Posture-Based Leader-Follower Formation For Multiple Mobile Robot Systems Using Data-Based Adaptive Iterative Learning Control

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

Posture-based Leader-follower Formation for

Multiple Mobile Robot Systems Using Data-based


Adaptive Iterative Learning Control
Zhuolei Chaochen Yaonan Wang Haoran Tan*
College of Electrical and College of Electrical and College of Electrical and
Information Engineering Information Engineering Information Engineering
2022 China Automation Congress (CAC) | 978-1-6654-6533-5/22/$31.00 ©2022 IEEE | DOI: 10.1109/CAC57257.2022.10055775

Hunan University Hunan University Hunan University


National Engineering Research National Engineering Research National Engineering Research
Center for Robot Visual Center for Robot Visual Center for Robot Visual
Perception and Control Perception and Control Perception and Control
Changsha, China Changsha, China Changsha, China
chaochenzhuolei@hnu.edu.cn yaonan@hnu.edu.cn tanhaoran@hnu.edu.cn

Jiaming Zhou Qiong Lin


College of Electrical and College of Electrical and
Information Engineering Information Engineering
Hunan University Hunan University
National Engineering Research National Engineering Research
Center for Robot Visual Center for Robot Visual
Perception and Control Perception and Control
Changsha, China Changsha, China
zhouj1am1ng@hnu.edu.cn linqiong@hnu.edu.cn

Abstract—Multiple mobile robot systems (MMRSs) have great behavioral-based [5], virtual structure [6], and leader-follower
potential in various fields and their reliable application depends methods [7], [8]. Each of them has its pros and cons. The
on accurate formation among robots. Focusing on the formation behavioral-based method describes the behavior of mobile
problem of MMRSs, a posture-based adaptive iterative learning
control scheme (PB-AILC) is proposed in this paper. Firstly, robots but is difficult to analyze mathematically, which leads
the leader-follower formation model is built. Then PB-AILC to a lack of formation accuracy. As for the virtual structure
scheme is introduced to optimize the control linear and angular method, a virtual structure is built to represent the entire
velocity of follower robots at each iteration based on the robots’ formation instead of assigning desired trajectories to every
global posture only, guaranteeing that followers can steadily track single robot. However, the virtual structure method seems easy
their leader with prescribed relative distance and orientation.
Meanwhile, we analyze its feasibility mathematically. Finally, to realize but leads to a lack of flexibility and adaptability. In
numerical simulations illustrate the validity of the PB-AILC the leader-follower method, generally, each robot in MMRSs
scheme and we compare PB-AILC with an iterative learning is classified as either leader or follower. The leader works
control (ILC) method to demonstrate its superiority. according to the predefined trajectory, while the followers
Index Terms—adaptive control, formation control, iterative utilize accessible information to adjust their position in order
learning control, multiple mobile robot systems
to track the leader with prescribed offsets. Only through the
I. I NTRODUCTION control of followers, the leader-follower method can achieve
the formation that leaders and followers work with a specific
Multiple mobile robot systems (MMRSs) have become a
relative angle and distance from each other. It allows simple
significant research topic in the industry area in recent years
implementation and extensible application which is why it is
due to their expandable applications in synergetic processing
of particular interest to us.
[1], group hunting [2], and cooperative transportation [3].
Traditional leader-follower formation methods usually use
The accurate formation control of the MMRS is the premise
sliding mode control, fuzzy control, and proportional-integral
of completing the corresponding tasks safely and effectively,
(PI) control to design the controller. Tutuko [9] used an interval
which leads to the problem we solved here. The formation
fuzzy type-2 controller for the leader-follower control problem
control schemes are usually divided into three types [4]:
of mobile robots. Zhao [10] took into account both kinematics
Haoran Tan is the corresponding author (email: tanhaoran@hnu.edu.cn). and dynamics in controller design based on the sliding mode

978-1-6654-6533-5/22/$31.00 ©2022 IEEE 6463

Authorized licensed use limited to: UNIVERSITE DE CARTHAGE. Downloaded on June 25,2024 at 11:13:41 UTC from IEEE Xplore. Restrictions apply.
control. A PI predictive control strategy was employed in [11] II. P ROBLEM F ORMULATION
multi-agent systems (MAS). All of the above methods can be A. Leader–Follower Kinematics
categorized as a kind of progressive control implemented in
As shown in Fig. 1, The basic leader-follower setup consists
the time domain, which generally leads to low accuracy of for-
of one leader robot Rl and one follower robot Rf .
mation. Thus, intelligent control methods have been proposed.
Hou [12] introduced iterative learning control (ILC) to deploy
the tracking algorithm of followers. The tracking accuracy was
improved with the increasing number of iterations. However,
due to the constancy of ILC’s gain parameters, the complexity
and uncertainty of a strong nonlinear MMRS are difficult
to capture accurately and comprehensively. To refine this
shortcoming, a model-free adaptive control method (MFAC)
[13] is proposed by Hou where a pseudo-partial derivative
(PDD) was updated online in every iteration and utilized to
quantify the nonlinearity and uncertainty dynamically of the
system. On this basis, Tan [14] proposed a networked model-
free adaptive predictive control method for nonlinear MASs.
Chi [15] combined MFAC and ILC to construct a model-free
adaptive iterative learning control method (MFAILC) recently.
Just like ILC algorithm, MFAILC is also constructed in the
iteration domain but its learning gain can be optimized online
based on the adaptive adjustment of PPDs’ value, which breaks Fig. 1. The basic leader-follower setup in world coordinates system.
the previously mentioned shortcoming of ILC. It has captured
widespread attention within both academia and industry. In First, we model each robot individually, without considering
the field of MASs control theory research, Bu [16] proposed the side-slip of tires, the kinematics of each robot Rk , k ∈
a novel distributed MFAILC for the coordination problem. In {l, f} at time t can be described by:
the medical field, Ai [17] improved the tracking performance ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
xk (t + 1) xk (t) cos θk (t) 0  
of pneumatic artificial muscles using a model-free adaptive ⎣ yk (t + 1) ⎦ = ⎣ yk (t) ⎦ + ⎣ sin θk (t) 0 ⎦ vk (t)
iterative learning controller with high-order pseudo partial ωk (t)
θk (t + 1) θi (t) 0 1
derivative. (1)
where vk (t) and ωk (t) are respectively the linear and angular
Inspired by the above approach, we propose a PB-AILC
velocities of robot Rk . xk (t), yk (t), θk (t) are global posture
leader-follower formation scheme. The information obtained
of robot Rk which means x-y position and orientation respec-
from previous iterations modifies unsatisfactory control signals
tively. Define the output and input vectors as:
at the current iteration to enhance tracking accuracy. No model
information is used but only the robots’ global posture is ok (t) = [xk (t), yk (t), θk (t)]T , uk (t) = [vk (t), ωk (t)]T (2)
required and the introduction of PPD provides the adaptive
capability of PB-AILC. Finally, numerical simulations show Then the global posture difference value between Rl and Rf
its advancement and effectiveness. The main points of this is denoted as: olf (t) = ol (t) − of (t) = [xl (t) − xf (t), yl (t) −
paper are as follows: yf (t), θl (t) − θf (t)]T .
In formation task, follower Rf tracks leader Rl with a fixed
(1) PB-AILC scheme focuses on the MIMO MMRSs instead relative tracking distance ddlf and angle ϑdlf . According to Fig.
of SISO MASs. 1, we define a tracking error ef of Rf at time t as follows:

(2) As a data-based method, PB-AILC works based on the ef (t) = odf (t) − of (t)
⎧ d  
robots’ global posture of robots only. ⎨ dlf sin θlf (t) + ϑdlf − π2  + xlf (t) (3)
= dd cos θlf (t) + ϑdlf − π2 + ylf (t)
(3) Draw upon the MFAILC, PB-AILC breaks the limitation ⎩ lf
θlf (t)
of the constant gain parameters during the iterative process by
using PPD. It means being able to describe the nonlinearity Equation. (3) tells us that there is a nonlinear relationship
of the MMRS. between olf and ef , which we describe it as a function Ψ (·).
So the (3) can be simplified as:
The rest of this paper is organized as follows. Section ef (t) = Ψ(olf (t)) (4)
II models the system and illustrates the scheme. Numerical
simulations are presented to demonstrate the efficiency and Our goal is to design suitable controllers to adjust the
superiority of PB-AILC scheme in Section III. Section IV are input signal uf (t) in order to achieve limt→∞ ef (t) =
conclusions of this paper. limt→∞ Ψ(olf (t)) = , where  denotes a minimum value.

6464

Authorized licensed use limited to: UNIVERSITE DE CARTHAGE. Downloaded on June 25,2024 at 11:13:41 UTC from IEEE Xplore. Restrictions apply.
It needs to be clarified that the kinematics model we Prove Theorem 1 as follows.
constructed here is only for producing the output information Proof: Expanding the (7) is that
which we need for numerical simulations. There is no reliance
Δof (i, t + 1)
on mathematical models for practical applications.
= of (i, t + 1) − of (i − 1, t + 1)
B. PB-AILC scheme design = g(of (i, t), uf (i, t)) − g(of (i, t), uf (i − 1, t)) (10)
+ g(of (i, t), uf (i − 1, t))
Designing PB-AILC scheme based on the above. Assuming
time range in [0, T] and iterations number I. During the − g(of (i − 1, t), uf (i − 1, t))
iteration i ∈ {1, I}, at moment t ∈ {1, T}, we describe the Since Assumption 2 is satisfied, the above (10) can continue
nonlinear mobile robot model as follows according to (2): to be written as
of (i, t + 1) = g(of (i, t), uf (i, t)) (5) ∂g ∗
Δof (i, t + 1) = (u (i, t) − uf (i − 1, t))
ˆ f (i, t) f
∂u (11)
And (5) it satisfies assumptions 1 and 2: + ξf (i, t)
Assumption 1: The partial derivatives of g(·) to control inputs
uf (i, t) is continuous. where ∂g ∗ \∂u ˆ f (i, t) is the value of g(·)’s gradient ma-
Assumption 2: The (5) is generalized Lipschitz on the itera- trix to uf (i, t), ξf (i, t) = g (of (i, t), uf (i − 1, t)) −
tion axis with i ∈ {1, I} and t ∈ {1, T}, g (of (i − 1, t), uf (i − 1, t)).
Import a matrix ζ f (i, t) with the same dimension as
Δof (i, t + 1) ≤ b Δuf (i, t) (6) ∂g ∗ \∂uˆ f (i, t),the following equation is given:

where ξf (i, t) = ζ f (i, t)Δuf (i, t) (12)

Δof (i, t + 1) = of (i, t + 1) − of (i − 1, t + 1) Because of Δuf (i, t) = 0, the (12) is necessarily solvable
(7) which means there exists a solution of ζ f (i, t) that makes (13)
Δuf (i, t) = uf (i, t) − uf (i − 1, t)
satisfied
and b is a positive constant number. ∂g∗
Φf (i, t) = + ζ f (i, t) (13)
Remark 1: The reasonability of assumptions 1 and 2 for ∂uf (i, t)
mobile robot systems has been discussed in [18]. Combining (12) and (13), (14) can be further reformulated as:
Remark 2: The change rate Δof is restricted when uf
increases by assumption 2, which makes sure an infinite Δof (i, t + 1)
variation of mobile robots’ posture despite the finite variation ∂g ∗
= (u (i, t) − uf (i − 1, t)) + ξf (i, t)
of linear and angular velocities, and it is apparently in accord ∂u ˆ f (i, t) f
with practical situations. ∂g ∗
Theorem 1: For (5) satisfying assumptions 1 and 2, under the = Δuf (i, t) + ζ f (i, t)Δuf (i, t) (14)
ˆ f (i, t)
∂u
premise that Δuf (i, t) = 0, a Pseudo Jacobi Matrix (PJM) + ,
Φf (i, t) which has the same function as PPD but in a matrix ∂g ∗
= + ζ f (i, t) Δuf (i, t)
form is introduced as follows (The term (i, t) is omitted from ˆ f (i, t)
∂u
the equation (8)):
= Φf (i, t)Δuf (i, t)
⎡ ⎤
φ11,f · · · φ1m,f · · · φ1M,f as shown in (9).
⎢ .. .. .. .. .. ⎥ Combining ef (i − 1, t + 1) mentioned in (4),the learning
⎢ . . . . . ⎥
⎢ ⎥
Φf = ⎢⎢ φ n1,f · · · φ nm,f · · · φ nM,f

⎥ (8) control algorithm is defined as follows:
⎢ . . . . . ⎥ T
⎣ .. .. .. .. .. ⎦ ρΦ̂f (i, t)ef (i − 1, t + 1)
uf (i, t) = uf (i − 1, t) + - -2
φN1,f · · · φNm,f · · · φNM,f - -
λ + -Φ̂f (i, t)-
then the (5) can be reformulated by the dynamical lineariza- T
ρΦ̂f (i, t)Ψ(olf (i − 1, t + 1))
tion: = uf (i − 1, t) + - -2 .
- -
Δof (i, t + 1) = Φf (i, t)Δuf (i, t) (9) λ + -Φ̂f (i, t)-
(15)
Obviously in (8), N = 3, M = 2, which represent the From (15), we can find that at moment t, the i-th control signal
dimensions of output and input respectively. of follower is determined by the control signal uf (i − 1, t)
Remark 3: Φf (i, t) is a complicated representation of the and the posture difference value olf (i − 1, t + 1) between
system’s nonlinearity, complexity, and unmodelable features leader and follower. ρ and λ > 0 is a controller parameter
of the MMRS. The sign of its elements is unchanged. and a weight factor respectively which need to be fixed

6465

Authorized licensed use limited to: UNIVERSITE DE CARTHAGE. Downloaded on June 25,2024 at 11:13:41 UTC from IEEE Xplore. Restrictions apply.
appropriately. Φ̂f (i, t) is the estimated value of Φf (i, t) by
(16)
Φ̂f (i, t) = Φ̂f (i − 1, t)
f (i − 1, t)
ηΔuT
+ 2 (Δof (i − 1, t + 1) (16)
μ + uf (i − 1, t)
−Φ̂f (i − 1, t)Δuf (i − 1, t)
where η is the step factor, μ > 0 is a weight factor. In
order to make sure that the Δuf (i, t) = 0 for any i iteration,
conditions are formulated to constrain each element φnm,f of
matrix Φf (i, t) :
φ̂nm,f (i, t) = φ̂nm,f (0), n = 1, . . . , N, m = 1, . . . , M Fig. 3. The desired formation trajectory based on seted initial value.
. . . . (17)
. . . .
if .φ̂nm,f (i, t). ≤  or .φ̂nm,f (i, t). ≥ W

where φ̂nm,f (0),  and W are predefined initial value, a mini-


mum value and a large value respectively.
Theorem 2: Under the satisfaction of assumptions 1 and 2,
the PB-AILC scheme we proposed in (15)-(17) can guarantee
that limt→∞ ef (t) = limt→∞ Ψ(olf (t)) =  holds and
of (i, t), uf (i, t) are bounded.
Proof : See Appendix C in Hou [19].
III. S IMULATION R ESULTS
In numerical simulations, we chose five mobile robots
whose interaction topology is shown in Fig. 2 in formation.
Robot L is used as the leader, robots F1 and F2 are the fol-
lowers of robot L, while robots F3 and F4 are the followers of
robots F1 and F2 respectively. Suppose the leader’s trajectory
is a circle:

⎨ xl (t) = cos(πt)
yl (t) = sin(πt) , t ∈ [0, 2] (18)

θl (t) = πt + π/2

Fig. 2. Topological relationships among five mobile robots.

Fig. 4. Formation results of PB-AILC scheme in different iterations.


Clearly, the initial position of the leader in the world
coordinates is (1, 0). We set the initial coordinates of fol-
lower 1-4 respectively to be (1.1, 0), (0.9, 0), (1.3, −0.1) and [v1,2,3,4 (i, 0), ω1,2,3,4 (i, 0)]T = [2, 3]T and four fixed parame-
(0.8, −0.2). The relative tracking distance and angle are ddl1 = ter of each follower are respectively set as η = 1, μ = 10, λ =
ddl2 = dd13 = dd24 = 0.1, ϑdl1 = ϑdl2 = 30◦ , ϑd13 = ϑd24 = 45◦ . 5, ρ = 1. The initial value of PJM is as follow:
The desired formation trajectory is shown in Fig. 3, where the  
solid black line demonstrates the trajectory of robot L. The red, 0.5 1 0
Φ1,2,3,4 (i, 0) = (19)
blue, yellow and green dashed lines represent the trajectories 0 0 10
of robot 1-4 respectively. Points with their corresponding color Sampling time is 0.01s.
represent the initial positions. Fig. 4 shows how the formation results change as the itera-
Then the initial conditions of PB-AILC scheme are as tions are increasing. Clearly that with the increase of iterations,
follows. The initial input is given as u1,2,3,4 (i, 0) = the formation tends to increasingly fit the desired. In Fig. 5,

6466

Authorized licensed use limited to: UNIVERSITE DE CARTHAGE. Downloaded on June 25,2024 at 11:13:41 UTC from IEEE Xplore. Restrictions apply.
error curves converge to zero and the curves in Fig. 6 show the
adaptive tuning of non-zero PPD in PJM along the time axis,
which explained and adaptivity of PB-AILC. We can find that
the higher iterations are, the lower the error is, and the better
the formation works, which is indeed theoretically correct.
However, it does not mean that the higher iteration number
is better because in practical applications the computational
complexity needs to be considered and redundant iterations
obviously will cost more computational resources, which we
do not expect. Combining the convergence curve in Fig. 5
and simulation results in Fig. 4, we choose 800 iterations
to achieve relatively good performance while keeping the
iteration number as low as possible.

Fig. 5. Convergence curve of posture error using PB-AILC scheme.

Fig. 7. Formation result at the 800th iteration using ILC.

To show the advancement of PB-AILC, a ILC method in


(20) is chosen as a comparison. Other initial conditions of ILC
method are the same as PB-AILC.
 
0.5 1 0
uf (i, t) = uf (i − 1, t) + ef (i − 1, t) (20)
0 0 1
Fig. 7 shows the formation result of ILC. Comparing Fig. 7
with Fig. 4 (c), we found the improvement of PB-AILC with
respect to ILC. In the same 800 iterations, PB-AILC shows a
better formation performance than ILC. From Fig. 8, we can
see that PB-AILC makes the overshoot less and has a slight
superiority in settling time.
In conclusion, the above numerical simulations prove the
effectiveness and advancement of PB-AILC in dealing with
formation problems. The convergences of formation error and
the comparison results with ILC show its superiority.
IV. C ONCLUSION
For the leader-follower formation control of MMRSs, we
introduce a posture-based adaptive iterative learning scheme
called PB-AILC. We use mathematical analysis and numerical
simulations to prove its feasibility, effectiveness, and advance-
ment. The merits of PB-AILC scheme are that 1) None of the
Fig. 6. Adaptive parameter Φ tuning along the time axis. model information but only real-time measurement I/O data

6467

Authorized licensed use limited to: UNIVERSITE DE CARTHAGE. Downloaded on June 25,2024 at 11:13:41 UTC from IEEE Xplore. Restrictions apply.
[7] J. Lin, Z. Miao, H. Zhong, W. Peng, Y. Wang and R. Fierro, “Adaptive
Image-Based Leader–Follower Formation Control of Mobile Robots
With Visibility Constraints,” IEEE Transactions on Industrial Electron-
ics, vol. 68, no. 7, pp. 6010-6019, July 2021.
[8] Z. Miao, Y.-H. Liu, Y. Wang, G. Yi and R. Fierro, “Distributed estima-
tion and control for leader-following formations of nonholonomic mobile
robots”, IEEE Transactions on Automation Science and Engineering,
vol. 15, no. 4, pp. 1946-1954, Oct. 2018.
[9] B. Tutuko, S. Nurmaini and G. F. Fitriana, “Tracking control en-
hancement on non-holonomic leader-follower robot,” in International
Conference on Electrical Engineering and Computer Science (ICECOS),
2017, pp. 83-862.
[10] Y. Zhao, D. Park, J. Moon and J. Lee, “Leader-follower formation
control for multiple mobile robots by a designed sliding mode controller
based on kinematic control method,” in 56th Annual Conference of the
Society of Instrument and Control Engineers of Japan (SICE), 2017, pp.
186-189.
[11] L. Cao, G. Liu and D. Zhang, “A leader-follower formation strategy
for networked multi-agent systems based on the PI predictive control
method,” in 40th Chinese Control Conference (CCC), 2021, pp. 4763-
4768.
[12] R. Hou and X. Bu, “Leader-follower iterative learning formation control
for robots with arbitrary initial position,” Computer Engineering and
Applications, vol. 56, no. 20, pp. 226-231, 2020.
[13] Z. Hou, R. Chi and H. Gao, “An Overview of Dynamic-Linearization-
Based Data-Driven Control and Applications,” IEEE Transactions on
Industrial Electronics, vol. 64, no. 5, pp. 4076-4090, 2017.
[14] H. Tan, Y. Wang, M. Wu, Z. Huang and Z. Miao, “Distributed Group
Coordination of Multiagent Systems in Cloud Computing Systems Using
a Model-Free Adaptive Predictive Control Strategy,” IEEE Transactions
on Neural Networks and Learning Systems, vol. 33, no. 8, pp. 3461-
3473, Aug. 2022.
[15] R. Chi and Z. Hou, “A new adaptive iterative learning control motivated
by discrete-time adaptive control,” International Journal of Innovative
Computing, Information and Control, vol. 4, no. 6, pp. 1267-1274, 2008.
[16] X. Bu, Q. Yu, Z. Hou and W. Qian, “Model Free Adaptive Iterative
Learning Consensus Tracking Control for a Class of Nonlinear Multia-
gent Systems,” IEEE Transactions on Systems, Man, and Cybernetics:
Fig. 8. Performance comparison of posture error between ILC and PB-AILC. Systems, vol. 49, no. 4, pp. 677-686, April 2019.
[17] Q. Ai, D. Ke, J. Zuo, W. Meng, Q. Liu, Z. Zhang and S. Xie, “High-
Order Model-Free Adaptive Iterative Learning Control of Pneumatic
Artificial Muscle With Enhanced Convergence,” IEEE Transactions on
of the MMRS is used in the control process. 2) It allows the Industrial Electronics, vol. 67, no. 11, pp. 9548-9559, Nov. 2020.
learning gain in the iteration process to be tuned online based [18] X. Bu, Z. Hou, R. Chi, “Model Free Adaptive Iterative Learning Control
for Farm Vehicle Path Tracking,” IFAC Proceedings Volumes, vol. 46,
on the estimated value of PJM, which improves the ability no. 20, pp. 153-158, 2013.
to describe the nonlinearity and uncertainty of the MMRS. [19] Z. Hou and S. Jin, “Data-Driven Model-Free Adaptive Control for a
However, we have not taken into account the impacts of Class of MIMO Nonlinear Discrete-Time Systems,” IEEE Transactions
on Neural Networks, vol. 22, no. 12, pp. 2173-2188, Dec. 2011.
convergence speed and time spent. Meanwhile, we have not
deployed PB-ILC scheme in a real experiment. Those will be
parts of improvement in our further study.

R EFERENCES
[1] X. Zhao, B. Tao and H. Ding, “Multimobile Robot Cluster System for
Robot Machining of Large-Scale Workpieces,” IEEE/ASME Transac-
tions on Mechatronics, vol. 27, no. 1, pp. 561-571, Feb. 2022.
[2] Z. Wu, Z. Cao, Y. Yu, L. Pang, C. Zhou and E. Chen, “A multi-robot
cooperative hunting approach based on dynamic prediction of target
motion,” in IEEE International Conference on Robotics and Biomimetics
(ROBIO), 2017, pp. 587-592.
[3] Z. Yan, N. Jouandeau and A. Ali-Chérif, “Multi-robot heuristic goods
transportation,” in 6th IEEE International Conference Intelligent Sys-
tems, 2012, pp. 409-414.
[4] S. Ge, “A survey and a simple simulation of leader- follower formation
control with MatLab,” in 3rd International Symposium on Robotics and
Intelligent Manufacturing Technology (ISRIMT), 2021, pp. 127-130.
[5] T. Balch and R. C. Arkin, “Behavior-based formation control for
multirobot teams,” IEEE Transactions on Robotics and Automation, vol.
14, no. 6, pp. 926-939, Dec. 1998.
[6] L. Chen and M. Baoli, ”A nonlinear formation control of wheeled
mobile robots with virtual structure approach,” in 34th Chinese Control
Conference (CCC), 2015, pp. 1080-1085.

6468

Authorized licensed use limited to: UNIVERSITE DE CARTHAGE. Downloaded on June 25,2024 at 11:13:41 UTC from IEEE Xplore. Restrictions apply.

You might also like