30
Gaussian function, the initial weights W are set as random
numbers between [-10 10] . The hidden cell center was set as 20
ci = [ −2 −1 0 1 2] , bi = 3.0 ,(i = 1,2,3,4 5), each
10
joint position error and initial velocity are set to zero.
0
The simulation results are shown below: 0 0.2 0.4 0.6 0.8 1 1.2
Time/s

5
Theory
Angular displacement of hip joint/deg

0 Simulation
Fig.6. Knee position tracking of computed torque control with RBF neural
network compensator
-5 Figures 3 and figure 4 are the simulation results of
Computed torque control method, while figures 5 and figure 6
-10
are the simulation results of computed torque controller with
-15 RBF neural network compensator. The analysis results indicate
that the PD feedback control method based on the calculation
-20
of torque can achieve the expected control effect to some
-25 extent. However, the effect of this control method is
susceptible to the following factors: ķ system modeling error
-30
0 0.2 0.4 0.6 0.8 1 1.2 and disturbance, as well as some uncertainties .ĸ Additional
Time/s
error caused by imprecise quality distribution and inertia
Fig.3 Hip position tracking of computed torque control calculation as well as ignored impact of small parts of the
rehabilitation robot in simplifying or approximation algorithms
Angular displacement of knee joint/deg

60 on the torque calculation.

Simulation
50 Theory Comparing the analysis results, angular displacement error
40
of the rehabilitation robot which use the RBF neural network
compensator to the computed torque control method becomes
30 smaller. what’s more, it’s almost nearly zero. In addition, the
20 precision of the torque calculation has been improved to some
extent, as well as, the rehabilitation robot can quickly track the
10
desired joint angular displacement, so that the rehabilitation
0 robot joint trajectory tracking capability has been improved
significantly.
-10
0 0.2 0.4 0.6 0.8 1 1.2
Time/s V. CONCLUSION
Fig.4. Knee position tracking of computed torque control A new trajectory tracking control scheme for rehabilitation
5 robot, which is based on RBF neural network compensation, is
Angular displacement of hip joint/deg

Theory
developed in this paper. The proposed scheme consists of a
0 Simulation
well-known computed torque controller, which is based on the
-5 known nominal robot dynamics model, and a compensator,
which is based on the RBF neural network. The compensator
-10 based on RBF is used to compensate the uncertainty part of
-15
computed torque as well as the disturbance of rehabilitation
robot system. The neural network is trained on-line based on
-20 Lyapunov theory and learning convergence is thus guaranteed.
The simulation results have demonstrated the efficiency of the
-25
proposed scheme. It is believed that the proposed scheme can
-30 improve the trajectory tracking capability significantly for
0 0.2 0.4 0.6 0.8 1 1.2 rehabilitation training.
Time/s

ACKNOWLEDGMENT
The authors are very grateful for the research supported by
Fig.5. Hip position tracking of computed torque control with RBF neural the National High Technology Research and Development
network compensator

363
Program 863(No.2015AA042301) and the National Natural [7] WANG Binrui, JIN Yinglian, XU Hong, XU Xinhe.Computed
Science Fund (No.91420301). Torque and Proportion-differential Feedback Control for Robot Bionic
Knee Joint [J]. Chinese Journal of Mechanical Engineering,2008,
44(1):179-183.
[8] DAI Ying.A Class of Robust Control Strategies for Robot [J].Acta
REFERENCES Automatica Sinica, 1999,25(2):204-209.
[1] Riener R, Lunenburger L, Jezernik S, et al. Patient-cooperative [9] Wijesoma W S, Kodagoda K R S. Synthesis of stable fuzzy
strategies for robot-aided treadmill training: first experimental results PD/PID control laws for robotic manipulators from a variable
[J]. IEEE Transactions on Neural Systems and Rehabilitation structure system standpoint[C]. Proceedings of the 6th International
Engineering, 2005, 3(13):380-394. Conference on Computational Intelligence, Theory and Applications:
Fuzzy Days, 1999:495-511.
[2] Jezernik S, Colombo G, Morari M. Automatic gait-pattern
adaptation algorithms for rehabilitation with a 4-DOF robotic [10] Antonio V, Giovanni L 㧚 On the trajectory tracking control of
orthosis [J]. IEEE Transaction on Robotics and Automation, 2004, industrial SCARA robot manipulators[J] 㧚 IEEE transactions on
20(3):574-582. industrial electronics, 2002, 49(1):224-232.
[3] Agrawal S K, Banala S K, Fattah A. Assessment of motion of a swing [11] Wai R J. Tracking control based on neural network strategy for
leg and gait rehabilitation with a gravity balancing exoskeleton [J]. robot manipulator [J]. Neurocomputing, 2003, 51:425-445.
IEEE Transactions on Neural Systems and Rehabilitation [12] Roselito A, Braga A P, Benjamim. Control of a robotic
Engineering, 2007, 15(3):410-420. manipulator using artificial neural networks with on-line adaptation
[J]. Neural Processing Letters, 2004, 12:19-31.
[4] Banala S K, Kulpe A, Agrawal S K. A powered leg orthosis for gait
rehabilitation of motor-impaired patients[C]. IEEE International [13] Hu S, Ang M, Krishnan H. On-line neural network compensator
Conference on Robotics and Automation, 2007: 4140-4145 for constrained robot manipulators[C]. In Proc. of the 3rd Asian
Control Conference, Shanghai, 2000, 1621-1627.
[5] Asseldonk E H, Veneman J F, Ekkelenkamp R. The effects on
kinematics and muscle activity of walking in a robotic gait trainer [14] Khemaissia S, Morris A S. Neural-adaptive control of robot
during zero-force control [J]. IEEE Transactions on Neural Systems and manipulators [J]. Robotica, 1993, 11:456-473.
Rehabilitation Engineering, 2008, 16(4):360-370. [15] Gang F. A compensating scheme for robot tracking based on
[6] Ekkelenkamp R, Veneman J F, Kooij van der H. LOPES: selective neural networks [J]. Robotics and Autonomous System, 1995,
control of gait functions during the gait rehabilitation of CVA patients 15(3):199-206.
[C]. IEEE 9th International Conference on Rehabilitation Robotics,
2005:361-364.

364