Professional Documents
Culture Documents
Adaptive Neural Network Force Tracking Control of Flexible Joint Robot With An Uncertain Environment
Adaptive Neural Network Force Tracking Control of Flexible Joint Robot With An Uncertain Environment
Adaptive Neural Network Force Tracking Control of Flexible Joint Robot With An Uncertain Environment
This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3290250
Abstract—In this paper, a control scheme of the flexible joint works in robotic mechanical design [9]. In [10], the robot
robot contacting with an unknown environment is proposed consisted of flexible joints, which make the robot have passive
to realize force tracking. Tracking performance is ensured by compliant. It can improve the security in physical interaction
designing the force-based outer loop and the position-based inner
loop of the controller. The reference trajectory is obtained from between human and robot effectively. When joint actuators
the outer loop based on interaction force error and the estimated are subjected to external shocks, they can rely on physical
environment stiffness. The inner loop controller of the flexible flexibility to avoid damage. Despite the fact that there have
joint robot based on the singular perturbation method is designed been many control algorithms for controlling robots [11]–[13],
to achieve precise position tracking performance, and neural there are the difficulty of control design due to the introduction
network is utilized to compensate for uncertainties in robotic
dynamics. The stability of the control system is strictly proven by of springs and dampers in flexible joints. [14], [15] found that
the Lyapunov method. The effectiveness of the proposed method joint flexibility was the major impediment to the development
is verified by simulations and experiments on the flexible joint of high-performance controllers in applications. [16] proposed
robot. a flexible joint model, and since then, a large number of
Index Terms—Adaptive force tracking control, Flexible joint theoretical and experimental studies have focused on how to
robot, Singular perturbation, Neural networks control control the flexible joint robot. [17] proposed an adaptive
control law for flexible joints and analyzed the stability of
I. I NTRODUCTION the controller based on the singular perturbation method. [18]
With the development of robotics and artificial intelligence proposed a feedback linearization method and designed a
technology [1]–[3], robots have been rapidly developed and sliding mode controller to obtain robust tracking performance.
applied in various fields [4]. Due to high efficiency, good Based on the backstepping method, an adaptive controller was
stability and high repeatability accuracy, robots have been designed using neural network, which solved the problem of
widely used in automobile manufacturing, aerospace, welding, linearization of unknown parameters caused by the backstep-
handling and other industrial fields [5]. To a large extent, ping method [19]. In addition, there were various methods
robots have replaced human to complete simple and repetitive, for dealing with control issues, such as classical control [20]–
high-intensity, high-risk tasks. The position-based controllers [24], adaptive control [25]–[28], intelligent control [29]–[31]
are usually used in such works to ensure high-precision and other effective control methods [32]–[34].
tracking, so robots are required to have high stiffness and Due to the passive compliance of the flexible joint, the
good immunity to interference [6]. Robots are more and more interaction force generated when the robot contacts with the
widely used in physical interaction tasks, such as assembly environment could be buffered by its inherent impedance [35].
and limb rehabilitation [7], [8]. In these scenarios, robots in- The introduction of spring and damper has brought flexibility
evitably have direct physical interaction with the environment to the robot, however, the value of the spring and damping
or human, while traditional robots may be risky because of coefficient are fixed, hence the flexibility is limited [36]. The
their high stiffness and weight. emergence of active compliance control solved the issue of
In order to ensure safety in physical interaction tasks, passive compliance and has been widely used in human-
robots need to have compliance. Some researchers have done robot interaction tasks. [37] introduced the active compliance
control to force signal into control systems, so that the robot
X. Yu, S. Liu, S. Zhang, W. He and H. Huang are with the School of can respond to the change of force in real-time. Impedance
Intelligence Science and Technology, Beijing Advanced Innovation Center
for Materials Genome Engineering, University of Science and Technology control and hybrid force/position control are typical methods
Beijing, Beijing 100083, China; the Key Laboratory of Intelligent Bionic for achieving force tracking [38]–[40].
Unmanned Systems, Ministry of Education, University of Science and Tech- [41] proposed a novel hierarchical human-in-the-loop
nology Beijing, Beijing 100083, China; Institute of Materials Intelligent
Technology, Liaoning Academy of Materials, Shenyang 110004, China. paradigm and introduces the presented impedance learning
This work was supported in part by the National Natural Science Foundation strategy to regulate the impedance model. [42] employed adap-
of China under Grant 62225304, 62061160371 and 62003032, in part by the tive fuzzy neural networks to compensate for the uncertainty
Beijing Natural Science Foundation under Grant JQ20026, in part by the
China Postdoctoral Science Foundation under Grant 2020TQ0031 and Grant in the robotic system to improve compliance when the robot
2021M690358, in part by Scientific and Technological Innovation Foundation collides with a wall. [43] proposed a neural network based
of Shunde Graduate School, USTB under Grant BK21BF003, and in part by admittance control scheme for robotic manipulators. There are
the Fundamental Research Funds for the China Central Universities under
Grant FRF-TP-22-003C2. some works integrating impedance control into force/position
The corresponding author is W. He, Email: weihe@ieee.org. controller. Based on force/position control, [44] proposed a
Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on November 13,2023 at 19:57:23 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3290250
sensorless control method for light electric ac-tuators, which where q, q̇, q̈ ∈ Rn×1 are the robot joint angle, velocity and
enabled the robot to adopt different stiffness when interacting acceleration vectors, respectively. θ ∈ Rn×n is the motor an-
with objects. [45] proposed a new hybrid force/position control gular displacement vector, τ ∈ Rn×1 is the motor input torque
method based on a fuzzy observer. And the designed system vector, M (q) ∈ Rn×n represents the rigid link inertia matrix,
is consisted of a position observer and a contact force/torque C(q, q̇)q̇ ∈ Rn×1 is the Coriolis and centrifugal force vector,
estimator. [11] provided a parametric representation that al- G(q) ∈ Rn×1 is the gravitational force vector, B ∈ Rn×n
lowed the robot to learn compliant manipulation skills based represents the motor inertia matrix, and K ∈ Rn×n represents
on motion and force information collected from human. the joint stiffness matrix.
In this paper, we use impedance control to realize the The original flexible joint model is imperceptible to the
force tracking when the flexible joint robot contacts with an standard form of the singularly perturbed model. We set the
unknown environment. We combine the active compliance joint position q is the slow variable the joint torque τ is the
with passive compliance to ensure safety. When the controller fast variable, then we can get the standard singular perturbation
design cannot guarantee safety, the passive flexible joint can model.
also work. We design an adaptive NN controller based on the First we analyze the simplified system. Substituting (3) into
singular perturbation method [46]. Based on these discussions, (2), the joint torque equation can be obtained as:
we highlight our contributions as follows:
1) A control method for estimating the environment stiff- BK −1 τ̈m + τm = τ − B q̈ (4)
ness and implementing force tracking in an unknown
We introduce small positive parameters ε to make K =
environment is proposed for flexible joint robot;
Kε /ε2 , where Kε is a positive definite diagonal matrix, and
2) An adaptive controller for the flexible joint robot is pro-
ε is a very small constant. Then (4) can be rewritten as:
posed, which can achieve good tracking performance. It
can also handle the robotic dynamics with uncertainties ε2 B τ̈m + Kε τm = Kε (τ − B q̈) (5)
using neural networks;
3) The proposed controller can guarantee the uniform By setting ε = 0, (5) can be expressed as: τ̄m = τ̄ − J q̄ ¨,
boundness of the system on an infinite time interval substituting it into (1), the quasi-steady state equation of the
through the Lyapunov method. system can be obtained as:
¨ + C(q̄, q̄)
[M (q̄) + B]q̄ ˙ q̄˙ + G(q̄) = τ̄ − τe (6)
II. P ROBLEM FORMULATION
Then we analyze the boundary layer ayatem. (1) can be
A. Dynamic Model and Singular Perturbation Control of
rewritten as:
Flexible Joints
q̈ = M (q)−1 [τm − τe − C(q, q̇)q̇ − G(q)] (7)
Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on November 13,2023 at 19:57:23 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3290250
B. Force Tracking Impedance Control parameter of environmental stiffness will also have a great
Interaction force is generated when end-effector contacts impact on the control effect. In order to reduce the above
with the environment. As shown in (12) and (13), there is impact, we used the interaction force feedback mechanism.
a certain relationship between the deformation and interaction We introduced the interaction force error and force tracking
force of object and robot, so we can get the desired interaction error into the control loop to modify the desired trajectory after
force by controlling the relative positions. obtaining the reference trajectory by estimation. Furthermore,
the singular perturbation method is used to design the position
fe = Ke (x − xe ) (12) controller of the flexible joint to realize the real-time force
tracking.
−fe = Md (ẍ − x¨d ) + Dd (ẋ − x˙d ) + Kd (x − xd ) (13)
where fe ∈ Rn denotes the interaction force. x, xd ∈ Rn xr x
represent the actual and expected position of the effector re- Reference Controller Robot
Trajectory
spectively. The environment can be regarded as a linear spring ef Force fe
xd
model, and Ke , xe denote environmental stiffness and posi- Sensor
C. Environmental Stiffness Estimation Fig. 2. The block diagram of force tracking control.
Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on November 13,2023 at 19:57:23 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3290250
Through (24), we can calculate the correction of the reference is the input vector, S(·) is the activation function in neural
position ex after obtaining the force deviation ef and compen- networks, ψ is a n × 1 minimum neural networks approximate
sate the reference position xd to obtain the expected position error vector, which is small and bounded.
xr . If the robot did not contact the external environment, the ˙
reference force fd and the actual force fe will be 0, so ex Ŵi = −Γi [Si (Z)si + σi Ŵi ] (31)
will also be 0. The reference trajectory does not need to be where Γi is the constant gain matrix, and σi is small positive
adjusted, that is xr = xd . constants.
B. Singular Perturbation Control Method for Flexible Joints The controller can be designed as:
τs = −Kp s + τe + τada + Ŵ T S(Z) − kr sgn(z) (32)
qr
Neural where Ŵ T S(z) is the estimate of W T S(z), and sgn(z) is
Networks Fe
t NN J T Force the sliding control term for the neural networks approximate
Sensor
te
qr s t q errors.
Kp Robot q
t ada tm
Adaptive
tf
eKf
In order to provide sufficient damping to the boundary layer
Updater
system, the controller τf is proposed as
τf = −εKf τ̇m (33)
Fig. 3. Position controller of flexible joint robot.
Substituting (33) into (25), we have
We set τf = τ − τ̄ is proposed to guarantee the exponential
η 00 + J −1 KKf τ̇m + Kε [J −1 + M (q)−1 ]η = 0 (34)
stability of the boundary layer system. Define dη/dτ = η, and
the boundry layer system model is described as Then, the boundary layer system can easily be guaranteed the
exponential stability for appropriate control gain Kf .
η 00 + Kε [J −1 + M (q)−1 ]η = J −1 Kε τf (25)
So the flexible joint controller can be designed as:
The equations of boundary layer system and quasi-steady
state equation are expressed by (6) and (11), and the controllers τ = τs + τf (35)
can be designed respectively.
τs = τ̄ is uesd to guarantee the exponential stability of the
quasi-steady-state reduced system. Substituting s̄ = q̄˙ − q̄˙ r
into (9), we can obtain:
MB (q̄)(s̄˙ + q̄
¨r ) + C(q̄, q̄)(s̄
˙ + q̄˙ r ) + G(q̄) = τs − τe
(26) C. Stability analysis
We set MB (q̄) = M (q̄) + B and propose the control input
of robot as:
In this section, we analyze the stability of our proposed
method. In particular, we consider a Lyapunov function can-
τs = − Kp s + τe + τada
didates L as follows:
˙ q̇r + G(q)
+ [M (q) + B]q̈r + C(q̄, q̄) (27)
n×n
L = Le + Lw + Lf (36)
where Kp ∈ R is diagonal positive definite matrices and
the term −Kp s realizes tracking control tasks, and τada is an where:
adaptive item and the updated law is written as:
1 1 T
Le = sT Mbk s + τ τada
2 2β ada
τ̇ada (t + T ) = −βs(t) (28) n
1 X T −1
Lw = W̃ Γ W̃i
Substituting (27) into (26), we obtain: 2 i=1 i i
˙ + Kp ]s = τada 1 T 1 1
[M (q) + B]ṡ + [C(q̄, q̄) (29) Lf = (ė Md ėx + eTx Kd ex ) + eTf Ke−1 ef + ϑ̃T Υ−1 Qϑ̃
2 x 2 2
which is the dynamic equation of the subsystem.
˙ G(q) are unknown, we can use where Mbk = MB + Kp , W̃i = Ŵi − Wi represents the
Since M (q) + B, C(q̄, q̄),
RBFNN weight error, ef = fe − fd is force tracking error,
RBF neural network to approximate.
ex = x − xd is position tracking error, ek = K̂e − Ke is
˙ q̇r + G(q) = W ∗ S(Z) + ψ
[M (q) + B]q̈r + C(q̄, q̄) the error of environmental stiffness estimates, and ϑ̃ = ϑ̂ − ϑ
(30) donates the estimation error of environmental stiffness Ke .
−1
where W T ∗ S(Z) is an n × 1 ideal neural networks approxi- From (29) we can get ṡ = Mbk ˙ + Kp )s].
[τada − (C(q̄, q̄)
mate vector, Z = [q, q̇, −k1 ∗(q−qr )+ q̇r , −k2 ∗(q̇− q̇r )+ q̈r ] Substituting it and (28) into the time derivative of Le , we
Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on November 13,2023 at 19:57:23 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3290250
f d xd fe xe
Ke
We consider that ϑ is constant, then its derivation ϑ̇ = 0.
And take the time derivative of Lf and substitute (24), we
obtain:
Fig. 4. Control scheme flow chart in Simulation.
˙
L̇f = −ėTx (Dd ėx + ef ) + eTf Ke−1 ėx + ϑ̃T Υ −1 Qϑ̃
˙ In the simulation part, we designed two simulations using
= −ėTx Dd ėx − ėTx ef + eTf ėx + ϑ̃T Υ −1 Qϑ̂ a 2-DOF manipulator to verify the impact performance and
= −ėT Dd ėx − ϑ̃T Υ −1 Q(Υϕfˆe )
x
tracking performance respectively. In this simulation scenario,
we set up an uneven wall at y = 0.4 m, which the manipulator
= −ėTx Dd ėx − f˜eT Qf˜e < 0 (38)
cannot obtain the position. In order to ensure that the effector
Taking the time derivative of Lw and substituting adaptation can exert a desired force as it moves along the wall, we adopt
law (31), we obtain: dif-ferent control strategies: position control on the x-axis and
n impedance control on the y-axis, as shown in Fig. 4.
X ˙
L̇w = W̃iT Γ−1
i Ŵi
i=1 A. Impact Simulation Results
n
X
= W̃iT Γ−1
i [−Γi (Si (Z)si + σi Ŵi )]
i=1
(39) v q t20 ( xt f , yt f ) l2 q t2f
l2
( xt0 , yt0 )
Considering (37), (38) and (39), we obtain:
l1 l1
L̇ =L̇e + L̇w + L̇f y
q t10 y qt1f
T
= − s Kp s − ef Qef − ėTx Dd ėx − x x
n
X T
W̃i Γ−1
i [−Γi (Si (Z)si + σi Ŵi )] Fig. 5. Illustration of Impact.
i=1
≤ − sT Kp s − (fˆe − fe )2 − ėTx Dd ėx Fig. 5 verifies the impact performance of the designed
Xn
σi Xn
σi 1 controller. We hope the effector moves from the initial position
− kW̃i k2 + kWi k2 + kk2 of x = −0.2 m, y = 0.3 m to the wall until the interaction
2 2 2
i=1 i=1 force with the wall is the desired 10 N. In the process, the
≤ − ρL + C (40) manipulator does not know the position of the wall.
Fig. 6 shows the position and angle tracking performances,
where
which the blue line is the desired position or angle and the
ρ =min(2λmin (Kp ), 2λmin (Q), 2λmin (Dd ), red line is actual. As shown in Fig. 6(c), the interaction force
σi tracks the expected interaction force after 0.04 s with an error
min( )) (41)
λmax (Γ−1
i )
of 0.01 N in simulation.
n
X σi 1
C= kWi k2 + kk2 (42) B. Force Tracking Simulation Results
i=1
2 2
Fig. 7 verifies the force tracking performance. After the
where λ(A) denotes the eigenvalue of matrix A. We choose robot completes the impact, the effector will move along the
Kp and Dd following as (43) to make ρ > 0: wall and continue to exert 10 N interaction force during this
process.
λmin (Kp ) > 0, λmin (Q) > 0, λmin (Dd ) > 0 (43)
We compare our proposed method with existing PID-based
Based on the above analysis, we can see that the s, ef , force tracking method in the revised manuscript. Fig. 8 shows
ex and W̃i are semiglobally uniformly bounded. It can be the values of interaction forces under our proposed method and
concluded that the position x, angular velocity q̇, external PID-based method, respectively, in which the blue line repre-
force fe and the weight estimation error of RBFNN W̃i are sents the desired tracking force, and the red line represents
Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on November 13,2023 at 19:57:23 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3290250
-0.18 2 2
position[m]
angle[rad]
angle[rad]
1.35
angle[rad]
-0.19 1.5 1.5
1.3
-0.2 1 1
0.6 2 2 2
1.8 1.5
position[m]
angle[rad]
angle[rad]
0.5
angle[rad]
1.5
1.6 1
0.4 1
1.4 0.5
(a) Position tracking performance (b) Angle tracking performance (a) Angle of our proposed method (b) Angle of PID-based method
12
8
11 0.5
0.5
force[N]
10.5 0.4
6 0.4
10
0.3
0.7s 0.3
9.5
-0.18
4 1.05s 0.2
0.2
-0.19
0 0.2 0.4 0.6 0.8 1
0.1
position[m]
-0.2
position[m]
0.1
-0.21
2
2.4 2.5 2.6 2.7
0
0
0.28
-0.1
-0.1
0 0.24
0 0.5 1 1.5 2 2.5 3
-0.2 0.2 -0.2
time[s]
0.16
-0.3 7.2 7.4 7.6 7.8 8 8.2 8.4 -0.3
-0.5 -0.5
0 1 2 3 4 5 6 7 8 9 10
0 1 2 3 4 5 6 7 8 9 10
time[s] time[s]
Fig. 6. The position and angle tracking performance of impact simulation.
(a) Position tracking of our method(b) Position tracking of PID-based
method
( xt0 , yt0 ) v
q t20
l2
(xt f , yt f ) Fig. 10. The position tracking on X-axis.
l2 q t2 f
y q t10 y l1
l1
q t1 f
x x V. EXPERIMENTS
We design a force tracking experiment based on the Baxter
Fig. 7. Illustration of Force Tracking Simulation.
collaborative robot in this section, and the experiment setup is
shown in Fig. 12.
the actual external force. We can see the external force has The goal of the experiment is to control the robot to move
a mutation at 2.5 s and 7.5 s because the wall is not smooth from any initial point and clean the glass with a constant
in two places. But the interaction force changes quickly to interaction force, and it can be divided into three steps:
restore the tracking performance seen from Fig. 8. However, 1) The robot moves from the initial point to the starting
the overshoot of PID-based method is larger. point (The trajectory can be calculated by the glass
position);
11
14 2) After arrived the starting point, the robot drops the
10.8
10.6
10.2
10.1
12
appropriate position where the interaction force is re-
0.1N
10.4
9.9
10 10
lated with the desired force value obtained using the
8
impedance control method;
Force[N]
10
6
9.8
9.6
10.1
4
3) The robot cleans the glass with constant interactive force
9.4
10
9.9 0.15N
2
(The cleaning trajectory is given but the position of the
9.2
9
0 1 2 3 4 5 6
9.8
7
7
7.5 8
8
8.5
9 10
0
glass is not fixed).
0 1 2 3 4 5 6 7 8 9 10
time[s] time[s]
In the first step of the experiment, we used the position
(a) Force tracking of our method (b) Force tracking of PID-based method tracking controller on the x, y, z-axis; In the second and
Fig. 8. The performance in force tracking. third steps, the process is similar to simulation, we used the
position tracking controller on the x, y-axis, and force tracking
The blue line in Fig. 9 is the desired joint angle under controller on z-axis respectively. Since the small position
our proposed method and PID-based method, and the red tracking error will cause a large force tracking error, we must
line is the actual angle of the manipulator. We can see that ensure that the robot has good position tracking performance.
the controller has a satisfactory tracking performance, even if Fig. 13(a) shows the interaction force tracking performance,
there exists unsmooth angle change, it shows good tracking we can see that there are overshoots in the interaction force at
performance without vibration. Compared with PID-based the initial moment, but the tracking performance is quickly
method, our proposed method shows precise tracking. restored with little error. Fig. 13(b) and Fig. 13(c) show
Figs. 10 and 11 show tracking performance under our position or part of angles joint tracking performance, where
proposed method and PID-based method, where the blue line the blue line repre-sents the desired position and the red line
is desired position and the red line is actual position in X-axis represents the actual position. We can see that the controller
and Y-axis, respectively. Compared with PID-based method, has good position and force tracking performance.
our proposed method has higher tracking accuracy.
Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on November 13,2023 at 19:57:23 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3290250
0.52 0.6
0.6
0.5
0.48
0.46
0.55
0.55 0.4s
position[m]
0.44
position[m]
0.5
0.5
0.45
0.45
0.52
0.5
7 7.5 8 8.5
0.35 0.35
0 1 2 3 4 5 6 7 8 9 10
0 1 2 3 4 5 6 7 8 9 10
time[s] time[s]
Fig. 11. The position tracking on Y-axis. Fig. 13. The tracking performance of experiment.
Joint S0 Joint S1
step 1: Robot moves to the starting point. (T = 0 - 3.2s) 0.5 0.1 -0.8 0.1
0.2
0.05 -0.9 0.05
Angle error[rad]
Angle error[rad]
Angle[rad]
Angle[rad]
-0.1
0 -1 0
-0.4
1.3
0.05 1.9 0.05
step 3: Robot cleans the glass with constant force. (T = 6.7 ā 20s)
Angle error[rad]
Angle error[rad]
Angle[rad]
Angle[rad]
1.2
0 1.7 0
1.1
-0.575
-0.6
Angle error[rad]
Angle error[rad]
Angle[rad]
Angle[rad]
Fig. 12. Illustration of the experiment setup. -0.625 0 1.2 0
-0.65
-0.675
In this paper, we proposed a control scheme for the flexible (e) Joint W0 (f) Joint W1
joint robot to realize force tracking in an unknown environ- 0.7
Joint W2
0.1
0.1
According to the force error and the estimated stiffness of the -0.2
time[s]
singular perturbation ensured the tracking performance. The
(g) Joint W2
RBFNN in the controller was used to compensate for uncer-
tainties of robotic dynamics. And we used Lyapunov method to Fig. 14. The tracking performance of 7 joints in experiment.
prove the stability and verified the effectiveness by simulations
and experiments.
In the future, considering the scenario where the flexible [2] C. Cui, B. Wang, Y.-x. Zhao, Y.-J. Zhang, and L.-m. Xue, “Risk man-
joint robot contacts with the complex environment of un- agement for mine closure: A cloud model and hybrid semi-quantitative
decision method,” International Journal of Minerals, Metallurgy and
known contacting surface and time-varying stiffness, we aim Materials, vol. 27, no. 8, pp. 1021–1035, 2020.
to propose a variable impedance control strategy to maintain a [3] Z. Xu, Y. Kang, Y. Cao, and Z. Li, “Deep amended copert model
constant interaction force, which means the stiffness parameter for regional vehicle emission prediction,” Science China Information
Sciences, vol. 64, no. 3, pp. 1–3, 2021.
is time-varying and the contacting surface is unknown. [4] W. He, Z. Li, and C. L. P. Chen, “A survey of human-centered intelligent
robots: issues and challenges,” IEEE/CAA Journal of Automatica Sinica,
R EFERENCES vol. 4, no. 4, pp. 602–609, 2017.
[5] P. Maurice, M. E. Huber, N. Hogan, and D. Sternad, “Velocity-curvature
[1] D. Huang, B. Li, Y. Li, and C. Yang, “Cooperative manipulation of patterns limit human–robot physical interaction,” IEEE robotics and
deformable objects by single-leader–dual-follower teleoperation,” IEEE automation letters, vol. 3, no. 1, pp. 249–256, 2018.
Transactions on Industrial Electronics, vol. 69, no. 12, pp. 13 162– [6] H. Lu, W. Shang, H. Xie, and Y. Shen, “Ultrahigh-precision rotational
13 170, 2022. positioning under a microscope: Nanorobotic system, modeling, control,
Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on November 13,2023 at 19:57:23 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3290250
and applications,” IEEE Transactions on Robotics, vol. 34, no. 2, pp. joint robots with uncertain parameters,” IEEE Transactions on Control
497–507, 2018. Systems Technology, vol. 26, no. 6, pp. 2259–2267, 2018.
[7] M. Pi, Y. Kang, C. Xu, G. Li, and Z. Li, “Adaptive time-delay balance [28] C. Zhu, Y. Jiang, and C. Yang, “Fixed-time neural control of robot
control of biped robots,” IEEE Transactions on Industrial Electronics, manipulator with global stability and guaranteed transient performance,”
vol. 67, no. 4, pp. 2936–2944, 2020. IEEE Transactions on Industrial Electronics, vol. 70, no. 1, pp. 803–812,
[8] J. C. Pulido, C. Suárez-Mejı́as, J. Dorado, A. D. Ruiz, and F. Fernández, 2023.
“A socially assistive robotic platform for upper-limb rehabilitation: A [29] C. Sun, W. He, and J. Hong, “Neural network control of a flexible robotic
longitudinal study with pediatric patients,” IEEE Robotics & Automation manipulator using the lumped spring-mass model,” IEEE Transactions
Magazine, vol. 26, no. 2, pp. 24–39, 2019. on Systems, Man, and Cybernetics: Systems, vol. 47, no. 8, pp. 1863–
[9] G. Khademi, H. Mohammadi, H. Richter, and D. Simon, “Optimal mixed 1874, 2017.
tracking/impedance control with application to transfemoral prostheses [30] L.-H. Kong, W. He, W.-S. Chen, H. Zhang, and Y.-N. Wang, “Dynamic
with energy regeneration,” IEEE Transactions on Biomedical Engineer- movement primitives based robot skills learning,” Machine Intelligence
ing, vol. 65, no. 4, pp. 894–910, 2017. Research, pp. 1–12, 2023.
[10] C.-C. Cheah and D. Wang, “Learning impedance control for robotic [31] C. Cheng, H. Zhang, H. Peng, Z. Zhou, B. Chen, Z. Zeng, and H. Lu,
manipulators,” IEEE Transactions on Robotics and Automation, vol. 14, “Stability control for end effect of mobile manipulator in uneven terrain
no. 3, pp. 452–465, 1998. based on active disturbance rejection control,” Assembly Automation,
[11] C. Zeng, Y. Li, J. Guo, Z. Huang, N. Wang, and C. Yang, “A unified vol. 41, no. 3, pp. 369–383, 2021.
parametric representation for robotic compliant skills with adaptation [32] G. Yu, J. Reis, D. Cabecinhas, R. Cunha, and C. Silvestre, “Reduced-
of impedance and force,” IEEE/ASME Transactions on Mechatronics, complexity active disturbance rejection controller for quadrotor-slung-
vol. 27, no. 2, pp. 623–633, 2021. load transportation,” IEEE Transactions on Systems, Man, and Cyber-
[12] J. Xie, S. Y. Liu, and J. X. Chen, “A framework for distributed semi- netics: Systems, 2023.
supervised learning using single-layer feedforward networks,” Machine [33] H.-R. Yang and W. Ni, “Continuous-time distributed heavy-ball algo-
Intelligence Research, vol. 19, no. 1, pp. 63–74, 2022. rithm for distributed convex optimization over undirected and directed
[13] Z. Li, B. Huang, Z. Ye, M. Deng, and C. Yang, “Physical human–robot graphs,” Machine Intelligence Research, vol. 19, no. 1, pp. 75–88, 2022.
interaction of a robotic exoskeleton by admittance control,” IEEE [34] X. Jiang, X. Zeng, J. Sun, J. Chen, and Y. Wei, “A fully distributed hy-
Transactions on Industrial Electronics, vol. 65, no. 12, pp. 9614–9624, brid control framework for non-differentiable multi-agent optimization,”
2018. IEEE/CAA Journal of Automatica Sinica, vol. 9, no. 10, pp. 1792–1800,
[14] L. Sweet and M. Good, “Redefinition of the robot motion-control 2022.
problem,” IEEE Control Systems Magazine, vol. 5, no. 3, pp. 18–25, [35] C. Lv, Q. Zhu, and R. Xiong, “An adaptive compliance position control
1985. based on ekf for series elastic actuation,” in Proceeding of the 11th
[15] M. Good, L. Sweet, and K. Strobel, “Dynamic models for control system World Congress on Intelligent Control and Automation. IEEE, 2014,
design of integrated robot and drive systems,” Journal of Dynamics pp. 3100–3106.
Systems, Measurement and Control, vol. 107, no. 1, pp. 53–59, 1985. [36] K. Haninger and M. Tomizuka, “Robust passivity and passivity re-
laxation for impedance control of flexible-joint robots with inner-loop
[16] M. W. Spong, “Modeling and control of elastic joint robots,” Journal
torque control,” IEEE/ASME Transactions on Mechatronics, vol. 23,
of Dynamics Systems, Measurement and Control, vol. 109, no. 1, pp.
no. 6, pp. 2671–2680, 2018.
310–319, 1987.
[37] A. Calanca, R. Muradore, and P. Fiorini, “A review of algorithms for
[17] Z.-H. Jiang and L. Jiang, “Impedance control of flexible joint robots
compliant control of stiff and fixed-compliance robots,” IEEE/ASME
based on singular perturbation method,” in 2019 IEEE International
Transactions on Mechatronics, vol. 21, no. 2, pp. 613–624, 2016.
Conference on Cybernetics and Intelligent Systems (CIS) and IEEE
[38] C. Zhu, Y. Jiang, and C. Yang, “Fixed-time neural control of robot
Conference on Robotics, Automation and Mechatronics (RAM). IEEE,
manipulator with global stability and guaranteed transient performance,”
2019, pp. 548–552.
IEEE Transactions on Industrial Electronics, vol. 70, no. 1, pp. 803–812,
[18] M. Yallala and S. Mija, “Path tracking of differential drive mobile robot 2022.
using two step feedback linearization based on backstepping,” in 2017 [39] Z. Li, J. Liu, Z. Huang, Y. Peng, H. Pu, and L. Ding, “Adaptive
International Conference on Innovations in Control, Communication and impedance control of human–robot cooperation using reinforcement
Information Systems (ICICCI). IEEE, 2017, pp. 1–6. learning,” IEEE Transactions on Industrial Electronics, vol. 64, no. 10,
[19] J. Zietkiewicz and A. Owczarkowski, “Direct nonlinear model predictive pp. 8013–8022, 2017.
control and predictive control with feedback linearization. a comparison [40] Y. Zhang, L. Kong, S. Zhang, X. Yu, and P. Yang, “Tracking control
of the approaches,” in 2017 18th International Carpathian Control of a robotic system with deferred constraints and actuator faults,” IET
Conference (ICCC). IEEE, 2017, pp. 451–455. Control Theory & Applications, vol. 15, no. 9, pp. 1257–1269, 2021.
[20] H. Sun, L. Huang, and L. He, “Research on the trajectory tracking [41] Z. Li, X. Li, Q. Li, H. Su, Z. Kan, and W. He, “Human-in-the-loop
control of a 6-dof manipulator based on fully-actuated system models,” control of soft exosuits using impedance learning on different terrains,”
Journal of Systems Science and Complexity, vol. 35, no. 2, pp. 641–659, IEEE Transactions on Robotics, vol. 38, no. 5, pp. 2979–2993, 2022.
2022. [42] W. He and Y. Dong, “Adaptive fuzzy neural network control for a
[21] L. Zhang, Z. Li, and C. Yang, “Adaptive neural network based vari- constrained robot using impedance learning,” IEEE Transactions on
able stiffness control of uncertain robotic systems using disturbance Neural Networks and Learning Systems, vol. 29, no. 4, pp. 1174–1186,
observer,” IEEE Transactions on Industrial Electronics, vol. 64, no. 3, 2017.
pp. 2236–2245, 2017. [43] G. Peng, C. P. Chen, W. He, and C. Yang, “Neural-learning-based force
[22] Y. Liu, X. Chen, Y. Mei, and Y. Wu, “Observer-based boundary control sensorless admittance control for robots with input deadzone,” IEEE
for an asymmetric output-constrained flexible robotic manipulator,” Transactions on Industrial Electronics, vol. 68, no. 6, pp. 5184–5196,
Science China Information Sciences, vol. 65, no. 3, pp. 1–3, 2022. 2020.
[23] G. Duan and G.-P. Liu, “Attitude and orbit optimal control of combined [44] B. Huard, M. Grossard, S. Moreau, and T. Poinot, “Sensorless
spacecraft via a fully-actuated system approach,” Journal of Systems force/position control of a single-acting actuator applied to compliant
Science and Complexity, vol. 35, no. 2, pp. 623–640, 2022. object interaction,” IEEE Transactions on Industrial Electronics, vol. 62,
[24] H. Huang, W. He, J. Wang, L. Zhang, and Q. Fu, “An all servo-driven no. 6, pp. 3651–3661, 2014.
bird-like flapping-wing aerial robot capable of autonomous flight,” [45] Z. Lu, P. Huang, Z. Liu, and H. Chen, “Fuzzy-observer-based hybrid
IEEE/ASME Transactions on Mechatronics, 2022. force/position control design for a multiple-sampling-rate bimanual
[25] H. Huang, W. He, Q. Fu, X. He, and C. Sun, “A bio-inspired flapping- teleoperation system,” IEEE Transactions on Fuzzy Systems, vol. 27,
wing robot with cambered wings and its application in autonomous no. 7, pp. 1383–1396, 2019.
airdrop,” IEEE/CAA Journal of Automatica Sinica, vol. 9, no. 12, pp. [46] M. Spong, K. Khorasani, and P. Kokotovic, “An integral manifold
2138–2150, 2022. approach to the feedback control of flexible joint robots,” IEEE Journal
[26] M. Yuan and X. Zhang, “Stability and fast transient performance oriented on Robotics and Automation, vol. 3, no. 4, pp. 291–300, 1987.
motion control of a direct-drive system with modeling uncertainties,
velocity, and input constraints,” IEEE/ASME Transactions on Mecha-
tronics, vol. 27, no. 6, pp. 5926–5935, 2022.
[27] L. Le-Tien and A. Albu-Schäffer, “Robust adaptive tracking control
based on state feedback controller with integrator terms for elastic
Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on November 13,2023 at 19:57:23 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Industrial Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIE.2023.3290250
Xinbo Yu (Member, IEEE) received the B.E. degree Haifeng Huang (Member, IEEE) received the
in control technology and instrument and the Ph.D. B.Eng. degree in automation and the Ph.D. degree in
degree in control science and engineering from the control science and engineering from the School of
School of Automation and Electrical Engineering, Automation and Electrical Engineering, University
University of Science and Technology Beijing, Bei- of Science and Technology Beijing, Beijing, China,
jing, China, in 2013 and 2020, respectively. in 2016 and 2022, respectively.
He is currently working as an Associate Professor His current research interests include bionic
with the School of Intelligence Science and Technol- flapping-wing robot, neural network control, and
ogy, University of Science and Technology Beijing. fault-tolerant control.
His current research interests include adaptive NNs
control, robotics and human-robot interaction. His
current research interests include intelligent robot control and human-robot
collaboration.
Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on November 13,2023 at 19:57:23 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.