This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 1

A Robust Human–Robot Collaborative Control

Approach Based on Model Predictive Control
Tianyi Zeng , Abdelkhalick Mohammad , Andres Gameros Madrigal , Dragos Axinte ,
and Max Keedwell

Abstract—Human skill-based robotic control to perform z̄  z-axis component of the nominal position.
critical manufacturing operations (e.g., repair and inspec-
tion for high-value assets) can reduce scrap rates and in- Disturbances
crease overall profitability in the industrial community. In dx  x-axis component of the external disturbance.
this study, a human–robotic collaborative control system is
developed for accurate path tracking subject to unknown
dy  y-axis component of the external disturbance.
external disturbances and multiple physical constraints. dz  z-axis component of the external disturbance.
This is achieved by designing a model predictive control System Parameters
with a sliding-mode disturbance rejection term. To rule out
the possibility of the constraints violation caused by ex- Ts  Sampling time.
ternal disturbances, tightened constraints are formulated Controller
to generate the control input signal. The proposed con-
troller drives the robotic system remotely with enhanced ux  control input along x-axis.
smoothness and real-time human modification on the out- uy  control input along y-axis.
putted performance so that the human experience can be
fully transferred to robotic systems. The efficacy of the uz  control input along z-axis.
proposed collaborative control system is verified by both ux  nominal control input along x-axis.
Monte–Carlo simulation with 200 cases and experimental uy  nominal control input along y-axis.
results including tungsten inert gas welding based on a uz  nominal control input along z-axis.
universal robot 5e with 6 degree-of-freedom.
Index Terms—Disturbance rejection, model predictive
control (MPC), remote control, universal robot (UR). I. INTRODUCTION
NDUSTRIAL robots are widely used in manufacturing and
NOMENCLATURE I assembly lines, such as the automotive and white-goods
industries, as they offer major improvements in productivity,
Variables in Actual System safety, ease of programming, portability, and cost saving [1], [2].
x x-axis component of the position. Much effort has been put into the design of robotic solutions, e.g.,
y y-axis component of the position. 6 degree-of-freedom (DoF) with payload capacities improve-
z z-axis component of the position. ments to assist a wide range of applications and industries [3].
Nevertheless, their controllers still seem to work based on
Variables in Nominal System
point-to-point vectoring of the desired paths as this is acceptable
x̄  x-axis component of the nominal position. for many applications, such as spot welding, assembly, part
ȳ  y-axis component of the nominal position. manipulations, and even machining where such operations are
often used in the automotive industry. However, the stability of
moving robotic end effectors is very sensitive when it comes to
Manuscript received 3 November 2022; revised 15 February 2023 continuous operations with high-accuracy requirements, such
and 4 May 2023; accepted 13 July 2023. This work was supported as welding/gluing for curved paths where a small change in
by Innovate U.K. through REINSTATE: Repair, Enhanced Inspection,
and Novel Sensing Techniques for increased Availability and reduced both the distance and the velocity between the end effector
Through life Expense under Grant TS/V005103/1. (Corresponding au- and the component may produce a considerable variation in
thor: Abdelkhalick Mohammad.) the process outcomes [4], [5]. Taking further the example of
Tianyi Zeng, Abdelkhalick Mohammad, Andres Gameros Madri-
gal, and Dragos Axinte are with the Rolls-Royce UTC in Manu- robotic welding/gluing as processes requiring “smooth” and
facturing and On-Wing Technology, University of Nottingham, NG8 continuum end effector manipulations, the robot should keep
1BB Nottingham, U.K. (e-mail: [email protected]; abd. moving after each waypoint without any stop during the whole
[email protected]; [email protected];
[email protected]). process. Due to the shortage of the skilled worker, especially
Max Keedwell is with the Rolls-Royce plc., BS34 7QE Bristol, U.K. the experienced welder (e.g., left-handed welder), it would be
(e-mail: [email protected]). desirable that industrial robots can provide continuous velocity
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TIE.2023.3299046. to guarantee this smoothness so that it can mimic or replicate
Digital Object Identifier 10.1109/TIE.2023.3299046 the characteristics that human limbs posses.
0278-0046 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on September 28,2023 at 13:48:23 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Therefore, concern over control performance with the A remote control scheme based on MPC was proposed for
smoothness not only in the path but also in velocity to guarantee welding [18]. The influence caused by the external disturbance
the manufacturing quality and fast dynamic response result in a of the system was remained to be further investigated. A control
set of challenging problems for robotic control. scheme combines MPC and sliding mode control (SMC) was
Currently, most applications of industrial robots are based on developed in [19] so that the MPC can be achieved based on
movement among waypoints, which results in relative “rugged” a system with reduced uncertainty. Such control scheme has
performance, especially when it comes to accurate and continu- been applied to the robotic system [20]. However, although
ous path tracking. For example, current industrial welding robots many control methods have been proved that can be used in
are articulated arms with a preprogrammed set of movements open architectures [21], [22], [23], such as servo motors, how to
based on separate waypoints, such as “MoveL” command of utilize this method to real robot and guarantee the smoothness
universal robot (UR), which leads to a discontinuous movement based on such control scheme is still to be investigated. Most of
as the robot has to stop after each waypoint. Although “MoveP,” the industrial robots’ controllers including the UR, are closed
a movement type intended for processes, can be employed to architecture (i.e., the control designer has no access to the low-
achieve a continuous movement, it will lose the real-time input level control). Furthermore, existing standards [24] and methods
to modify the path during the whole process, which means of human–robot collaborative systems are focused on safety
the human–robot collaboration is failed to be provided. This concerns, such as obstacle avoidance [25]. To introduce human
emphasizes the need for advanced control methods that can skills and experience into the robotic system, the collaboration
make the robots more able to mimic the capabilities (e.g., path between human and robots should not only remain on the ob-
and velocity “smoothness”) of human operators and be open stacle avoidance, but also be extended to response human input
to accept real-time corrections to fulfill critical applications. in the real time with satisfactory control performance. Despite
A radial basis function neural network was introduced for in- MPC and SMC has been used in industrial system, up to date,
dustrial robot manipulators to deal with uncertain dynamical there is no industrial robot that able to accept human inputs
environments [6]. The adaptive parameter estimation and control (e.g., fine adjustment from expert welder) during executing a
for nonlinear robotic systems were also developed in [7]. An predefined path with given speed and trajectory. Design robotic
adaptive control method was developed to improve the tracking systems with human–robot collaboration and capability of mim-
performance of 6 DoF industrial robots. Compared with the icking the operation of skilled operators will dramatically reduce
conventional PID control, the proposed control method can the maintenance cost and downtime in the industry as they can
achieve a smaller tracking error when external disturbances provide efficient critical operations (e.g., repair for specific com-
and parametric uncertainties are taken into consideration [8]. ponent shapes), reduce scrap rates, increase overall profitabil-
A sliding-mode controller is applied as position controller for ity, and reduce CO2 emissions during the whole maintenances
compliant mechanics to improve the tracking and disturbance process.
response [9]. However, these research studies focused on posi- To address these gaps stated above, this article reports on
tion control. How to regulate the velocity and perform smooth a control algorithm that allows the smooth arbitrary path of
movement during the operation process still remains to be further multi-DoF robots capable to mimic human movement. We took
investigated. If the robots could present similar smoothness in further this new capability and integrate the robot with human
movements as human limbs, this would provide a direct link experiences (i.e., a database of human hand movement with an
from human intelligence to robotics. end effector) so that the trajectories of humans can be replicated
In relation to smoothness/stable motion of robots, a fuzzy by the robot in real time. This opens the avenue that actually
logic-based controller was proposed to reduce motion chattering the robot movements can be corrected in real time when the
phenomena on a wheeled mobile robot [10]. Also, time-optimal human/any other sensing device can observe deviations not only
controller that allows the smooth control of rigid robotic ma- in trajectory but also in “smoothness.” The main contributions
nipulators was designed based on high-order kinematic vari- of this article can be summarized as follows.
ables [11]. Jerk-limited trajectory planning methods have also 1) A human–robot collaborative system is developed with
been proposed to reduce vibrations of robotic motions [12], enhanced smoothness and robustness so that the input
[13], [14]. However, these research results focus on path plan- from the skilled operator can be responded in the real
ning rather than the actual output with smoothness for robotic time during the path tracking process.
systems. 2) A control scheme based on MPC and the sliding-mode
Model predictive control (MPC) has received considerable compensator is implemented for both position tracking
attention in past decades due to its online optimization capability and speed regulation of industrial robots to follow a
considering state and input constraints [15], [16]. Compared predefined trajectory while allowing the expert user to
with traditional optimal control, the online optimization fea- deviate from this desired trajectory in the real time and in
ture of MPC can update the optimal control sequence of the a smooth manner.
system [17]. Therefore, control strategies based on the MPC 3) The effectiveness of the proposed control scheme is ex-
concept have found wide acceptance in industrial applications perimentally verified to guarantee the practical value of
and have been widely studied by academia. Particularly, the such collaborative robotic systems with a wide range of
MPC approaches have been developed for industrial robotics. applications.

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on September 28,2023 at 13:48:23 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZENG et al.: ROBUST HUMAN–ROBOT COLLABORATIVE CONTROL APPROACH BASED ON MODEL PREDICTIVE CONTROL 3

The rest of this article is organized as follows. The control- B. Physical Constraints
oriented model considering the constraints and external distur-
During the operation, several physical constraints should be
bances is established in Section II. The control scheme including
considered. The position should be within safe range, which
online optimization based on MPC and disturbance compensator
leads to the following state constraints:
is developed in Section III. Simulation and experimental results
are demonstrated in Section IV to verify the efficacy of the x ∈ X, X = {x ∈ R : xmin ≤ x ≤ xmax } (5)
control proposed strategy. Finally, Section V concludes this y ∈ Y , Y = {y ∈ R : ymin ≤ y ≤ ymax } (6)
article.
z ∈ Z, Z = {z ∈ R : zmin ≤ z ≤ zmax } (7)
II. CONTROL-ORIENTED MODEL ESTABLISHMENT where xmin , xmax , ymin , ymax , zmin , and zmax denotes the upper
A control-oriented model is established for the path tracking and lower boundaries of the position along three axes, which are
control of the industrial robot. Both the external disturbance and constants.
physical constraints are considered in this section. The industrial Due to the actuation limitation, the velocities, which are also
robot is connected to a computer by an Ethernet cable so that the control inputs should be in reasonable range as follows:
the remote operation can be enabled with an advanced control ux ∈ Ux , Ux = {ux ∈ R : |ux | ≤ ux,max } (8)
strategy.
uy ∈ Uy , Uy = {uy ∈ R : |uy | ≤ uy,max } (9)
A. Dynamic Model uz ∈ Uz , Uz = {uz ∈ R : |uz | ≤ uz,max } (10)
To achieve a high-performance control for the position of the
where ux,max , uy,max , and uz,max are the maximum velocities that
end effector, the kinetics of the UR and its transformation to the
can be provided along three axes, which are constants.
end effector are, respectively, modeled. This is to consider the
Assumption 1: The external disturbance caused by vibrations,
control signal is actuated to manipulate the UR directly and to
etc., along the i-axis is bounded by a constant Di > 0 with
take the vibration from the end effector as well as other external
i = x, y, z.
disturbances into account.
The velocity along i-axis is the control input defined as ui
III. MODEL PREDICTION CONTROL WITH DISTURBANCE
with i = x, y, z. Considering the vibration of the end effector
REJECTION
and other external disturbances, we have the following dynamic
model: In this section, an MPC with sliding mode (SM) compensator
strategy is proposed to track a preplanned trajectory subject
ẋ = ux + dx to unknown external disturbances, uncertainties, and multiple
ẏ = uy + dy constraints. The compensator works on actual model providing
disturbance rejection so that MPC can work on the nominal
ż = uz + dz (1) model coping with constraints. Due to the disturbance rejection
and its discrete-time model obtained by zero-order holder is performance of the integral sliding mode (ISM), and tightened
constraints, the original constraints are satisfied and the whole
xk+1 = xk + Ts uk,x + Ts dk,x system is robust against disturbances.
yk+1 = yk + Ts uk,y + Ts dk,y
A. Overall Strategy
zk+1 = zk + Ts uk,z + Ts dk,z (2)
The control system diagram is shown in Fig. 1. A preplanned
where Ts is the sampling time, xk , yk , and zk are the system trajectory is considered as a desired signal fed into the controller,
states at time t = tk , and uk,x , uk,y , and uk,z are the system and with the real-time feedback from the UR, the proposed
inputs at time t = tk . controller determines the control action to stabilize the system.
The nominal model (disturbance-free model) of (1) is The control signal of the i-axis contains two components with
ẋ = ux i = x, y, z. One control component, which is the nominal control
input, is to stabilize the system and cope with the constraints,
ẏ = uy and the other one is to cope with the external disturbances, which
drive the state trajectory away from the desired or preplanned
ż = uz . (3)
path. Furthermore, the constraints considered to generate the
The nominal model (disturbance-free model) of (2) is nominal control input are also tightened from the original con-
straints to ensure that the original constraints are satisfied in all
xk+1 = xk + Ts uk,x
cases subject to external unknown disturbances.
y k+1 = y k + Ts uk,y The control strategy along the i-axis is proposed as
z k+1 = z k + Ts uk,z (4) ui = uMPC,i + uISM,i (11)
where xk , y k , and z k are the nominal states at time t = tk , and where uMPC,i is the nominal control input for the velocity
uk,x , uk,y , and uk,z are the nominal control inputs at time t = tk . control and constraints handing, i.e., ui = uMPC,i , and uISM,i

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on September 28,2023 at 13:48:23 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

where N is the prediction horizon.

Choose the new state vector as
X = [xk+1 , xk+2 , . . . , xk+N ]T (14)
and the new input vector as
Ux = [uk,x , uk+1,x , . . . , uk+N −1,x ]T . (15)
The optimization problem over the N -step horizon is
PN,x = arg min Jx
[uk,x ,...,uk,x+N −1 ]

s.t. xk+i ∈ X
uk+i,x ∈ U x (16)
where
Jx = X T Qx X + UxT Rx Ux (17)
and U x is the tightened constraint on the input, X is the tight-
ened constraint on the state; Qx and Rx as constant weight
matrices that are positive definite. Consider the nominal model
(4), and ⎡define the system matrix
⎤ as Ax = [1, 1 · · · 1]T and
1 0 ··· ··· 0
⎢1 1 0 · · · 0⎥
⎢ ⎥
B x = Ts ⎢ . .. .. ⎥, the cost function (17) can be
Fig. 1. Control strategy scheme of SM compensator-based MPC with ⎣ .. . . 0⎦
constraints and uncertainties.
1 ··· 1
rewritten as
is the compensator for coping with external disturbances, with ∗
J x = UxT Hx Ux + FxT Ux (18)
i = x, y, z.
In the following sections, for simplification purposes, we where
consider the model of x-axis and design a controller, and then, 1 T
similar results are given for other two axes. Hx = B x Qx B X + R x (19)
2
and
B. MPC Design 1
FxT =x0 Ax Qx Bx . (20)
The nominal control input ui = uMPC,i is generated based 2
on the nominal model (4). The working principle of MPC is Therefore, an optimization can be undertaken based on a
based on a receding horizon strategy where a control sequence single input variable Ux , and in MATLAB, it is achieved by
is generated by an online optimization, and only the first element “quadprog” program. Only the first element of the resulted
of the control sequence is applied to the system. optimal sequence Ux = [uk,x , uk+1,x , . . . , uk+N −1,x ]T , which
According to the nominal model (4), we have the following is uk,x , is applied to the system as the current nominal control
state evolutions along the x-axis: input at time t = tk .
Remark 1 (Tightened constraints): Since a compensation
xk+1 = xk + Ts uk,x
control input is introduced to cope with the external distur-
xk+2 = xk+1 + Ts uk+1,x bances, the input constraints for the MPC subsystem should be
tightened to ensure that the total control input does not exceed
= xk + Ts uk,x + Ts uk+1,x
the limitation. In the sequel, the amplitude of the compensation
.. control input is designed as Kism,i with i = x, y, z, therefore,
. the tightened input constraints U i ∈ Ui are calculated as
xk+N = xk+N −1 + Ts uk+N −1,x (12) U i = {ui : |ui | ≤ ui,max − Kism,i } (21)
which gives with i = x, y, z.
⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤ In the following section, we prove that the nominal state
xk+1 1 1 0 ··· ··· 0 uk,x
⎢ xk+2 ⎥ ⎢1⎥ ⎢1 1 0 ··· 0⎥ ⎢ ⎥ approximates the actual state by applying the compensator, so
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ uk+1,x ⎥ the tightened constraints on the state are calculated as X = X,
⎢ .. ⎥ = ⎢ .. ⎥ xk + ⎢ .. .. .. ⎥⎢ .. ⎥ Ts
⎣ . ⎦ ⎣.⎦ ⎣. . . 0⎦ ⎣ . ⎦ Y = Y , and Z = Z.
xk+N 1 1 ··· 1 uk+N −1,x Following the similar procedure, nominal control inputs along
(13) y-axis and z-axis can be achieved.

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on September 28,2023 at 13:48:23 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZENG et al.: ROBUST HUMAN–ROBOT COLLABORATIVE CONTROL APPROACH BASED ON MODEL PREDICTIVE CONTROL 5

C. Disturbance Compensator Design

The compensation control input uISM,x is designed as follows:
uISM,x = −Kism,x sign(sx ) (22)
where Kism,x is a positive constant satisfying Kism,x ≥ Dx and
sx is sliding variable defined by
t
sx = Lx x(t) − x(t0 ) − uMPC,x (τ )dτ (23)
t0

with Lx as a positive constant, and the function of sign(Δ) is

defined by
Δ
Δ , Δ = 0
sign(Δ) = (24)
0 , Δ = 0. Fig. 2. System overview of the proposed control scheme.
Theorem 1: By using the compensation control input (22), the
unknown external disturbances can be eliminated and the closed-
loop dynamics of the actual model approximates the nominal Remark 2: The implementation for practical robotic systems
model. will be influenced by the complexity of the proposed control
Proof: The proof is divided into two parts. First, we prove scheme. SMC is verified to be computational lightweight, but
that the proposed sliding variable converges to zero and stays at MPC is expected with a higher computational cost. However,
zero afterwards. with simple constrains and models, the computational burden
Select a Lyapunov candidate as V = 12 s2x , and its first time can be further reduced. A quantified complexity analysis can
derivative is be found in [20]. Meanwhile, the prediction horizon should be
V̇ = sx Lx (ẋ − uMPC,x (τ )) set with a tradeoff between control performance and calculation
speed. By choosing a smaller prediction horizon, a faster calcu-
= sx Lx (uISM,x + dx ) lation speed can be achieved but it will lead to a worse control
≤ − Kism,x Lx |sx | + Dx Lx |sx |. (25) performance. A better control performance can be achieved if a
larger prediction horizon is selected as more future information
Since Kism,x is set to be Kism,x > Dx , we have that V̇ < can be introduced during the optimization. However, a slower
0 holds for any sx = 0 and V̇ = 0 holds only if sx = 0 calculation speed will be obtained. In this study, a 500 Hz
holds. Therefore, the sliding variable s converges to zero and rate, which is the highest data exchanging rate of UR, can be
sx = ṡx = 0 holds afterwards. guaranteed by the proposed control scheme.
Second, we prove that once the sliding mode is maintained, the
closed-loop dynamics of the actual model becomes the nominal
model. IV. SIMULATION AND EXPERIMENT VALIDATION
From sx = ṡx = 0, we have In this section, we will testify if the proposed control system
ṡx = ẋ − uMPC,x = 0 (26) can provide a high-enough level of repeatability, accuracy, and
dynamic response, particularly in highly stressed and difficult
which yields ẋ = uMPC,x , and since uMPC,x = ux holds, we have to access locations.
ẋ = uMPC,x . Therefore, the closed-loop dynamics approximates
the nominal dynamics. 
Similarly, the compensation control inputs along y-axis and A. Test Rig Set-Up
z-axis can be obtained as The human–robot collaborative is established as shown in
uISM,y = −Kism,y sign(sy ) (27) Fig. 2. A UR-5e with 6 DoF is employed as case studies.
The UR-5e is connected by an ethernet cable with an Intel
and Xeon 3.5 GHz computer by running a Python program as the
uISM,z = −Kism,z sign(sz ) (28) developing environment. By using the current setup, no obvious
latency is achieved as the communication frequency between
where Kism,y and Kism,z are positive constants satisfying the UR-5e and the computer is 500 Hz, which guarantees a fast
Kism,y ≥ Dy and Kism,z ≥ Dz , and sy and sz are sliding vari- real-time data exchange. An Xbox One joystick linked to the
ables defined by computer is employed as the real-time human input. The input
t from the joystick is introduced into the control scheme so that
sy = Ly y(t) − y(t0 ) − uMPC,y (τ )dτ (29) it can be responded during the path tracking process. Multiple
t0
kinds of end effectors can be attached to the robot arm so that the
t
proposed control method can be extended to wide inspection and
sz = Lz z(t) − z(t0 ) − uMPC,z (τ )dτ . (30)
t0 repair tasks even in hazardous environments, such as nuclear and

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on September 28,2023 at 13:48:23 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

TABLE I standard deviation of the initial position is 0.4, 0.2, and 0.025 m,
CONTROLLER PARAMETERS
respectively for three axes. Meanwhile, a time-varying external
disturbance d = 0.5 sin(t) +  is introduced in the simulation.
Here,  ∼ N (0, 0.1) is a white noise.
Fig. 4 shows that all the axes can converge after 1 s and
the largest tracking error in the static state can be guaranteed
within 3 mm for all three axes. Hence, the satisfactory control
performance can be achieved with the initial perturbation and
the external disturbance, which guarantees the robustness and
repeatability of the proposed control strategy. Meanwhile, the
active input saturations are demonstrated in Fig. 4, which proves
the effectiveness of the proposed MPC. We can see that the input
constraints are active, which bounds the actual control input
in the range of [−0.6, 0.6], [−0.6, 0.6], and [−0.2, 0.2] m/s for
three axes, respectively. The constraints are set as 60% of UR’s
speed limitation to make sure the constraints are active. On the
other hand, if loose constraints are set, a better dynamic response
and less computational burden can be achieved with a shorter
convergence process.
Fig. 3. Tracking performance of the proposed controller in comparison
with PID and MPC.
C. Experimental Results
1) Case 1: Comparison With Linear-Move Strategy: In
extreme temperatures, where accurate remote control and col- practical implementation, the signum function can be modified
1
laborative robotics based on human experience are needed. The as sign(Δ) = Δ+µ , where μ is an arbitrarily small constant. In
appropriate selection of Qi and Ri should find a tradeoff between this way, the chattering problem can be coped with. Linear-move
the control effort and the performance of the system. The MPC control command [26], [27] is usually employed to drive the
parameters have been selected as Qi = diag(10, 10, 10) and robot move from one waypoint to another in the industrial appli-
Ri = 0.1 for each axis, and the rest of controller parameters are cations. A comparison of the velocity by using the linear-move
shown in Table I. Specifically, Kism,x ≥ Dx , Kism,y ≥ Dy , and control command and the velocity by using the proposed MPC
Kism,z ≥ Dz should be satisfied as they are gains of disturbance is shown in Fig. 5.
rejection terms. Lx , Ly , and Lz are positive constants that Although the path tacking can be realized by the linear move
can be tuned carefully to achieve a satisfactory compensation control command naturally provided by the UR-5e, it should
performance. be noticed that the velocity is unsmooth. During the operation
process, the quality depends on not only the position control
performance but also the velocity. Therefore, the proposed MPC
B. Numerical Monte–Carlo Analysis is more suitable than linear move control for the application
To verify the effectiveness of the proposed control scheme, that needs smooth movement, such as welding and glueing.
a simulation of path tracking by using the proposed control This is because the proposed method can drive the robot for
scheme and MPC only considering initial error, control input path tracking by the velocity regulation based on the real-time
saturation, and external disturbance is demonstrated in Fig. 3. tracking error. Furthermore, the proposed control scheme is
Meanwhile, a PID controller, which is widely used in robotics more reliable as the velocity chattering of the linear move is
control, is introduced as a comparison. avoided for accurate path tracking.
It is shown in Fig. 3 that with input constraints and external 2) Case 2: Comparison With Open-Loop Velocity Con-
disturbances, a smaller tracking error can be achieved by the trol, PID, and MPC: Path tracking experimental results are
proposed MPC. When the robot is driven to track a path with achieved, as shown from Figs. 6 and 7. The desired path is
higher dynamic performance requirement, the superiority of the recorded in the free-drive mode of the UR-5e with a 500 Hz data
developed MPC scheme is more obvious as PID will fail to track exchange rate. To test the robustness of the proposed control
the desired path when input constraints and external disturbance scheme, the installation settings of UR are set incorrect on
are taken into consideration. Compared with MPC, the proposed purpose as the uncertainties of the system, and a1 kg extra
control scheme can further reduce the tracking error as SMC is payload is added during the tracking process as the external
introduced to tackle uncertainties and external disturbances, and disturbance.
guarantee the robustness of the control system. As shown above, compared with PID, both MPC and the
To further test the robustness of the proposed control strategy, proposed control scheme can drive the robot for accurate path
a numerical Monte–Carlo simulation of 200 cases with random tracking. However, when a heavier payload (5 kg) is employed
initial states and multiple reference signals is demonstrated in as the external disturbance, the proposed control scheme can still
Fig. 4. The mean value of the initial position is 0 mm, and the guarantee the tracking performance, as shown in Figs. 8 and 9.

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on September 28,2023 at 13:48:23 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZENG et al.: ROBUST HUMAN–ROBOT COLLABORATIVE CONTROL APPROACH BASED ON MODEL PREDICTIVE CONTROL 7

Fig. 4. Tracking performance of Monte–Carlo simulation with random initial states, uncertainties, and multiple reference signals of 3 axes by using
the proposed controller.

Fig. 5. Velocity comparison between the proposed controller and

point-to-point linear-move control strategy.
Fig. 7. Tracking error by using the proposed controller in comparison
with PID and MPC.

Fig. 6. Tracking trajectory by using the proposed controller in compar-

ison with PID and MPC.

Although PID, MPC, and the proposed method can guarantee

Fig. 8. Tracking trajectory by using the proposed controller in compar-
a smooth path motion compared with linear-move control com- ison with PID and MPC with larger disturbance.
mand, the proposed control scheme can guarantee the accuracy
of the path tracking with robustness against uncertainties and
disturbances. This is because the tracking error is employed as application as the control architecture is closed compared to
the input of the proposed controller and a disturbance compen- simulation and lab environment (e.g., servo motors) where the
sator is introduced in the proposed control scheme so that the controller is open architecture.
robot can be driven accordingly. The robustness provided by 3) Case 3: Path Tracking With Real-Time Human Input:
the proposed control scheme is desirable for industrial robots To modify the recorded path in real time during a process

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on September 28,2023 at 13:48:23 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Fig. 9. Tracking error by using the proposed controller in comparison

with PID and MPC with larger disturbance.
Fig. 11. Tracking performance of 3 axes with joystick real-time input.

Fig. 10. 3-D tracking trajectory with real-time human input by a joy-
stick.

Fig. 12. UR-5e with TIG welding torch.

(e.g., welding) that require high smoothness in path/velocity, a
joystick that is linked with the host PC is introduced in the control
scheme. The tracking performance of three axes is shown from
Figs. 10 and 11.
As shown above, the real-time human intervention can be
achieved with a fast dynamic response, which guarantees the
collaboration between the operator and the robotic system. The
maximum distance that can be modified corresponding to the
original path is set as 10 mm in this study and it can be tuned
according to different scenarios. Meanwhile, it worth to be
noticed that the human input will not cause the jerk movement.
Therefore, the smoothness can be guaranteed by the proposed
human–robot collaborative control system.
4) Case 4: Tungsten Inert Gas (TIG) Welding Test: To
further evaluate the practical effectiveness of the proposed con-
trol scheme, a group of TIG welding tests on curved path is
presented. The welding torch is attached to UR-5e as shown in
Fig. 12. The welding results are shown in Fig. 13.
In Case A, the linear move control based on ten waypoints
Fig. 13. TIG welding results comparison. (a) Welding based on lin-
is employed. The speed vector is set as 1 mm/s during Phase ear move control. (b) Welding based on the proposed control scheme.
A1, which is the desire speed according to experienced welder, (c) Desired path created for case A and B.

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on September 28,2023 at 13:48:23 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZENG et al.: ROBUST HUMAN–ROBOT COLLABORATIVE CONTROL APPROACH BASED ON MODEL PREDICTIVE CONTROL 9

real-time human modification on the outputted tracking perfor-

mance. The path tracking with smoothness was empowered by
the proposed MPC and the sliding-mode disturbance rejection
term was designed to compensate the external disturbance so that
the control performance of the robotic system can be further
improved. Meanwhile, fast dynamic response to the real-time
human input can be guaranteed by the proposed control scheme
to meet the unique requirement of human–robot collaborative
industrial manufacturing and inspection. Simulation and exper-
imental results were demonstrated to verify the effectiveness of
the proposed control strategy.

REFERENCES
[1] M. Ratiu and M. A. Prichici, “Industrial robot trajectory optimization—A
Fig. 14. TIG welding (a) Welding follow the original path. (b) Welding review,” in Proc. MATEC Web Conf., 2017, vol. 126, Art. no. 02005.
with real-time human inputs. [2] K. Guo, Y. Zhang, and J. Sun, “Towards stable milling: Principle and
application of active contact robotic milling,” Int. J. Mach. Tools Manuf.,
vol. 182, 2022, Art. no. 103952.
however, the speed is too low that leads too much wire fed in [3] C. B. Smith, “Robotic friction stir welding using a standard industrial
robot,” Kei Kinzoku Yosetsu (J. Light Metal Weld. Construction), vol. 42,
the weld pool. The widest part is 7.29 mm as shown in the no. 3, pp. 40–41, 2004.
result. The welding result in Phase A2 is achieved with a higher [4] J. N. Pires, A. Loureiro, and G. Bölmsjo, Welding Robots: Technology,
speed but the speed is not constant, which makes the welding System Issues and Application. New York, NY, USA: Springer, 2006.
[5] Q. Chu et al., “In-depth understanding of material flow behavior and
lack of quality. The narrowest part can be only 1.50 mm. This refinement mechanism during Bobbin tool friction stir welding,” Int. J.
is because linear move command focus on position control and Mach. Tools Manuf., vol. 171, 2021, Art. no. 103816.
the speed cannot be constant among waypoints. Note that the [6] V. T. Yen, W. Y. Nan, and P. Van Cuong, “Robust adaptive sliding mode
neural networks control for industrial robot manipulators,” Int. J. Control,
speed can be continuous by using “Move Process” strategy, but Autom. Syst., vol. 17, no. 3, pp. 783–792, 2019.
it will lose the real-time human input during the whole process. [7] J. Na, M. N. Mahyuddin, G. Herrmann, X. Ren, and P. Barber, “Robust
Case B is achieved by using the proposed control scheme with adaptive finite-time parameter estimation and control for robotic systems,”
Int. J. Robust Nonlinear Control, vol. 25, no. 16, pp. 3045–3071, 2015.
1 mm/s as the desired speed. The smooth movement and stable [8] X. Yin and L. Pan, “Enhancing trajectory tracking accuracy for indus-
speed regulation provided by the proposed method guarantees trial robot with robust adaptive control,” Robot. Comput.- Integr. Manuf.,
the quality of the TIG welding with a 4.00 mm fixed width vol. 51, pp. 97–102, 2018.
[9] M. Neubauer, F. Brenner, C. Hinze, and A. Verl, “Cascaded sliding mode
welding result. position control (SMC-PI) for an improved dynamic behavior of elastic
Based on the same desired path shown in Fig. 13(c), TIG feed drives,” Int. J. Mach. Tools Manuf., vol. 169, 2021, Art. no. 103796.
welding based on the proposed control scheme with real-time [10] Z. Jia, Y. Song, and W. Cai, “Bio-inspired approach for smooth motion con-
trol of wheeled mobile robots,” Cogn. Comput., vol. 5, no. 2, pp. 252–263,
human input via a joystick is shown in Fig. 14. We try to add 2013.
real-time human inputs to weld the Branch 1 and 2 shown in [11] Y.-S. Lu and Y.-Y. Lin, “Smooth motion control of rigid robotic manipu-
Fig. 14(a) based on the live via a welding camera shown in lators with constraints on high-order kinematic variables,” Mechatronics,
vol. 49, pp. 11–25, 2018.
Fig. 14(c)–(f). [12] R. Bearee and A. Olabi, “Dissociated jerk-limited trajectory applied to
In this case, the robot follows the original curved path during time-varying vibration reduction,” Robot. Comput.-Integr. Manuf., vol. 29,
Phase B1 and human inputs are added at point P1 in Phase B2 no. 2, pp. 444–453, 2013.
[13] R. Béarée, “New damped-jerk trajectory for vibration reduction,” Control
to follow Branch 1. After welding Branch 1, there is no human Eng. Pract., vol. 28, pp. 112–120, 2014.
input from point P2 and the robot is back to follow the original [14] W. Gao, Q. Tang, J. Yao, and Y. Yang, “Automatic motion planning for
path. During Phase B3, human inputs are added again at point complex welding problems by considering angular redundancy,” Robot.
Comput.- Integr. Manuf., vol. 62, 2020, Art. no. 101862.
P3 to follow Branch 2. It can be seen in Fig. 14(c) that both the [15] C. E. Garcia, D. M. Prett, and M. Morari, “Model predictive control:
positions and rotations of the welding torch have been changed Theory and practice—A survey,” Automatica, vol. 25, no. 3, pp. 335–348,
according to the human input during the welding so that the 1989.
[16] S. Zhan, J. Na, G. Li, and B. Wang, “Adaptive model predictive control
relative position between the welding torch and the wire feed of wave energy converters,” IEEE Trans. Sustain. Energy, vol. 11, no. 1,
can be maintained to guarantee the welding quality. Therefore, pp. 229–238, Jan. 2020.
the experience of the welder can be introduced in the real time, [17] Y. Zhang, S. Zhan, and G. Li, “Model predictive control of wave energy
converters with prediction error tolerance,” IFAC-PapersOnLine, vol. 53,
which is critical for the welding in confined spaces that human no. 2, pp. 12289–12294, 2020.
cannot reach. [18] Y. Liu and Y. Zhang, “Toward welding robot with human knowledge: A
remotely-controlled approach,” IEEE Trans. Autom. Sci. Eng., vol. 12,
no. 2, pp. 769–774, Apr. 2015.
V. CONCLUSION [19] M. Rubagotti, D. M. Raimondo, A. Ferrara, and L. Magni, “Robust
model predictive control with integral sliding mode in continuous-time
A control strategy based on MPC and sliding-mode compen- sampled-data nonlinear systems,” IEEE Trans. Autom. Control, vol. 56,
sator was designed for the path tracking control of industrial no. 3, pp. 556–570, Mar. 2011.
[20] G. P. Incremona, A. Ferrara, and L. Magni, “MPC for robot manipulators
robot. The developed robotic system can be remotely driven by with integral sliding modes generation,” IEEE/ASME Trans. Mechatronics,
the proposed controller for accurate path tracking control with vol. 22, no. 3, pp. 1299–1307, Jun. 2017.

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on September 28,2023 at 13:48:23 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

[21] A. E. K. Mohammad, N. Uchiyama, and S. Sano, “Reduction of electrical Andres Gameros Madrigal received the M.Sc.
energy consumed by feed-drive systems using sliding-mode control with and Ph.D. degrees in manufacturing and
a nonlinear sliding surface,” IEEE Trans. Ind. Electron., vol. 61, no. 6, mechatronic engineering from the Monterrey
pp. 2875–2882, Jun. 2014. Institute of Technology and Higher Education,
[22] T. Zeng, X. Ren, and Y. Zhang, “Fixed-time sliding mode control and high- Monterrey, Mexico, in 2012 and 2016, respec-
gain nonlinearity compensation for dual-motor driving system,” IEEE tively.
Trans. Ind. Informat., vol. 16, no. 6, pp. 4090–4098, Jun. 2020. He has been a Research Fellow with the
[23] N. Uchiyama, Y. Ogawa, A. E. K. Mohammad, and S. Sano, “Energy saving Rolls-Royce University Technology Centre in
in five-axis machine tools using synchronous and contouring control and Manufacturing and On-Wing Technology, Uni-
verification by machining experiment,” IEEE Trans. Ind. Electron., vol. 62, versity of Nottingham, Nottingham, U.K. Cur-
no. 9, pp. 5608–5618, Sep. 2015. rently, he is an Assistant Professor of Intelligent
[24] B. Matthias and T. Reisinger, “Example application of ISO/TS 15066 to Tooling and Specialist Machines with the University of Nottingham.
a collaborative assembly scenario,” in Proc. 47st Int. Symp. Robot., 2016, His research interest includes mechanical design, structural analysis,
pp. 1–5. robotics for inspection/repair, and design of high precision mechatronic
[25] M. Rubagotti, T. Taunyazov, B. Omarali, and A. Shintemirov, “Semi- systems for production.
autonomous robot teleoperation with obstacle avoidance via model pre-
dictive control,” IEEE Robot. Autom. Lett., vol. 4, no. 3, pp. 2746–2753,
Jul. 2019.
[26] K. Miatliuk, A. Wolniakowski, M. Diaz, M. A. Ferrer, and J. J. Quintana,
“Universal robot employment to mimic human writing,” in Proc. 20th Int.
Carpathian Control Conf., 2019, pp. 1–5.
[27] S. Ponsà Cobas, “Strategies for remote control and teleoperation of a UR
robot,” Master’s thesis, Eng. Projects Dept., Universitat Politècnica de
Catalunya, Barcelona, Spain, 2020. Dragos Axinte received the M.Eng. degree
in manufacturing engineering in 1988, and the
Ph.D. degree in manufacturing engineering in
1996.
After graduating, he was with the R&D in in-
dustry for 10 years and then moved to academia
Tianyi Zeng received the B.Eng. degree in con- to lead research in the field of machining, pro-
trol science and engineering from the Harbin cess monitoring and design of innovative tool-
Institute of Technology, Harbin, China, in 2015 ing/robotics for in-situ repair especially related
and the Ph.D. degree in control science and to on-wing repair of aeroengines. He is currently
engineering from the Beijing Institute of Tech- a Professor of Manufacturing Engineering with
nology, Beijing, China, in 2020. the Faculty of Engineering, University of Nottingham, Nottingham, U.K.
He is currently a Research Fellow with the Since 2009, he has been a Director of Rolls-Royce University Tech-
Rolls-Royce University Technology Centre, Uni- nology Centre in Manufacturing and On-Wing Technology, Nottingham,
versity of Nottingham, Nottingham, U.K. His U.K.
research interest includes nonlinear system Prof. Axinte is a Fellow of the International Academy of Production
control, plant/controller codesign, and robotic Engineering (FCIRP). He is also the Editor-in-Chief of the International
control. Journal of Machine Tools and Manufacture.

Abdelkhalick Mohammad received the B.Sc.

degree in mechatronics engineering from As-
siut University, Assiut, Egypt, in 2006, and
the M.Eng. and Ph.D. degrees in robotics and Max Keedwell joined Rolls-Royce, Bristol, U.K.,
mechatronics from the Toyohashi University of in 2011 on the Higher Apprenticeship Scheme.
Technology, Toyohashi, Japan, in 2010 and He had set up the Eurofighter EJ200 engine
2013, respectively. overhaul facility in RR Filton and was subse-
He is currently an Assistant Professor of quently responsible for engine overhaul and
Mechatronics with the Department of Mechan- lineside repairs as a Manufacturing Engineer.
ical, Materials and Manufacturing Engineering, His current role in Capability Acquisition has
University of Nottingham, Nottingham, U.K. His included the development of welding processes
research interests include mechatronics system design, robotics, ma- for high-value components and the industrializa-
chine tool control, and control theory. tion of low-cost robotics into the MRO shops.

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on September 28,2023 at 13:48:23 UTC from IEEE Xplore. Restrictions apply.