Applied Intelligence

https://doi.org/10.1007/s10489-021-02459-3

Formation control of multiple mecanum-wheeled mobile robots

with physical constraints and uncertainties
Dongliang Wang1 · Wu Wei1 · Xinmei Wang2 · Yong Gao1 · Yanjie Li1 · Qiuda Yu1 · Zhun Fan3,4,5,6

Accepted: 20 April 2021

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2021

Abstract
Aiming at the formation control of multiple Mecanum-wheeled mobile robots (MWMRs) with physical constraints and
model uncertainties, a novel robust control scheme that combines model predictive control (MPC) and extended state
observer-based adaptive sliding mode control (ESO-ASMC) is proposed in this paper. First, a linear MPC strategy is
proposed to address the motion constraints of MWMRs, which can transform the robot formation model based on leader-
follower into a constrained quadratic programming (QP) problem. The QP problem can be solved iteratively online by a
delay neural network (DNN) to obtain the optimal control velocity of the follower robot. Then, to address the input saturation
constraints, model uncertainties and unknown disturbances in the dynamic model, an improved ESO-ASMC is proposed
and compared with the robust adaptive terminal sliding mode control (RATSMC) and the conventional sliding mode control
(SMC) to prove the effectiveness. The proposed scheme, considering the optimal control velocity obtained by the kinematics
controller as the given desired velocity of the dynamics controller, can implement precise formation control, while solving
various physical constraints of the robot, and eliminating the effects of model uncertainties and disturbances. Finally, through
a comparative simulation case, the effectiveness and robustness of the proposed method are verified.

Keywords Formation control · Model predictive control · Adaptive sliding mode control ·
Mecanum-wheeled mobile robot · Extended state observer

1 Introduction motion control of mobile robots, such as non-holonomic

wheeled mobile robots [2–5], aircraft [6, 7], flying robot [8,
Extensive research has been conducted on mobile robots 9], autonomous underwater vehicles [10, 11] and omnidirec-
in recent years [1]. Numerous studies have focused on the tional wheeled mobile robots [12, 13]. These strategies pre-
sented in these studies have proven to be extremely effective
 Wu Wei
in many applications. The Mecanum-wheeled mobile robot
weiwu@scut.edu.cn (MWMR) is an omnidirectional mobile robot [14]. Com-
pared with other types of omnidirectional mobile robot, the
1
MWMR has excellent stability and carrying capacity, and
School of Automation Science and Engineering, South China
University of Technology, Guangzhou 510006, China therefore have great application potential in high-end man-
2
ufacturing fields such as aviation, aerospace, and industrial
School of Automation, China University of Geosciences,
Wuhan, Hubei 430074, China
transportation [15].
3
The MWMR is comprised of four symmetrically arranged
School of Department of Electronic and Information
Engineering, Shantou University, 515063, Guangdong, China
Mecanum wheels, and the robot relies on the various veloc-
4
ities of the wheels to achieve omnidirectional movement.
Key Lab of Digital Signal and Image Processing of Guangdong
Province, Shantou University, 515063, Guangdong, China
The MWMR has a complicated wheel train structure, and
5
each Mecanum wheel is distributed with small driven
Key Laboratory of Intelligent Manufacturing Technology
(Shantou University), Ministry of Education, 515063,
rollers, which results in difficulties concerning the pre-
Guangdong, China cise control [14]. In one study [16], the dynamic equation
6 State Key Lab of Digital Manufacturing Equipment
of MWMRs was presented, which was derived using the
& Technology, Huazhong University of Science second Lagrange equation. In a separate study [17], the
and Technology, 430074, Wuhan, China influence of contact force on the rollers of the Mecanum
 D. Wang et al.

wheels of MWMRs was investigated and a theoretical for MWMRs. However, the above-mentioned studies only
model was proposed. Moreover, to effectively predict the consider the kinematic constraints of MWMRs. To achieve
energy consumption of movement process of MWMRs, a an improved control effect in practical applications, the
power consumption model for the movement of MWMRs dynamics must be considered [36].
was proposed [18]. In the work of [19], a two-level con- It is well known that the dynamic model of MWMRs
trol system was designed to realize robot indoor navigation. is highly nonlinear and contains model uncertain terms.
In [20], a fault-tolerant control method based on fuzzy was Additionally, it is necessary to consider the input satu-
proposed for the motion control of MWMRs. The above- ration constraint, which results in a great challenge for
mentioned methods can be used to achieve effective results dynamic control [37]. On the one hand, in order to deal
regarding the control of a single MWMR, however, more with model uncertainties, making the controller adaptive is
complex tasks must often be completed jointly by multi- an effective method [14]. The adaptive sliding mode con-
ple robots. For example, in industry, multiple MWMRs are trol (ASMC) is insensitive to disturbances and parameter
often used in cooperation to transport large object. And the variations and can effectively suppress the jitter due to the
formation control of robots is the premise of the cooperation discontinuity of the control law, thus, it is widely used in the
of multiple robots for the completion of complex tasks [21, design of complex nonlinear system controllers with model
22]. uncertainties [38, 39]. In the work of [14], an adaptive
Formation control of mobile robots addresses the prob- second-order sliding mode control strategy for the trajectory
lem of that multiple mobile robots keep a certain formation tracking control of MWMRs was proposed. In [40], a novel
and move synchronously, which is a hot research direction ASMC method was proposed, introducing adaptive rules
in the field of robotics [23]. Several control strategies aimed based on artificial neural networks to model and estimate
at the general formation control problem of multi-robot sys- various uncertain disturbances. Additionally, considering
tems have been developed, including the leader-follower unknown disturbances and uncertainties, a sliding mode
strategy, virtual structure schemes and behavioral strategy, control method based on extended state observers was pro-
etc [24]. In recent years, numerous in-depth studies have posed for the trajectory tracking of MWMRs can be found
been conducted on the formation control of MWMRs. In in [41]. In [42], a robust adaptive terminal sliding mode
[25], a collision avoidance strategy based on artificial poten- control (RATSMC) strategy was proposed for the trajectory
tial field was integrated into the cooperative formation tracking of an omnidirectional mobile robot. On the other
control of MWMRs. Moreover, to realize the coopera- hand, numerous studies focused on the input saturation
tive transportation of large objects by multiple MWMRs, constraints in the dynamic model. For example, to approx-
a kinematics model of a multi-robot system with velocity imate the unknown input constraint, a smooth non-affine
compensation was established [26]. In the work of [27], function of the input signal was used [43]. Furthermore,
a distributed formation control method based on a fuzzy an unknown dynamic model was obtained through an adap-
wavelet neural network with obstacle avoidance and col- tive controller, and an auxiliary system was designed to
lision avoidance for MWMRs was proposed. In [28], an address the saturation constraints of the actuator in [36].
adaptive non-singular terminal sliding mode control method Although model uncertainties, unknown disturbances,
was proposed for the coordinated formation control prob- or various physical constraints including kinematics and
lem of MWMRs. The existing literature on the formation dynamics constraints was considered in the existing
control of MWMRs demonstrates promising control results, research, few studies on MWMRs considered these factors
nevertheless, the various constraints of MWMRs were not simultaneously. However, ignoring these factors with regard
considered or directly ignored in these methods. Note that to robot control may reduce the stability of the system [44].
the MWMR is a complex mechanical system with vari- Motivated by [34, 41], the combination of MPC and ASMC
ous physical constraints including kinematics and dynamics will be helpful for the design of a formation controller for
constraints. It is well known that ignoring constraints may MWMRs that accounts for the above-mentioned factors.
reduce the robustness of the system [29]. The model pre- In this paper, considering various physical constraints of
dictive control (MPC) strategy has been widely used in MWMRs including kinematics and dynamics constraints,
previous studies to deal with linear and nonlinear systems model uncertainties and unknown disturbances, a robust
under various constraints [30–33]. In [34], in order to real- control strategy combining MPC and ESO-ASMC is pro-
ize the precise trajectory tracking control of MWMRs while posed for the formation control of MWMRs. At the kine-
considering the velocity constraints, a robust MPC strategy matics level, a linear model predictive controller is used
was proposed. In [35], a model predictive fault tolerant con- to deal with the kinematics constraints of MWMRs; at the
trol scheme that satisfied the input constraints was designed dynamics level, an improved ESO-ASMC strategy is used
Formation control of multiple mecanum-wheeled mobile robots with physical...

to deal with the input saturation constraints, model uncer- 2.1 Kinematics
tainties and unknown disturbances in the dynamic model.
Moreover, the comparative studies with the RATSMC in Figure 1 shows the motion schematic of the MWMR in
[42] and conventional sliding mode control (SMC) are the inertial coordinate system XOY . L and l denote the
carried out to verify the effectiveness and superiority. longitudinal and lateral distances of the four Mecanum
In comparison with the existing research on MWMRs, wheels with regard to the center of mass of MWMRs,
the key contributions are as follows: i) A linear formation respectively. The position and direction angle of the
model of MWMRs with constraints is established for the MWMR can be represented as [x y θ ]T . By expressing
first time, and the computational complexity can be reduced the rotation velocities of the four Mecanum wheels as wi ,
by linearizing the formation model. Based on this, a linear i = 1, 2, 3, 4, the velocities of the robot can be expressed
MPC strategy is proposed to generate the desired velocities as [vx vy ω]T . The offset angles of the small rollers on the
of the follower robot. ii) An improved ESO-ASMC method four Mecanum wheels are βi (i = 1, 2, 3, 4), henceforth, a
is proposed to track the desired velocities. The ESO can typical offset angle βi = π/4 is adopted in the paper. The
estimate the unknown dynamics and disturbances, thereby composite velocities of the four Mecanum wheels along the
improving the robustness of the system. Compared with the longitudinal and lateral of MWMRs can be expressed as Vix
RATSMC in [42] and conventional sliding mode control and Viy , respectively. Meanwhile, the translation velocities
(SMC), the ESO-ASMC reduces the jitter of torques and of the four Mecanum wheels are Viw and the tangential
increases the convergence speed of the system. iii) A novel velocities produced by the roller are Vir , i = 1, 2, 3, 4.
robust control scheme that combined MPC and ESO-ASMC It is assumed that the robot does not slide during
is proposed for the formation control. The proposed scheme movement, that is, there is always rolling contact between
can not only effectively address various constraints of the wheel and the ground. By the decomposition and
MWMRs, including kinematics and dynamics constraints, synthesis of the velocities of the body and each Mecanum
but also eliminate the influence of model uncertainties and wheel and the elimination of intermediate variables, the
unknown disturbances. iv) The cooperative transportation of relationship between the velocities of the MWMR and the
multiple MWMRs, relying on precise formation control, has velocities of the four Mecanum wheels can be obtained as
great application potential in industry. The proposed control follows [34]:
strategy in this paper can be applied to various types of ⎡ ⎤
⎡ ⎤ ⎡ ⎤ w1
MWMRs, which can effectively solve a class of engineering vx 1 1 1 1 ⎢ w2 ⎥
issues. ⎣ vy ⎦ = R ⎣ −1 1 1 −1 ⎦ ⎢ ⎥ (1)
4 −1 1 −1 1
⎣ w3 ⎦ ,
The remaining sections of this article are organized ω L+l L+l L+l L+l w4
as follows. The kinematics and dynamics of MWMRs
are modeled in Section 2. In Section 3, a linear model where R represents the radius of the wheels. The transfor-
predictive controller is designed. In Section 4, an adaptive mations of the velocities of the four Mecanum wheels and
sliding mode controller based on an extended state observer the velocities of the MWMR are derived by solving the
is designed, and the stability of the controller is proved generalized inverse of (1) as follows:
by the Lyapunov function. In Section 5, a formation ⎡ ⎤ ⎡ ⎤
control contrast simulation case is presented to prove the w1 1 −1 −(L + l) ⎡ ⎤
⎢ w2 ⎥ ⎢ vx
effectiveness of the proposed control strategy. Finally, in ⎢ ⎥ = 1 ⎢1 1 L+l ⎥ ⎥ ⎣ vy ⎦ .
⎣ w3 ⎦ R ⎣ 1 1 −(L + l) ⎦ (2)
Section 6, the whole article is summarized. ω
w4 1 −1 L + l
By combining formulas (1) and (2), it can be deduced that
2 System model description the following velocity constraint relationship exists during
the movement of the robot:
The MWMR included in this paper has a typical axisym-
metric structure, and four Mecanum wheels are installed w1 + w2 = w3 + w4 . (3)
symmetrically around the robot. The wheels of MWMRs
are driven by four separate motors. During movement, the Remark 1 Equation (3) demonstrates that there is a velocity
wheel will rotate along the axis, and the small rollers constraint, i.e. the sum of the velocities of the first two
distributed on the wheel will rotate in their respective direc- Mecanum wheels should be equal to the sum of the
tions due to ground friction. By controlling the rotation velocities of the latter two Mecanum wheels. To improve
velocities of the four wheels, the MWMR could be moved the robustness of the system, it is necessary to consider this
in any direction on a plane without a change in posture. velocity constraint during movement.
 D. Wang et al.

Fig. 1 The model of MWMRs

Then, through the decomposition and synthesis of and the plane where the robot is located is selected as the
velocity, the kinematic model of the MWMR in the inertial zero potential energy surface. According to the Lagrangian
coordinate system can be obtained as: function, the torques experienced by each Mecanum wheel
⎡ ⎤ ⎡ ⎤ are:
ẋ cos θ − sin θ 0
⎣ ẏ ⎦ = R ⎣ sin θ cos θ 0 ⎦ d ∂K ∂K
4 τi = − , (6)
θ̇ 0 0 1 dt ∂ q̇i ∂qi
⎡ ⎤
⎡ ⎤ w1
1 1 1 1 ⎢ w2 ⎥ where τi , i = 1, 2, 3, 4 is the torque vector of the four
⎣ −1 1 1 −1 ⎦ ⎢ ⎥
⎣ w3 ⎦ . (4) Mecanum wheels, and qi , i = 1, 2, 3, 4 is the angle vector
−1 1 −1 1
L+l L+l L+l L+l of the four Mecanum wheels. Additionally, the robot is
w4
affected by the ground friction, and the friction torque
is proportional to the velocity of the wheel movement,
2.2 Dynamics
therefore, the driving torque of each Mecanum wheel can be
recalculated as:
In this part, to establish the dynamic model of the MWMR,
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
it is assumed that the robot moves on a plane and the τ1 C −D D F q¨1 q̇1
effects of air resistance and the inertia of the small rollers of ⎢ τ2 ⎥ ⎢ −D C F D ⎥ ⎢ q¨2 ⎥ ⎢ q̇2 ⎥
⎢ ⎥=⎢ ⎥⎢ ⎥ ⎢ ⎥
mechanical wheels can be ignored. The total kinetic energy ⎣ τ3 ⎦ ⎣ D F C −D ⎦ ⎣ q¨3 ⎦ + λ ⎣ q̇3 ⎦ , (7)
of the robot during motion includes the rotational kinetic τ4 F D −D C q¨4 q̇4
energy of the robot, the translational kinetic energy of the
robot, and the rotational kinetic energy of each wheel, which where λ = diag(λ1 , λ2 , λ3 , λ4 ) is the viscous friction
can be calculated as [45]: coefficient of each wheel, and

1 
4
1 1 mR 2 Iz R 2
K = m vx 2 + vy 2 + Iz ω2 + Iw wi2 , (5) C= + + Iw ,
2 2 2
i=1
8 16(L + 1)2
Iz R 2
where m represents the mass of the robot, Iz and Iw are D= ,
the moment of inertia of the robot and the moment of 16(L + l)2
inertia of the Mecanum wheels respectively. Since the robot mR 2 Iz R 2
F = − . (8)
moves on a plane, the potential energy remains unchanged, 8 16(L + l)2
Formation control of multiple mecanum-wheeled mobile robots with physical...

By converting (7) into the standard Lagrangian dynamic 3 Design of linear model predictive
equation, the nominal dynamic model of the robot can be controller
obtained as:
3.1 Linearization of the formation model
τ = Mn q̈ + Hn q̇, (9)
The positional relationship between the follower robot RF
where Mn denotes the inertial matrix of the robot, and and the leader robot RL in the inertial coordinate system
Hn denotes the diagonal matrix of the viscous friction is shown in Fig. 2. The position and direction angle of
coefficient of each wheel. However, the actual movement RL and RF are represented by [xl , yl , θl ] and [xf , yf , θf ],
of the robot will be affected by model uncertainties respectively. It is assumed that the parameters of the leader
and unknown external disturbances. Equation (9) can be robot and the follower robot are the same. According to (4),
rewritten to represent a dynamic model with uncertainties in the inertial coordinate system, the kinematics model of
and unknown disturbances, as: RL can be expressed as:
⎡ ⎤ ⎡ ⎤
ẋl cos θl − sin θl 0
⎣ ẏl ⎦ = R ⎣ sin θl cos θl 0 ⎦
(Mn + M)q̈ + (Hn + H )q̇ = τ + ddis , (10)
4
θ˙l 0 0 1
where M and H are the uncertain terms of the dynamic ⎡ ⎤
⎡ ⎤ w1l
equation, and ddis is the unknown disturbances. So (10) can 1 1 1 1 ⎢ w2l ⎥
be rewritten as:
⎣ −1 1 1 −1 ⎦ ⎢ ⎥
⎣ w3l ⎦ , (12)
−1 1 −1 1
L+l L+l L+l L+l w4l
ẇ = Mn−1 τ − Mn−1 Hn w + fdis , (11)
where [w1l , w2l , w3l , w4l ]T are the angular velocities of
four wheels of RL . The state of RF can be expressed as
where fdis = Mn−1 (ddis − M q̈ − H q̇) ∈ R4×1 denotes
the state space expression x˙f f (xf , uf ). By expanding the state
the external disturbances and model uncertainties of robot,
of RF with Taylor series at the trajectory points (xl , ul )
which can be regarded as the total disturbances of the
and ignore the high-order terms [46], it can be obtained that:
system. The vector w = [w1 , w2 , w3 , w4 ]T and wi = q˙i ,
i = 1, 2, 3, 4. ∂f (xl , ul )
ẋf = f (xl , ul ) + (xf − xl )
∂xl
Assumption 1 The uncertain terms M, H and the ∂f (xl , ul )
+ (uf − ul ), (13)
unknown disturbances ddis are all bounded. In addition, the ∂ul
total disturbances fdis is also bounded. where

⎡ ⎤⎡ ⎤
0 0 (w1l + w4l )(cos θl − sin θl ) − (w2l + w3l )(cos θl + sin θl ) xf − xl
∂f (xl , ul ) R
(xf − xl ) = ⎣ 0 0 (w1l + w4l )(cos θl + sin θl ) + (w2l + w3l )(cos θl − sin θl ) ⎦ ⎣ yf − yl ⎦ ,
∂xl 4
00 0 θf − θl
⎡ ⎤
⎡ ⎤ w1f − w1l
cos θl + sin θl cos θl − sin θl cos θl − sin θl cos θl + sin θl ⎢
∂f (xl , ul ) R w2f − w2l ⎥
(uf − ul ) = ⎣ sin θl − cos θl sin θl + cos θl sin θl + cos θl sin θl − cos θl ⎦ ⎢
⎣
⎥,
∂ul 4 w3f − w3l ⎦
− L+l
1 1
− L+l
1 1
L+l L+l w4f − w4l

where [w1f , w2f , w3f , w4f ]T are the angular velocities of The time derivative of formula (14) is calculated to
four wheels of RF . Define xoy is the leader robot coordinate l˙x = ly θ̇l − (ẋl − ẋf ) cos θl − (ẏl − ẏf ) sin θl ,
system, lxy is the line connecting the center of mass of l˙y = −lx θ̇l − (ẋl − ẋf ) sin θl − (ẏl − ẏf ) cos θl , (15)
RL and RF , indicating the relative distance between the θ̇e = wf − wl .
two robots, lx and ly are the projection of lxy in the leader
where wl and wf are the angular velocities of RL and
robot coordinate system, respectively. Then, according to
RF respectively. With the formulas (12)–(15), the linear
the geometric transformation we can get:
leader-follower formation control model can be obtained as:
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ R ⎤
lx = (xf − xl ) cos θl + (yf − yl ) sin θl , l˙x 0 wl R4 a lx 4c
ly = (xl − xf ) sin θl + (yf − yl ) cos θl , (14) ⎣ l˙y ⎦ = ⎣ −wl 0 R b ⎦ ⎣ ly ⎦ + ⎣ R d ⎦ , (16)
4 4
R
θe = θf − θl . θ̇e 0 0 0 θe 4(L+l) e
 D. Wang et al.

Fig. 2 Leader-follower
formation configuration

where (n ≥ 2) follower robots. The state space equation for the for-
mation control of a multi-robot system can be written as:
a = w1l − w2l − w3l + w4l ,
b = w1l + w2l + w3l + w4l , ẋm = Āxm + B̄um . (17)
c = w1f + w2f + w3f + w4f − w1l − w2l
−w3l − w4l , where Ā = diag[A1 , ...Ai , ...An ], and xm = [x1 T , ...xi T ,
d = w1l − w2l − w3l + w4l − w1f + w2f ...xn T ]T denotes the state vector between the follower robots
+w3f − w4f , and the leader robot. Among them, xi = [lix , liy , θie ]T
e = w1l − w2l + w3l − w4l − w1f + w2f denotes the state vector between the ith follower robot and
−w3f + w4f . the leader robot, u̇m = [u1 T , ...ui T , ...un T ]T denotes the
input vector of the follower robots; and ui represents the ith
subject to robot’s input vector.
w1l + w2l = w3l + w4l ,
w1f + w2f = w3f + w4f . 3.2 Model predictive control strategy

Remark 2 The approximately linear formation model (16) The MPC strategy can be expressed as an online iterative
with constraints can be obtained through the expansion of process of solving the optimal solution [47]. To design a
the first-order Taylor series. Linearization of the model can linear model predictive controller to realize the formation
convert the nonlinear MPC to a linear MPC, which can control of multiple MWMRs, the continuous formation
reduce computational complexity [32, 46]. model (17) needs to be discretized at first. Let the sampling
time of the discrete system be Ts , the state equation of its
Equation (16) can be rewritten as the state space equation discrete prediction model can be written as:
as ẋ = Ax + Bu, where the state vector x = [lx , ly , θe ]T , xm (k + 1) = ξ(xm (k)) + η(xm (k))um (k), (18)
and the input vector u = [ R4 c, R4 d, 4(L+l)
R
e]T . Through
controlling the input vector u, the state vector of the system where ξ(xm (k)) = xm (k) + Ts Āxm (k) and η(xm (k)) =
can be converged, so as to realize formation control. Ts B̄. The notation xm (k) and um (k) represent the state
The above-mentioned formation model (16) can be extended vector and the input vector of the system at time k,
to a multi-robot formation model of a leader robot and n respectively. The input vector um (k) = um (k − 1) +
Formation control of multiple mecanum-wheeled mobile robots with physical...

um (k), where um (k) represents the input increment at (18), the predicted state variable at the future moment based
time k obtained at time k − 1. Define the prediction horizon on the current information can be calculated as:
of the system as Np , then the prediction vectors for each ⎡ ⎤ ⎡ ⎤
xm (k + 1 |k) ξ(xm (k |k − 1))
variable can be defined as: ⎢ xm (k + 2 |k) ⎥ ⎢ ξ(xm (k + 1 |k − 1)) ⎥
⎢ ⎥ ⎢ ⎥
⎢ .. ⎥=⎢ .. ⎥
x̃ m (k) = [xm (k + 1 |k)T , xm (k + 2 |k)T , . . . , ⎣ . ⎦ ⎣ . ⎦
xm (k + Np |k) ξ(xm (k + Np − 1 |k − 1))
xm (k + Np |k)T ]T , ⎡ ⎤
η(xm (k |k−1)) ··· 0
ũm (k) = [um (k |k)T , um (k + 1 |k)T , . . . , ⎢ 0 η(xm (k+1 |k−1)) ··· ⎥
(19) ⎢ ⎥
+⎢ .. . ⎥
um (k + Np − 1 |k)T ]T , ⎣ .
. .. .. ⎦
ũm (k) = [um (k |k)T , um (k + 1 |k)T , . . . , 0 ··· η(xm (k+Np −1 |k−1))
⎡ ⎤
um (k)
um (k + Np − 1 |k)T ]T . ⎢ um (k + 1) ⎥
⎢ ⎥
The notation xm (k + j |k) , j = 1, 2, ...Np , denotes ⎢ .. ⎥,
⎣ . ⎦
the predicted state variable at time k + j based on the um (k + Np − 1)
current information at time instant k. Since there are various (21)
physical constraints during the robot’s movement, at any where
time k, the following constraints should be satisfied:
um (k) = um (k − 1) + um (k |k) ,
x̃ mmin ≤ x̃ m (k) ≤ x̃ mmax , um (k + 1) = um (k − 1) + um (k |k) + um (k + 1 |k) ,
um (k + Np − 1) = um (k − 1) + um (k + 1 |k)
ũmmin ≤ ũm (k) ≤ ũmmax , (20)
+ · · · + um (k + Np − 1 |k) .
ũmmin ≤ ũm (k) ≤ ũmmax ,
The predicted state variable can be expressed as follows:
where x̃ mmin , x̃ mmax , ũmmin , ũmmax , ũmmin and ũmmax x̃ m (k) = ξ̃ + η̃um (k − 1) + η̃Mũm (k), (22)
are the lower and upper bounds of the corresponding
variable respectively. According to the prediction model where

⎡ ⎤ ⎡ ⎤
ξ(xm (k |k − 1)) 1 ··· 0
⎢ ξ(xm (k + 1 |k − 1)) ⎥ ⎢1 1 ···⎥
⎢ ⎥ ⎢ ⎥
ξ̃ = ⎢ .. ⎥, M=⎢. .. .. ⎥ ,
⎣ . ⎦ ⎣ .. . . ⎦
ξ(xm (k + Np − 1 |k − 1)) 1 ··· 1
⎡ ⎤
η(xm (k |k − 1)) ··· 0
⎢ 0 η(xm (k + 1 |k − 1)) ··· ⎥
⎢ ⎥
η̃ = ⎢ .. .. .. ⎥.
⎣ . . . ⎦
0 ··· η(xm (k + Np − 1 |k − 1))

Then, the cost function of the system can be expressed as where Q and P are the weighting matrix of the corre-
the following constrained optimization problem: sponding dimension. After several algebraic operations,
the optimization problem (23) can be rendered into the
following constrained standard QP problem:
min J (ũm (k), x̃ m (k)) = ||ξ̃ + η̃um (k − 1) (23)
+η̃Mũm (k)||2Q +||ũm (k)||2P , 1
min ũm (k)T Y ũm (k) + γ T ũm (k), (24)
2
subjected to
subject to
x̃ mmin ≤ ξ̃ + η̃um (k − 1) + η̃Mũm (k) ≤ x̃ mmax ,
ũmmin ≤ ũm (k) ≤ ũmmax , Sũm  ψ,
 D. Wang et al.

where by summing the input increment at time instant k and the

Y = 2((η̃M)T Qη̃M + P ), input vector at time instant k − 1. Then through (16), the
γ = 2(η̃M)T Q(ξ̃ + η̃um (k − 1)), desired velocity of the follower robot can be calculated.
T
S = −Ẽ Ẽ −η̃M η̃M ,
⎡ ⎤
−ũmmin + ũm (k − 1)
⎢ ⎥ 4 Design of adaptive sliding mode controller
ũmmax − ũm (k − 1)
ψ =⎢ ⎥
⎣ −x̃ mmin + ξ̃ + η̃um (k − 1) ⎦ ,
In Section 3, the desired velocity of the follower robot in the
x̃ − ξ̃ − η̃um (k − 1)
⎡ mmax ⎤ formation control can be calculated. However, in the actual
E 0 ··· 0 movement of the robot, the influence of its own model
⎢E E ··· 0 ⎥
⎢ ⎥ uncertainties and external disturbances can not be ignored.
Ẽ = ⎢ . . . . ⎥ .
⎣ .. .. . . .. ⎦ Inspired by previous work [41], an improved ESO-ASMC
E E E E approach is proposed to track the desired velocity of the
follower robots.
Among them, E is the identity matrix of appropriate dimen-
sion. In the design of MPC, Q and P are selected as positive
4.1 Design of extended state observer
definite, thus, Y is also positive definite. Therefore, the QP
problem (24) can be seen as a convex optimization problem,
Considering the input saturation constraints of the actuator
and the solution is unique.
of the MWMR, the dynamic model (11) of the MWMR can
be rewritten as:
3.3 Neurodynamic optimization approach
ẇ = Mn−1 sat (τ ) − Mn−1 Hn w + fdis , (29)
In this section, the constrained QP problem is solved online where sat (τ ) = [sat (τ1 ), sat (τ2 ), sat (τ3 ), sat (τ4
)]T , and
through a delay neural network [34]. The QP problem (24) sat (τi ) can be defined as:
is equivalent to the following piecewise linear equation:
1 − sgn(τi − τimin )
sat (τi ) = τimin
PΦ (μ − α(Gμ + p)) = μ, (25) 2
1 + sgn(τi − τimax )
where + τimax (30)
  2
γ Y ST 1 − sgn(τi − τimin )sgn(τi − τimax )
p= ,G = . +
−ψ −S 0 2
τi ,
The upper and lower bounds of decision vector μ in (25) where τimin and τimax are the lower and upper bounds of
can be expressed as: the input, respectively. In real-life robot control, disturbance
T T fdis is often not directly available. In this paper, an
μ = ũm κ , μmax = ũmmax +κ + , μmin
extended state observer is used to approximate the unknown
T
= ũmmin −κ − . (26) term fdis . Define the extended state vector containing the
disturbances fdis as Z = [wT , fdis T ]T . Then, the extended
The convex set Φ can be be designed as Φ =
state vector can be written as:
μmin ≤ μ ≤ μmax . Where PΦ (μi ) is a saturation function
which can be defined as: ˙ ,
Ż = Az Z + Bz sat (τ ) + Hz fdis (31)
1 − sgn(μi − μimin ) where
PΦ (μi ) = μimin   
2 −Mn−1 Hn E1 Mn−1 04×4
1 + sgn(μi − μimax ) Az = , Bz = , Hz = .
+ μimax , (27) 04×4 04×4 04×4 E1
2
1 − sgn(μi − μimin )sgn(μi − μimax ) Among them, 04×4 represents zero matrix, and E1 ∈
+ μi R4×1 represents identity matrix. Define Ẑ = [ŵ T , fˆdis ]T
T
2
where sgn(·) is the sign function. Then, to solve the linear is the estimate state of Z, where ŵ and fˆdis are the estimates
projection (25) quickly, the DNN can be designed as: of w and fdis . Then, the extended state observer Ẑ can be
designed as:
μ̇ = −2μ + PΦ (μ − α(Gμ + p)) + μ(t − ε), (28)
Ẑ˙ = A Ẑ + B sat (τ ) + K w̃,
z z z (32)
where the transmission delay ε should satisfy: ε ≥ 0, and α
is a positive constant. In the actual control, by solving (28), where
  
the input increment at time instant k of the controller can be −Mn−1 Hn E1 Mn−1 Kz1
Az = , Bz = , Kz = .
obtained. The input vector at time instant k can be obtained 04×4 04×4 04×4 Kz2
Formation control of multiple mecanum-wheeled mobile robots with physical...

And w̃ = w − ŵ represents the velocity estimation error compensate, which can improve the stability and robustness
of the system. The matrix Kz1 and Kz2 are the observer of the system.
gain matrixes. Define the disturbance estimation error as
f˜dis = fdis − fˆdis = [f˜1dis , f˜2dis , f˜3dis , f˜4dis ]T . Then 4.2 Adaptive sliding mode control
combine (31) and (32), it yields
In this part, an improved ASMC based on the extended state
Z̃˙ = Ãz Z̃ + B̃z fdis
˙ , (33) observer (32) is implemented to address the input saturation
constraints and track the desired velocity of the four wheels
where
  of the follower robot. Define sat (τ ) = τ − τ , then (29)
w̃ −Mn−1 Hn − Kz1 E1 can be expressed as:
Z̃ = , Ãz = ,
f˜dis −Kz2 04×4
 ẇ = Mn−1 τ − Mn−1 Hn w + fdis − Mn−1 τ , (36)
04×4
B̃z = . where τ = τ − sat (τ ). Then, the error signals of four
E1
Mecanum wheels are defined as:

˙ e1 = w
Assumption 2 The first derivative of disturbances fdis d −w. (37)
˙  ≤ ς.
exist and are bounded, it can be expressed as fdis e2 = e1 dt
where w = [w1 , w2 , w3 , w4 ]T denote the actual velocities
Theorem 1 For the state observer (32), by designing the
of each Mecanum wheel, and wd = [w1d , w2d , w3d , w4d ]T
values of the observer gain matrices Kz1 and Kz2 so that
denote the desired velocities of w. Then the sliding surface
Ãz is Hurwitz, then the observation error (33) is bounded
˙ . function is designed as:
stable for any bounded fdis
s = e1 + K1 e2 , (38)
Proof Due to matrix Ãz is Hurwitz, for any symmetric where s = [s1 , s2 , s3 , s4 and K1 = diag(k11 , k12 , k13 ,
]T
positive definite matrix Q1 there exists a symmetric positive k14 ), k1i > 0, i = 1, 2, 3, 4. The compound reaching law
T
definite matrix P1 that satisfies the equation Ãz P1 + can be designed as:
P1 Ãz = −Q1 . Define a positive definite Lyapunov function
T ṡ = −K2 sat (s) − K3 s δ , (39)
as V = Z̃ P1 Z̃, then we have
where sat (s) = [sat (s1 ), sat (s2 ), sat (s3 ), sat (s4 )]T . And
V̇ = Z̃˙ T P1 Z̃ + Z̃ P1 Z̃˙
T
sat (si ), i = 1, 2, 3, 4, can be designed as: sat (si ) =
⎧
˙ )T P1 Z̃ + Z̃ T P1 (Ãz Z̃ + B̃z fdis
= (Ãz Z̃ + B̃z fdis ˙ ) ⎨1 , si /χ > 1
si /χ , −1 ≤ si /χ ≤ 1 , where χ > 0 is the thickness of
T T
˙ )T P1 Z̃ + Z̃ T P1 Ãz Z̃
= Z̃ Ãz P1 Z̃ + (B̃z fdis (34) ⎩
−1 , si /χ < −1
T
˙ the boundary layer.
+ Z̃ P1 B̃z fdis
The variable s δ = [s1δ , s2δ , s3δ , s4δ ]T , where siδ = si −
T
˙
= −Z̃ Q1 Z̃ + 2Z̃ P1 B̃z fdis
T
χsat (si ). Then, the following lemmas can be obtained.
Taking Q1 as a identity matrix, then
Lemma 1 s Tδ ṡ δ = s Tδ ṡ.
T
˙
T
V̇ = −Z̃ Z̃ + 2Z̃ P1 B̃z fdis
Lemma 2 siδ sat (si ) = |siδ |.
T
˙ 2
T
≤ −Z̃ 2 · Z̃2 + 2Z̃ 2 · P1 B̃z fdis (35)
T
˙ 2 ) The diagonal matrix K2 = diag(k21 , k22 , k23 , k24 ), k2i
= −Z̃ 2 (Z̃2 − 2P1 B̃z fdis
> 0, i = 1, 2, 3, 4. The diagonal matrix K3 = diag(k31 ,
T
It is obvious that −Z̃ 2 < 0. To make V̇ < 0, only k32 , k33 , k34 ), k3i > 0, i = 1, 2, 3, 4. The values of K2 and
˙ 2 . So, Z̃ is bounded, which
need Z̃2 > 2P1 B̃z fdis K3 can be designed by the following adaptive rate:
⎧ ⎧
implies fˆdis can approximate the true value of fdis . This ⎪ k21˙ = Υ1 s1δ sat (s1 ) ⎪ ˙ = ζ1 s 2
k31
⎪
⎨ ˙ ⎪
⎨ ˙ 1δ
end the proof. k22 = Υ2 s2δ sat (s2 ) k32 = ζ2 s2δ
2
˙ , ˙ = ζ3 s 2 , (40)
⎪
⎪ k23 = Υ3 s3δ sat (s3 ) ⎪
⎪ k33
⎩ ˙ ⎩ ˙ 3δ
Remark 3 Note that when the total disturbance term fdis k24 = Υ4 s4δ sat (s4 ) k34 = ζ4 s4δ
2
is too large, the output torque of the robot motor will cause
where Υi and ζi are constants greater than zero. Then, the
large chattering, resulting in the failure of the robot to move
time derivative of (38) can be calculated as
smoothly [41]. When designing the sliding mode controller,
the disturbance estimation value of the ESO can be used to ṡ = e˙1 + K1 e1 . (41)
 D. Wang et al.

From (29)–(41), the sliding mode control law can be = s Tδ (−K2 sat (s) − K3 s δ + fˆdis − fdis )
designed as: 4 4
+ (k2i − k̄2i )siδ sat (si ) + (k3i − k̄3i )siδ
2
τ = Mn [K1 e1 + w˙d +K2 sat (s)+K3 s δ − fˆdis ]+Hn w+τ . i=1 i=1
(42) 4 4
=− k̄2i siδ sat (si ) − 2
k̄3i siδ − s Tδ f˜dis
Remark 4 To ensure the convergence rate of the controller i=1 i=1
while weakening the effects of chattering, compared 4 4
with the conventional sliding mode control, the following ≤ − k̄2i siδ sat (si ) − 2
k̄3i siδ + |s Tδ | · |f˜dis |
improvements are made: i) An extended state observer is i=1 i=1
used to estimate the disturbance for compensation. ii) The 4
model parameters are modified to be adaptive and saturation ≤ − k̄2i siδ sat (si ) + |s Tδ | · |f˜dis |
functions are used instead of traditional symbolic functions i=1
in the design of the reaching law (39). Moreover, to ensure a 4
smooth trajectory of the robot, when designing parameters, =− k̄2i |siδ | + |s Tδ | · |f˜dis |
K3 should be increased while K2 should be decreased. i=1
4
Theorem 2 With input saturation constraints, unknown dis- = (−k̄2i |siδ | + |siδ | · |f˜idis |). (44)
turbances and model uncertainties considered, the adaptive i=1
sliding mode control law (42) based on ESO (32) can make
the closed loop system stable.
where |s Tδ | = [|s1δ |, |s2δ |, |s3δ |, |s4δ |] and |f˜dis | =
[|f˜1dis |, |f˜2dis |, |f˜3dis |, |f˜4dis |]T . To make V̇ < 0, just need
Proof Define vector K̄2 = [k21 ¯ , k22
¯ , k23
¯ , k24
¯ ]T and K̄3 = to design K̄2 so that |f˜idis | < k̄2i , i = 1, 2, 3, 4. It can
[k31 ¯ , k33
¯ , k32 ¯ , k34
¯ ] . The notation k¯2i and k¯3i are the
T be deduced that when t goes to infinity, s δ goes to zero.
nominal values of k2i and k3i respectively, i = 1, 2, 3, 4. When s δ goes to zero, it is easy to prove s converges to
Where k¯2i and k¯3i are both positive constants. Then, the inside of the boundary layer. Therefore, the system is
select the following positive definite function as Lyapunov stable. The complete steps of the developed robust control
function. method combining MPC and ESO-ASMC are shown in
Algorithm 1.
4 4
1 T 1 1
V = s sδ + (k2i −k̄2i )2 + (k3i −k̄3i )2 , (43)
2 δ 2Υi 2ζi
i=1 i=1

The time derivative of (43) is calculated to

4 4
1 1
V̇ = s Tδ ṡ + (k2i − k̄2i )k˙2i + (k3i − k̄3i )k˙3i
Υi ζi
i=1 i=1
4
= s Tδ (K1 e1 + w˙d − ẇ) + (k2i − k̄2i )siδ sat (si )
i=1
4
+ (k3i − k̄3i )siδ
2
x
i=1
= s Tδ (K1 e1 + w˙d − Mn−1 τ + Mn−1 Hn w
−fdis + Mn−1 τ )
4 4
+ (k2i − k̄2i )siδ sat (si ) + (k3i − k̄3i )siδ
2

i=1 i=1
= s Tδ (K1 e1 + w˙d − Mn−1 (Mn (K1 e1 + w˙d + K2 sat (s) + K3 s δ
−fˆdis ) + H n w + τ ) + M −1 Hn w − fdis + M −1 τ )
n n
4 4
+ (k2i − k̄2i )siδ sat (si ) + (k3i − k̄3i )siδ
2

i=1 i=1
Formation control of multiple mecanum-wheeled mobile robots with physical...

5 Simulation is the same as (42). The parameters of RATSMC are

selected as K2c = diag(0.1, 0.1, 0.1, 0.1), K3c = diag
In this section, a comparative simulation case considering (5, 5, 5, 5), c1 = c2 = 0.01, σ1 = 2, σ2 = 0.6, ν = 0.5 and
the physical constraints, external disturbances and model ς1 = −25.
uncertainties is used to verify the effectiveness and The objective of the simulation case is to use the robust
robustness of the proposed algorithm. The control effects of control method proposed in this paper to control the robots
the ESO-ASMC proposed in this paper are compared with in formation. To verify the robustness of the proposed
the conventional SMC and RATSMC. The simulation case algorithm, the multi-robot formation trajectory is designed
uses two identical MWMRs containing a follower robot and as a time-varying sinusoidal trajectory. Then, the angular
a leader robot. The main parameters and definition of the velocities of each wheel of the leader robot can be set to
formation system are shown in Table 1. follow the cosine function. The angular velocities of the four
The main parameters of the MPC are chosen as Np = 3, wheels of the leader robot are set as follows:
Q = 1200 I1 and P = 200 I2 , where I1 and I2 are the ⎧
⎪
⎪ w1l = 0.875 − 0.625 cos(0.1t)rad/s
identity matrices of appropriate dimension. The main ⎨
w2l = 0.875 + 0.625 cos(0.1t)rad/s
parameters of the ESO-ASMC are chosen as Kz1 = . (47)
⎪
⎪ w = 0.875 + 0.625 cos(0.1t)rad/s
diag(−20, −20, −20, −20), Kz2 = diag(−150, −150, ⎩ 3l
w4l = 0.875 − 0.625 cos(0.1t)rad/s
−150, −150) and K1 = diag(2, 2, 2, 2). The sliding
mode control adaptive rates are set as Υ1 = Υ2 = The formation trajectory is presented in Fig. 3, where RL
Υ3 = Υ4 = 0.2 and ζ1 = ζ2 = ζ3 = ζ4 = 0.1, and RF represent the leader robot and the follower robot,
respectively. Moreover, the total disturbances fdis are set as respectively. The initial coordinate and direction angle of
[− sin(0.2t), sin(0.2t), − cos(0.2t), cos(0.2t)]T . RL are (0,0,0), while the initial coordinate and direction
The conventional sliding mode controller is designed as angle of RF are (-0.2,0,0). The desired reference trajectory
follows: of RF is obtained via MPC. It is obvious that in the
presence of disturbances, the ESO-ASMC has almost no
τ = Mn [K1 e1 + w˙d + K2 sgn(s) + K3 s] + Hn w. (45)
tracking deviation. The trajectory of the follower robot can
The definition of all parameters is the same as in (42). completely overlap with the desired reference trajectory
The main parameters of the conventional SMC are selected when using the ESO-ASMC and maintain the desired
as K1 = diag(2, 2, 2, 2), K2 = diag(0.1, 0.1, 0.1, 0.1) and relative distance with the leader robot. In addition, it can
K3 = diag(1, 1, 1, 1). be observed from the partial enlarged figure in Fig. 3 that
The robust adaptive terminal sliding mode controller is the RATSMC can track the desired trajectory but has a
designed as follows: certain tracking error. And the conventional SMC can not
effectively track the desired trajectory, and it shows a large
τ = Mn [K2 |s|ν sgn(s) + K3 s + w˙d + c1 sig σ1 e1 deviation from the desired trajectory. The associated system
+c2 sig σ2 e1 − fˆdis ] + Hn w, (46) states are depicted in Fig. 4. It can be observed that the states
 of the formation system rapidly converge to corresponding
where s = e1 + (c1 sig σ1 e1 + c2 sig σ2 e1 ), and e1 = [e11 , desired value. The evolution of the optimal control input
e12 , e13 , e14 ]T . The function sigσ e1 = [|e11 |σ sgn(e11 ), |e12 |σ increments of model predictive controller is shown in Fig. 5.
sgn(e12 ), |e13 |σ sgn(e13 ), |e14 |σ sgn(e14 )]T . The adaptive It can be observed that the input increments converge to a
˙
control law fˆdis = ς1 s. The definition of other parameters stable value within several seconds.
The velocity constraint and the partial enlarged figure
Table 1 Parameter values of the formation system (3 ≤ t ≤ 7) of RF are presented in Figs. 6 and 7, where
[w1d , w2d , w3d , w4d ]T and [w1 , w2 , w3 , w4 ]T denote the
Definition Symbol Value Unit
desired angular velocities and the actual angular velocities
Mass of robot m 300 kg of the four wheels of RF , respectively. The results presented
Total control time Tc 80 s in Fig. 7 demonstrate that the constraint relationship (3) is
Sampling period Ts 0.1 s not satisfied in the initial stage of the formation when using
Radius of wheel R 0.05 m the SMC and RATSMC, whereas, the velocity constraint
Moment of inertia of robot Iz 200 kg·m2 relationship is always constant during the formation process
Moment of inertia of wheel Iw 5 kg·m2
when using the ESO-ASMC.
Longitudinal distance of wheels to centre L 0.8 m
The input torques are illustrated in Fig. 8, where [τ1 , τ2 ,
Lateral distance of wheels to centre l 0.5 m
τ3 , τ4 ]T are the control input torques of the four wheels
of RF . The input torques show an obvious impulse in the
Viscous friction coefficient of wheel λi 0.5 None
initial stage, however, they never exceed the torque limit
 D. Wang et al.

0.8 9
RL RATSMC SMC w 1d+ w 2d
8
0.6 R F reference ESO-ASMC w 3d+ w 4d
7
ESO-ASMC w 1+ w 2
0.4 6 ESO-ASMC w 3+ w 4

velocity (rad/s)
5 RATSMC w 1+ w 2
y (m)

0.2 RATSMC w 3+ w 4
4
SMC w 1+ w 2
0 3
SMC w 3+ w 4
-0.1
-0.11 2
-0.2 -0.12
2.04 2.06 2.08 1

-0.5 0 0.5 1 1.5 2 2.5 3 3.5 4 0

0 10 20 30 40 50 60 70 80
x (m)
time (s)
Fig. 3 Comparison of formation trajectory
Fig. 6 Velocity constraint versus time plot

0.25 2

0.2
1.8
0.15 lx ly e
1.6 w 1d+ w 2d
0.1
velocity (rad/s)

w 3d+ w 4d
0.05 1.4
state

ESO-ASMC w 1+ w 2
0 ESO-ASMC w 3+ w 4
1.2
-0.05 RATSMC w 1+ w 2
1 RATSMC w 3+ w 4
-0.1
SMC w 1+ w 2
-0.15 0.8
SMC w 3+ w 4
-0.2
0 10 20 30 40 50 60 70 80 0.6
3 3.5 4 4.5 5 5.5 6 6.5 7
time (s)
time (s)
Fig. 4 The states of formation system
Fig. 7 Velocity constraint versus time plot (3 ≤ t ≤ 7)

0.16
50
u1 u2 u3
0.14
1 2 3 4
0.12
control torque (N m)

0.1
input increment

0.08
0
0.06

0.04

0.02

0
-50
-0.02
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
time(s) time (s)

Fig. 5 The optimal control input increments of MPC Fig. 8 The control torques of ESO-ASMC
Formation control of multiple mecanum-wheeled mobile robots with physical...

range, indicating that the proposed method can effectively 6

solve the input saturation constraint problem. After a short ESO-ASMC

adjustment process, the input torques can rapidly converge 5 RATSMC
to a stable range and maintain stable throughout the SMC

control torque (N m)
formation process. The comparison of the control torques of
4 5.6
the four wheels is presented in Figs. 9, 10, 11, 12, 13, 14, 15
and 16. The partial enlarged figures in Figs. 9, 11, 13 and 15 5.4
3
demonstrate that when the ESO-ASMC is used, the control
39.8 40 40.2 40.4
torques reach stable values more rapidly compared with
the SMC and RATSMC. To compare the control torques 2

of the ESO-ASMC, RATSMC and SMC in greater depth,

the evolution of the control torques (31 ≤ t ≤ 41) is 1
presented in Figs. 10, 12, 14 and 16, respectively. It is
worth noting that when using the SMC and RATSMC, the
0
control torques will also fluctuate to a certain extent when 31 32 33 34 35 36 37 38 39 40 41
the system reaches a steady state, but there is almost no time (s)
fluctuation when using the ESO-ASMC. This implies the
Fig. 10 Comparison of control torque of wheel 1 (31 ≤ t ≤ 41)
ESO-ASMC proposed in this paper can reduce torque jitter.
Figures 17, 18, 19 and 20 show the disturbance
50
estimations of the four wheels of RF . It is worth noting
that the disturbance estimations are extremely close to ESO-ASMC
the disturbances, indicating that the designed ESO can RATSMC
SMC
control torque (N m)

approximate estimate the disturbances. Compared with

the RATSMC, the ESO-ASMC has superiorities in the
accuracy of the disturbance estimations and the disturbance
0
error convergence rate. To evaluate the accuracy of the
10
disturbance estimation of ESO-ASMC and RATSMC, the
integral of the absolute of disturbance estimation error and 0
the integral of square of the disturbance estimation error are
tabulated in Table 2. -10
3 4 5 6
The velocity errors of the four wheels of RF under the
-50
ESO-ASMC, RATSMC and SMC are shown in Figs. 21,
0 10 20 30 40 50 60 70 80
22, 23 and 24, respectively. It is worth noting that in
time (s)
the presence of disturbances, the velocity errors of the
SMC fluctuates greatly, and the RATSMC also has slight Fig. 11 Comparison of control torque of wheel 2

1
30
ESO-ASMC
ESO-ASMC
25
RATSMC 0 RATSMC
SMC SMC
20
control torque (N m)

4
control torque (N m)

-1
15
3
10 -4.4
-2
2.8 3 3.2 3.4 3.6 3.8
5 -4.6
-3
0 -4.8
39.5 40 40.5
-5
-4
-10

-15 -5
0 10 20 30 40 50 60 70 80 31 32 33 34 35 36 37 38 39 40 41

time (s) time (s)

Fig. 9 Comparison of control torque of wheel 1 Fig. 12 Comparison of control torque of wheel 2 (31 ≤ t ≤ 41)
 D. Wang et al.

3
50
ESO-ASMC ESO-ASMC
2
RATSMC
RATSMC
SMC 1 SMC
control torque (N m)

control torque (N m)
0

0 -4.3
-1
20 -4.4
-2 -4.5
10 31.8 32 32.2 32.4 32.6
-3
0
3 4 5 6
-4
-50
0 10 20 30 40 50 60 70 80 -5
31 32 33 34 35 36 37 38 39 40 41
time (s)
time (s)
Fig. 13 Comparison of control torque of wheel 3
Fig. 16 Comparison of control torque of wheel 4 (31 ≤ t ≤ 41)

6 1.5
ESO-ASMC Disturbance ESO-ASMC RATSMC
5
RATSMC
1
SMC
4
control torque (N m)

disturbance estimation

0.5
3 5.4

2 0
5.2
1
31.8 32 32.2 32.4 32.6 -0.5
0
-0.2
-1
-1 -0.3
32 32.5 33
-2 -1.5
31 32 33 34 35 36 37 38 39 40 41 0 10 20 30 40 50 60 70 80
time (s) time (s)

Fig. 14 Comparison of control torque of wheel 3 (31 ≤ t ≤ 41) Fig. 17 The disturbance estimation of wheel 1

6
30 Disturbance
4
ESO-ASMC ESO-ASMC
20
RATSMC 2 RATSMC
SMC
disturbance estimation

0
control torque (N m)

10 -2
1
-4
0
-6 0.8
-10 -2.5
-8
-3 0.6
-3.5
-10
-4
-20 4 5 6
-4.5
-12
3 4 5
-30 -14
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
time (s) time (s)

Fig. 15 Comparison of control torque of wheel 4 Fig. 18 The disturbance estimation of wheel 2
Formation control of multiple mecanum-wheeled mobile robots with physical...

5 Table 2 Comparison of disturbance estimation error

Disturbance ESO-ASMC RATSMC
Controller ESO-ASMC RATSMC
0
Error convergence [0.70, 0.80, 1.60, 1.90]T [4.10, 5.70, 5.50, 3.80]T
disturbance estimation

time
Integral absolute [1.38, 1.38, 1.73, 1.73]T [3.77, 14.4, 19.1, 4.28]T
-5
0 error
Integral square [0.03, 0.03, 0.29, 0.29]T [0.28, 82.7, 152, 0.79]T
-0.5 error
-10

-1

-15 -1.5 respectively. It can be easily found from the partial enlarged
3 4 5 6 figures in Figs. 29–32 that when the ESO-ASMC is used,
the sliding surfaces reach stable states faster than when the
-20
0 10 20 30 40 50 60 70 80 RATSMC and SMC are used. Moreover, when the system
time (s) reaches a steady state, the fluctuation range of the sliding
surfaces of the ESO-ASMC is smaller than that of the
Fig. 19 The disturbance estimation of wheel 3
SMC and RATSMC. It is indicating that in the presence of
disturbances, the sliding surfaces of the ESO-ASMC can
fluctuations. However, the velocity errors of the ESO- converge to the inside of the boundary layer and remain
ASMC can completely converge to zero, which suggests stable. In order to compare the control effects of the three
that the ESO-ASMC proposed in this paper can achieve control methods in greater depth, the integral of absolute
precise velocity tracking control. It is evident from the of velocity error(IAVE), the integral of square of velocity
partial enlarged figures in Figs. 22 and 23 that although both error (ISVE), the integral of absolute of angle error(IAAE)
control methods can achieve velocity tracking control, the and the integral of square of angle error (ISAE) of the
ESO-ASMC has a faster convergence rate than that of the entire formation process are calculated. Table 3 presents the
RATSMC and SMC. specific numerical comparison data. The results show that
Figures 25, 26, 27 and 28 show the comparison of the the ESO-ASMC has advantages in IAVE, ISVE, IAAE and
angle errors of the follower robot under different control ISAE compared with the RATSMC and SMC. It is worth
methods. It can be found that after the initial acceleration noting that when the ESO-ASMC is used to achieve robot
of the movement, the angle errors can converge to zero formation, the effect is far better than using the RATSMC
when the ESO-ASMC is used. The effect of using ESO- and SMC.
ASMC is better than RATSMC and SMC. The comparison Based on the above-mentioned simulation results, the
of sliding surfaces is shown in Figs. 29, 30, 31 and 32, following conclusions can be drawn: i) The designed

2 0.3
1
Disturbance ESO-ASMC
1.5 ESO-ASMC RATSMC
0.2
0.8 RATSMC SMC
disturbance estimation

1
velocity error (rad/s)

0.1
1 2 3 4
0.5
0
0

-0.1
-0.5

-1 -0.2

-1.5 -0.3
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
time (s) time (s)
Fig. 20 The disturbance estimation of wheel 4 Fig. 21 Comparison of velocity error of wheel 1
 D. Wang et al.

2
ESO-ASMC ESO-ASMC RATSMC SMC
1.5 RATSMC 0.4
SMC
1
velocity error (rad/s)

0.2

angle error (rad)

0.5

0 0

-0.5 -0.2
0.5
-1
0
-0.4
-1.5 -0.5
2 4 6
-2 -0.6
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
time (s) time (s)

Fig. 22 Comparison of velocity error of wheel 2 Fig. 25 Comparison of angle error of wheel 1

2.5 0.6

ESO-ASMC ESO-ASMC RATSMC SMC

2
RATSMC 0.4
1.5 SMC
velocity error (rad/s)

0.2
angle error (rad)

0.5
0
0

-0.5 -0.2
0.5
-1 0
-0.4
-1.5 -0.5
2 4 6
-2 -0.6
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
time (s) time (s)

Fig. 23 Comparison of velocity error of wheel 3 Fig. 26 Comparison of angle error of wheel 2

0.2 0.8
ESO-ASMC RATSMC SMC
0.6
0.1

0.4
velocity error (rad/s)

0
angle error (rad)

0.2
-0.1
0

-0.2
-0.2
ESO-ASMC
-0.3 RATSMC -0.4
SMC
-0.4 -0.6
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
time (s) time (s)

Fig. 24 Comparison of velocity error of wheel 4 Fig. 27 Comparison of angle error of wheel 3
Formation control of multiple mecanum-wheeled mobile robots with physical...

2.5
ESO-ASMC RATSMC SMC ESO-ASMC
2
0.4 0.2 RATSMC
0
1.5 -0.2 SMC
-0.4
0.2 2 4 6
angle error (rad)

0.5

s3
0
0

-0.2 -0.5

-1
-0.4
-1.5

-0.6 -2
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
time (s) time (s)

Fig. 28 Comparison of angle error of wheel 4 Fig. 31 Comparison of sliding surface s3

1 ESO-ASMC RATSMC SMC linear model predictive controller can effectively address
the kinematic constraints of MWMRs, and the designed
ESO-ASMC can effectively deal with the input torque
0.5 saturation constraints and eliminate the influence of
model uncertainties and unknown disturbances. ii) The
proposed ESO can approximate the unknown dynamics
s1

0
parameters and disturbances. iii) Considering various
physical constraints, model uncertainties and unknown
0.02 external disturbances, the robust control strategy proposed
-0.5
in this paper that combines MPC and ESO-ASMC can
0
achieve accurate formation Control of multiple MWMRs.
-1 -0.02 The trajectory and torque of the robot are smooth, and
2 3 4
chattering is extremely minimal. iv) Under the same initial
0 10 20 30 40 50 60 70 80
conditions, the simulation results show that the obvious
time (s)
improved performance of the ESO-ASMC compared with
Fig. 29 Comparison of sliding surface s1 the RATSMC and conventional SMC.

2.5
ESO-ASMC 1 ESO-ASMC RATSMC SMC
2
0.2 RATSMC
1.5 0 SMC
-0.2
-0.4 0.5
1 2 4 6

0.5
s2

s4

0
0

-0.5
-0.5 0.02
-1 0

-1.5 -0.02
-1
2 3 4
-2
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
time (s) time (s)

Fig. 30 Comparison of sliding surface s2 Fig. 32 Comparison of sliding surface s4

 D. Wang et al.

Table 3 Formation control

performance comparison Controller ESO-ASMC RATSMC SMC

IAVE [0.08, 1.21, 1.47, 0.10]T [0.48, 2.14, 2.39, 0.55]T [4.44, 5.74, 5.81, 4.75]T
ISVE [0.01, 1.03, 1.46, 0.02]T [0.01, 1.82, 2.25, 0.03]T [0.31, 1.51, 1.62, 0.38]T
IAAE [0.08, 0.50, 0.69, 0.08]T [2.01, 2.85, 3.04, 2.05]T [21.0, 21.6, 21.8, 20.9]T
ISAE [0.00, 0.13, 0.28, 0.00]T [0.06, 0.24, 0.38, 0.07]T [7.19, 7.45, 7.78, 7.09]T

6 Conclusion References

In this paper, we propose a novel robust control approach 1. Bencherif A, Chouireb F (2019) A recurrent tsk interval type-
combining MPC and ESO-ASMC for the formation control 2 fuzzy neural networks control with online structure and
parameter learning for mobile robot trajectory tracking. Applied
of multiple MWMRs. This paper focuses on implementing Intelligence
precise formation control, while addressing various physical 2. Dirik M, Castillo O, Kocamaz AF (2019) Global path planning
constraints of the robot including kinematics and dynamics and Path-Following for wheeled mobile robot using a novel
constraints, and eliminating the effects of model uncer- control structure based on a vision sensor. Int J Fuzzy Syst
22(6):1880–1891
tainties and unknown disturbances. To overcome control 3. Dirik M, Kocamaz AF, Castillo O (2020) Visual-Servoing Based
challenges, first, this paper establishes a multi-robot forma- global path planning using interval type-2 fuzzy logic control.
tion model that includes constraint relations based on the Axioms 8(58):1–16
kinematics model. Then, the torque saturation constraint, 4. Castillo O, Leticia AA, Castro JR, Garcia-Valdez M (2016) A
comparative study of type-1 fuzzy logic systems, interval type-2
the uncertainty of model parameters and unknown distur- fuzzy logic systems and generalized type-2 fuzzy logic systems in
bances are considered in the dynamic model. Next, in the control problems. Inf Sci:257–274
control realization part of the study, an MPC strategy is 5. Sanchez MA, Castillo O, Castro JR (2015) Generalized Type-2
used to generate the desired velocities of the follower robot Fuzzy Systems for controlling a mobile robot and a performance
comparison with Interval Type-2 and Type-1 Fuzzy Systems.
while solving the kinematic constraints, and an ESO is Expert Syst Appl 42(14):5904–5914
used to estimate the unknown dynamics parameters and 6. Castillo O, Cervantes L, Soria J, Sanchez M, Castro JR (2016) A
disturbances. Finally, an adaptive sliding mode controller generalized type-2 fuzzy granular approach with applications to
is designed based on the ESO to track the desired veloci- aerospace. Inf Sci:165–177
7. He H, Qi W, Kao Y (2021) HMM-based adaptive attack-resilient
ties while addressing the torque saturation constraint, and control for Markov jump system and application to an aircraft
the stability of the proposed controller is proved via the model. Appl Math Comput 392
Lyapunov function. The comparative study shows that the 8. Phan HV, Park HC (2020) Mimicking nature’s flyers: a review of
ESO-ASMC proposed in this paper shows more effective insect-inspired flying robots. Curr Opin Insect Sci 42:70–75
9. Lambert NO, Schindler CB, Drew DS, Pister KSJ (2021)
control than the RATSMC and SMC. Simulation results and Nonholonomic Yaw Control of an Underactuated Flying Robot
theoretical analysis are used to validate the accuracy and With Model-Based Reinforcement Learning. IEEE Robot Autom
effectiveness of the proposed method. Lett 6(2):1–7
Future research will focus on the formation control of 10. Nikou A, Verginis CK, Heshmati-alamdari S, Dimarogonas DV
(2020) A robust non-linear MPC framework for control of
multiple MWMRs with obstacle and collision avoidance in underwater vehicle manipulator systems under high-level tasks. Iet
complex dynamic environments, and the method proposed Control Theory Appl 15(3):323–337
in the present study will be applied to actual engineering. 11. Jia G, Li D, He B (2021) Intelligent Collaborative Navigation and
An additional topic that will be explored in greater depth is Control for AUV Tracking. IEEE Trans Indust Inf 17(3):1–1
12. Saenz A, Santibaez V, Bugarin E, Dzul A, Rios H, Villalobos-
the control of MWMRs for the completion of tasks with a Chin J (2021) Velocity Control of an Omnidirectional Wheeled
reduction in energy consumption. Mobile Robot Using Computed Voltage Control with Visual
This work was supported by the National Natural Science Feedback: Experimental Results. Int J Control Autom Syst 19:1–
Foundation of China (grant numbers 61573148, 61603358), 14
13. Taheri H, Zhao CX (2020) Omnidirectional mobile robots,
the Science and Technology Planning Project of Guang- mechanisms and navigation approaches. Mechanism and Machine
dong Province, China (grant numbers 2015B010919007, Theory
2019A050520001, 180917144960530, 2019A050519008), 14. Alakshendra V, Chiddarwar SS (2017) Adaptive robust control
the Project of Educational Commission of Guangdong of mecanum-wheeled mobile robot with uncertainties. Nonlinear
Dyn 87(4):2147–2169
Province of China (grant number 2017KZDXM032), and 15. Niu Z, Lu Q, Cui Y, Sun Z (2019) Fuzzy Control Strategy for
the State Key Lab of Digital Manufacturing Equipment & Course Correction of Omnidirectional Mobile Robot. Int J Control
Technology (grant number DMETKF2019020). Autom Syst 17(9):2354–2364
Formation control of multiple mecanum-wheeled mobile robots with physical...

16. Hendzel Z, Rykala L (2017) Modelling of Dynamics of a Wheeled 36. Chen Y, Li Z, Kong H, Ke F (2019) Model predictive
Mobile Robot with Mecanum Wheels with the use of Lagrange tracking control of nonholonomic mobile robots with coupled
Equations of the Second Kind. Int J Appl Mech Eng 22(1):87–89 input constraints and unknown dynamics. IEEE Trans Ind Inf
17. Bayar G, Ozturk S (2020) Investigation of The Effects of 15(6):3196–3205
Contact Forces Acting on Rollers Of a Mecanum Wheeled Robot. 37. Baek J, Jin M, Han S (2016a) A new adaptive sliding-mode control
Mechatronics 72 scheme for application to robot manipulators. IEEE Trans Indust
18. Zhang L, Kim J, Sun J (2019) Energy Modeling and Experimental Electron 63(6):3628–3637
Validation of Four-Wheel Mecanum Mobile Robots for Energy- 38. Zhang Q, Dong J (2019) Disturbance-observer-based adaptive
Optimal Motion Control. Symmetry 11(11):1372 fuzzy control for nonlinear state constrained systems with input
19. Qian J, Zi B, Wang D, Ma Y, Zhang D (2017) The design and saturation and input delay. Fuzzy Sets and Systems
development of an omni-directional mobile robot oriented to an 39. Boukattaya M, Mezghani N, Damak T (2018) Adaptive nonsingu-
intelligent manufacturing system. Sensors 17(9):2073 lar fast terminal sliding-mode control for the tracking problem of
20. Alshorman AM, Alshorman O, Irfan M, Glowacz A, Muhammad uncertain dynamical systems. Isa Trans:1–19
F, Caesarendra W (2020) Fuzzy-Based Fault-Tolerant Control for 40. Lu X, Zhang X, Zhang G, Fan J, Jia S (2019) Neural network
Omnidirectional Mobile Robot. Machines 8(55) adaptive sliding mode control for omnidirectional vehicle with
21. Agarwal M, Agrawal N, Sharma S, Vig L, Kumar N (2015) uncertainties. Isa Trans 86:201–214
Parallel multi-objective multi-robot coalition formation. Expert 41. Yuan Z, Tian Y, Yin Y, Wang S, Liu J, Wu L (2019) Trajectory
Syst Appl 42(21):7797–7811 tracking control of a four mecanum wheeled mobile platform: an
22. Wang Y, Shen T, Song C, Zhang Y (2020) Circle formation extended state observer-based sliding mode approach. Iet Control
control of second-order multi-agent systems with bounded Theory Appl 14(3):415–426
measurement errors. Neurocomputing 42. Feng X, Wang C (2020) Robust adaptive terminal sliding mode
23. Dai S, He S, Chen X, Jin X (2020) Adaptive leader-follower control of an omnidirectional mobile robot for aircraft skin
formation control of nonholonomic mobile robots with prescribed inspection. Int J Control Autom Syst:1–11
transient and steady-state performance. IEEE Trans Indust Inf 43. Chen Q, Tang X, Nan Y, Ren X (2017) Finite-time neural funnel
16(6):3662–3671 control for motor servo systems with unknown input constraint. J
24. Liu H, Wang Y, Lewis FL (2021) Robust Distributed Formation Syst Sci Complex 30(3):579–594
Controller Design for a Group of Unmanned Underwater Vehicles. 44. Huang Y, Cao Q, Leng C (2010) The path-tracking controller
IEEE Trans Syst Man Cybern Syst 51(2): 1215–1223 based on dynamic model with slip for one four-wheeled omr. Ind
25. Gronemeyer M, Horn J (2019) Collision avoidance for coop- Robot Int J 37(2):193–201
erative formation control of a robot group. IFAC-PapersOnLine 45. Vlantis P, Bechlioulis CP, Karras GC, Fourlas GK, Kyriakopoulos
52(8):434–439 KJ (2016) Fault tolerant control for omni-directional mobile
26. Li Y, Ge S, Dai S, Zhao L, Shi Y (2019) Kinematic modeling platforms with 4 mecanum wheels. pp 2395–2400
of a combined system of multiple mecanum-wheeled robots with 46. Wu H, Si Z, Li Z (2020) Trajectory tracking control for four-wheel
velocity compensation. Sensors 20(1):75 independent drive intelligent vehicle based on model predictive
27. Tsai C, Wu H, Tai F, Chen Y (2017) Distributed consensus control. IEEE Access 8(1):1–11
formation control with collision and obstacle avoidance for 47. Xiao H, Chen CLP (2019) Leader-follower consensus multi-
uncertain networked omnidirectional multi-robot systems using robot formation control using neurodynamic-optimization-based
fuzzy wavelet neural networks. Int J Fuzzy Syst 19(5):1375–1391 nonlinear model predictive control. IEEE Access 7:43581–43590
28. Tsai CC, Yu CC, Wu CW (2019) Adaptive distributed BLS-
FONTSM formation control for uncertain networking hetero-
geneous omnidirectional mobile multirobots. J Chin Inst Eng
43(2):171–185
29. Huynh HN, Verlinden O, Wouwer AV (2017) Comparative
Publisher’s note Springer Nature remains neutral with regard to
application of model predictive control strategies to a wheeled
jurisdictional claims in published maps and institutional affiliations.
mobile robot. J Intell Robot Syst 87(1):81–95
30. Dubois L, Suzuki S (2018) Formation control of multiple quad-
copters using model predictive control. Adv Robot 32(19):1037–
1046
31. Francisco M, Mezquita Y, Revollar S, Vega P, De Paz JF
(2019) Multi-agent distributed model predictive control with fuzzy
negotiation. Expert Syst Appl 129:68–83 Dongliang Wang received his
32. Yue M, An C, Sun J (2018) An efficient model predictive control bachelor degree in Beijing
for trajectory tracking of wheeled inverted pendulum vehicles with University of Chemical Tech-
various physical constraints. Int J Control Autom Syst 16(1):265– nology, China. He is currently
274 pursuing the Ph.D. from the
33. Hu Y, Su H, Fu J, Karimi HR, Knoll A (2020) Nonlinear School of Automation Sci-
Model Predictive Control for Mobile Medical Robot using Neural ence and Engineering, South
Optimization. IEEE Trans Ind Electron 99:1–1 China University of Technol-
34. Wang D, Wei W, Yeboah Y, Li Y, Gao Y (2019) A robust ogy. His research focus on
model predictive control strategy for trajectory tracking of omni- robot motion planning and
directional mobile robots. J Intell Robot Syst:1–15 control, Multi-robot collabo-
35. Karras GC, Fourlas GK (2020) Model predictive fault tolerant ration, intelligent control.
control for omni-directional mobile robots. J Intell Robot Syst
97:1–21
 D. Wang et al.

Wu Wei received a Ph.D. Yanjie Li is currently pur-

degree from Huazhong Uni- suing the Ph.D. from the
versity of Science and Tech- School of Automation Sci-
nology in 2000. From January ence and Engineering, South
2001 to December 2002, he China University of Tech-
was a postdoctoral researcher nology, Guangzhou, China.
at the Department of Automa- Her research focus on robot
tion at Tsinghua University. motion planning and control.
From 2003 to 2004, he worked
as a researcher at the Insti-
tute of Intelligent Systems,
The Hong Kong Polytechnic
University. He is also a mem-
ber of the International IEEE
Control Systems, Automotive
Technology, Signal Process-
ing, System People and Control Systems, Robotics and Automation Qiuda Yu received his bach-
Association. He is currently as a professor at the School of Automation elor degree in Hangzhou
Science and Engineering, South China University of Technology. His Dianzi University, Zhejiang,
research focus on intelligent control, intelligent systems, robot control China. He is currently pursu-
technology, pattern recognition and artificial intelligence. ing the Ph.D. from the School
of Automation Science and
Engineering, South China
Xinmei Wang received her University of Technology.
B.S. and M.S. degrees from His research focus on visual
Wuhan University of Technol- grasp, visual guided control
ogy, and Ph.D. degree from and deep-learning.
South China University of
Technology in 2009. She is
currently a lecturer at School
of Automation, China Uni-
versity of Geosciences. Her
research interests include sta-
bility analysis and control Zhun Fan received a Ph. D.
of time-delay system, robot degree from electrical engi-
visual servoing and switching neering, Michigan State Uni-
system. versity, USA, in 2004. From
2004 to 2007, he worked as
an assistant professor at the
Technical University of Den-
Yong Gao received his bach-
mark, Denmark. From 2007 to
elor’s degree from Hubei
2011, he was an associate pro-
Normal University, in 2016.
fessor at the Department of
He is currently pursuing
Mechanical Engineering and
a Ph.D. in the School of
the Department of Manage-
Automation Science and
ment Engineering at the Tech-
Engineering from South
nical University of Denmark.
China University of Technol-
He is currently as a professor
ogy. His principal research
at the College of Engineering,
interests are bionic robot
Shantou University. His research interest covers artificial intelligence,
technology including multi-
mechatronics design automation, intelligent robot system, computer
legged robot and snake-like
vision, and evolutionary computation.
robot, motion planning and
intelligent control.