Energy-optimal problem of multiple nonholonomic wheeled mobile robots via distributed event-triggered optimization algorithm
Zhang Ying-Wen1, Wang Jin-Huan1, †, Xu Yong1, Yang De-Dong2
School of Science, Hebei Province Key Laboratory of Big Data Calculation, Hebei University of Technology, Tianjin 300401, China
School of Artificial Intelligence, Hebei University of Technology, Tianjin 300401, China

 

† Corresponding author. E-mail: wjhuan228@163.com

Project supported by the National Natural Science Foundation of China (Grant No. 11701138) and the Natural Science Foundation of Hebei Province, China (Grant Nos. F2017202009 and F2018202075).

Abstract

The distributed event-triggered optimization problem for multiple nonholonomic robots has been studied to minimize the global battery energy consumption. Each robot possesses its own cost function which depends on the state of the hand position and represents battery energy consumption. By coordinate transformation, the dynamics of the hand positions can be formulated into two groups of first-order integrators. Then the distributed event-triggered optimization algorithm is designed such that the states of robots’ hand positions exponentially converge to the optimizer of the global cost function. Meanwhile, the velocity and orientation of each robot are ensured to reach zero and a certain constant, respectively. Moreover, the inter-execution time is lower bounded and the Zeno behavior is therefore naturally avoided. Numerical simulations show the effectiveness of the proposed algorithm.

1. Introduction

In robot systems, mobile robots are expected to operate in environments where it may not be easy to replace batteries or refill the fuel, such as hostile environments.[1] Therefore, to ensure that the robot team can complete their missions, energy consumption problems must be taken into account at the design phase. Energy consumption caused by both motion and communication belong to the field of energy-optimal problems.[27] In the existing research, the energy-optimal problems are mostly studied by optimal control theory. For example, in Ref. [4], the authors proposed a power control and channel allocation optimization algorithm with low energy consumption for channel allocation game model. Tokekar et al.[6] studied the problem of finding velocity profiles for car-like robots to minimize the energy consumption while traveling along a given path. In Ref. [7], Egerstedt et al. introduced an open-loop control protocol for multi-robots to achieve rendezvous while minimizing the overall energy consumption. However, this type of open-loop controller cannot achieve adjustment and compensation for deviations of the controlled variables. Therefore, the closed-loop controller is proven to be more favorable to actual robot systems.

Over the past few decades, distributed optimization problems have increasingly attracted attention in science and engineering. In distributed optimization problems, each agent has a local cost function, which is only known by itself. The goal of distributed optimization is to cooperatively optimize a global cost function in the form of the sum of local cost functions of all agents. In the process of optimization, each agent can only use the local information of its neighbors. There are a large number of studies for distributed optimization algorithms in the context of multi-agent systems. Most of these studies focus on the protocols with first-order or second-order linear systems. The adopted methods include sub-gradient method,[810] gradient method,[1113] primal-dual method[14,15], zero-gradient-sum (ZGS) method,[1618] and so forth. For example, Nedic et al.[8] first put forward the distributed discrete-time sub-gradient optimization algorithm with diminishing step size. In Ref. [12], a distributed gradient optimization algorithm was presented for continuous-time multi-agent systems with single-integrator dynamics, which took into account the gradient gain and finite-time convergence. In Ref. [13], the authors proposed a gradient-based distributed optimization algorithm for first-order integrators. A continuous-time zero-gradient-sum algorithm was presented in Ref. [16] to solve the time-varying distributed convex optimization problem. However, the ZGS algorithm needs to restrict the initial states of agents.

In practice, each agent is usually equipped with simple embedded micro-processor which has limited energy and communication bandwidth. Most existing algorithms are updated continuously, which might be inefficient or impractical since they can lead to large communication load and waste excessive resources. To reduce the unnecessary waste of communication and computation resources, event-triggered control laws have been applied to single and double-integrators.[1922] Similarly, event-triggered algorithms have also been applied to distributed optimization problems.[2329]et al.[23] focused on the distributed event-triggered sub-gradient algorithm for solving a class of convex optimization problem based on first-order discrete-time multi-agent systems. In Ref. [25], an event triggered zero-gradient-sum algorithm was proposed to solve the optimization problem. To avoid calculating the threshold of the triggering times continuously, Liu et al.[26] presented a novel distributed event-triggered optimization algorithm based on sampled-data state. In Ref. [29], the distributed event-triggered optimization algorithm was extended from first-order dynamics to second-order dynamics, thereby accomplishing the optimal coordination problems.

Motivated by these considerations, in this paper we focus on the distributed event-triggered optimization problem for multiple nonholonomic wheeled mobile robots. The local cost function of each robot represents its own battery energy consumption, which is related to the state of the hand position. By coordinate transformation, the dynamics of the hand positions of robots are formulated as two groups of first-order integrators. We will solve the energy-optimal problem by using the distributed event-triggered optimization algorithm, and at the same time the states of the hand positions exponentially converge to the optimal point.

The main contributions of this paper follow. (i) Most existing results related to distributed optimization problems have concentrated on first-order or second-order linear systems. However, there are abundant nonlinear and even nonholonomic systems in practice, such as mobile robots. Therefore, we propose the distributed event-triggered optimization algorithm to solve the energy-optimal problem for multiple nonholonomic mobile robots. (ii) A distributed event-triggered control scheme is designed to minimize the energy consumption. For each robot, the controller is updated only at its own triggering times. Moreover, the event-triggered algorithm can avoid the Zeno behavior. (iii) It can be proven that the optimal solution is the system equilibrium point and also the consensus point. Furthermore, the exponential convergence rates of position and velocity are presented with the help of Lyapunov stability theory, which is different from the asymptotical convergence in Ref. [13].

The rest of this paper is organized as follows. In Section 2, some basic concepts on algebraic graph theory are introduced and the problem is formulated. Section 3 shows the main results. The distributed event-triggered optimization algorithm for the energy-optimal problem of multiple robots and the detailed stability analysis are also presented. Simulations are given to validate the theoretical results in Section 4. Finally, Section 5 gives a conclusion of this paper.

2. Preliminaries and problem formulation
2.1. Algebraic graph theory

In this paper, we use a graph to model a communication network. represents an undirected graph, where and are the node set and the edge set, respectively. If exists, then the pair of nodes i and j are neighbors and they can transmit their information to each other. The neighbor set of node i is defined by . A=[aij]∈∝N × N is a weighted adjacency matrix of graph , where aij = aji > 0 if , otherwise aij = 0. The degree matrix of is given by D=diag{b1,…,bN}∈ℝN × N with . The Laplacian matrix L is defined as L = DA. For undirected graphs, L is symmetric and positive semi-definite. Graph is called connected if there is one path between any two nodes in . Then, the associated Laplacian matrix has a simple zero eigenvalue and 1N is the corresponding eigenvector; i.e., L1N = 0. Assume that the eigenvalues of L are λ1,…,λN with an increasing order, 0=λ1 < λ2λ3 ≤ · ≤ λN. More details can be found in Ref. [30].

2.2. Problem formulation

In this subsection, we put forward an energy-optimal strategy for robots to move some object to one location and ensure the battery energy consumption to reach the minimum.

First, consider a group of N nonholonomic wheeled mobile robots moving on a horizontal plane. The kinematics of robot i is described by where qi(t) = [qi1(t),qi2(t)]T ∈ ℝ2 and qi3(t) ∈ ℝ are the position and orientation of robot i, respectively. vi(t), ωi(t) ∈ ℝ are the linear and angular velocity control inputs of robot i, respectively.

Assume that the local cost function relies on the state of the hand position and shows the battery energy consumption. Combined with the practical significance of distributed optimization, the energy-optimal problem can be formulated as the following distributed optimization problem: where fi:ℝ2 → ℝ is the local cost function of robot i.

Next, we show the dynamics of the hand positions. According to the conversion of Ref. [32], assume the hand position of each robot is a point located at a fixed distance ρ > 0 from the centroid of the robot and on the axis of the orientation (see Fig. 1).

Fig. 1. Illustration of the hand position.

For robot i, the coordinates of the hand position are where ρ > 0 is a positive constant.

Differentiating Eq. (3) yields where ui1(t) and ui2(t) are the virtual control inputs. As ρ > 0, the above transformation is hence globally invertible, and the true control inputs vi(t) and ωi(t) can be deduced as

Hence, the dynamics of the hand positions can be expressed as

Obviously, the dynamics of the hand position of each robot has been modelled in the form of single integrator where ui(t) = [ui1(t),ui2(t)]T.

The distributed energy-optimal problem for multiple nonholonomic wheeled mobile robots is stated as: design an event-triggered control law, as well as an event-triggering condition for each robot such that the hand positions of all robots tend to the optimal point of the cost function while the global energy consumption reaches the minimum. In addition, the velocities and orientations converge to zero and some fixed values, respectively; that is, where , qis is a certain constant.

Assume that the local cost function fi is only available to robot i, and may be different. We make the following assumption on these cost functions.

3. Distributed event-triggered optimization algorithm for minimum energy consumption

In this section, a distributed event-triggered optimization algorithm is proposed to achieve the control objective in Eq. (8) and thus solve the optimization problem in Eq. (2).

To avoid the unnecessary resource waste, we construct the control protocol based on the distributed event-triggered strategy. For robot i, the control law is updated only at some discrete times , called the event-triggering sequence, satisfying a certain triggering condition.

To construct the triggering condition, we define the state measurement error for robot i as where is the k-th event-triggering time of robot i.

The event-triggering times for robot i are specified when the following triggering condition is violated; that is, where βi > 0 will be determined later and . For , is the latest triggering instant of robot j.

Then, the distributed event-triggered optimization control law for robot i is constructed as where α and γ are positive constants. The control law for each robot depends on itself and its neighbours’ latest event-triggering instants.

Let x = col(x1,…,xN), e = col(e1,…,eN), and F(x) = col(f1(x1),…,fN(xN)). The dynamics of the hand positions can be rewritten in the compact form

The following lemma illustrates a basic necessary condition for the optimality.

The following lemma indicates the relationship between the equilibrium point of system in Eq. (12) and the optimal solution of problem in Eq. (2).

Conversely, for any equilibrium point of system in Eq. (12), we have e = 0 once the equilibrium of system in Eq. (12) is achieved. Hence, we obtain From Assumption 1, we have . Multiplying the left of Eq. (13) by gives Therefore, which means According to Lemma 2, we obtain that the equilibrium point of system in Eq. (12) is the optimal solution of problem in Eq. (2).

Employ the transformation and system equation (12) is transformed into where h = h1 +h2 with h1 = F(x+e)−F(x) and h2 = F(x)−F(x∗). Let where X1,E1 ∈ ℝ2,X2,E2 ∈ ℝ2(N−1), and Q is determined by Remark 1. From Eq. (16) and Remark 1, we obtain

The derivative of V along the system Eq. (18) yields Recalling that h = h1 + h2 = (F(x + e) − F(x)) + (F(x) − F(x∗)), using Lemma 1 and Assumption 3 along with ||QI2|| = 1, we obtain the following estimation: In addition, Note that RTLR is symmetric and positive definite matrix with eigenvalues λ2,…,λN and thus the eigenvalues of are . Obviously, . Then we have Therefore, we obtain It follows from the triggering condition in Eq. (10) that Hence, where . Recall that ||QI2|| = 1, one has By substituting Eq. (22) into Eq. (21), we obtain Since , then Therefore, one has Denote , and we can select appropriate parameters α,γ, and βi such that μ > 0. One knows from Eq. (20) that where c = V(0). This implies that exponentially converges to 0, that is, xi exponentially converges to x∗. From Eq. (11) and F(x∗) = 0, u(t) = col(u1(t),…,uN(t)) can be written as According to Assumption 3, we obtain exponentially converges to zero and meanwhile e(t) exponentially tends to zero; therefore, u(t) exponentially reaches zero. From Eq. (5), we know that vi(t) and ωi(t) exponentially converge to zero. Hence, the orientation qi3(t) tends to a certain stationary value.

Next, we will show that any inter-event intervals are lower bounded by a positive number, which guarantees that there is no Zeno behavior.

By applying Newton–Leibnitz formula and noting that , we deduce that According to the above inequality, we have Note that on the basis of the triggering condition in Eq. (10), we hence obtain Multiplying Eq. (25) by and denote , one knows where , , and .

Since exp(q)≥ 1+q for any q ∈ ℝ, we have . For any robot i, In the conclusion, the Zeno behavior can be excluded.

4. Numerical simulations

In this section, some numerical results are given to verify the theoretical results.

Consider the unicycle system with five wheeled mobile robots satisfying Eq. (1), whose interaction topology is depicted in Fig. 2 with weights of all edges as 1. Thus, the largest eigenvalue of the Laplacian matrix is λ5 = 5. The initial states are chosen as (q11,q12,q13)(0) = (1, 0, 0), (q21,q22,q23)(0)=(1.5, 0.5, 0), (q31,q32,q33)(0)=(1, 1,π), , and (q51,q52,q53)(0)=(0.1, 1, 0). Define the hand position by ρ = 0.5. The local battery consumption functions are defined as fi(xi) = 50||xi(t)−xi(0)||2. It can be easily calculated that the optimal value x∗ = [1, 0.71]T. It is observed that the above local cost functions are strongly convex and their gradients are Lipschitz. After the calculation of the above local cost function , the minimum strongly-convex constant is χ = 100 and the maximum Lipschitz constant is σ = 200. Choose α = 0.6, γ = 0.01, β1 = 0.04, β2 = 0.04, β3 = 0.03, β4 = 0.02, and β5 = 0.045 satisfying Eq. (19). Then, by Theorem 1, the optimization problem in Eq. (2) can be solved by the distributed optimization control law in Eq. (11) with the event-triggering condition in Eq. (10). The simulation results are shown in Figs. 36, indicating the effectiveness of the proposed energy-optimal strategy. Figure 3 shows the states of five robots’ hand positions converge to the optimal point x∗ = [1, 0.71]T. Figure 4 shows the event-triggering times of five robots. The measurement error and threshold evolution of the robots are shown in Fig. 5. From Fig. 6, we can see that the orientations converge to fixed values, and the linear velocities reach zero; that is, they will keep still when the robots reach the optimal point.

Fig. 2. Communication topology for robots.
Fig. 3. State trajectories of xi.
Fig. 4. Event-triggering times of five robots.
Fig. 5. Measurement error and threshold evolution of robots.
Fig. 6. The evolution of orientation states and linear velocities of robot i.
5. Conclusion

In this paper, we have presented an energy-optimal strategy such that the states of the robots’ hand positions achieve consensus while minimizing the overall energy consumption of robots. We have formulated this problem as the distributed optimization problem. By coordinate transformation, the dynamics of the hand positions are formulated as two groups of first-order integrators. A distributed event-triggered optimization algorithm is hence presented and guaranteed that the equilibrium point is the optimal solution. Based on the Lyapunov theory, sufficient conditions are obtained to make the states of the robots’ hand positions exponentially converge to the optimal point of the global cost function. At the same time, the velocities and orientations converge to zero and certain fixed values, respectively. Moreover, the Zeno behavior is avoided.

Our further research considers that the robots rendezvous at some point while the overall battery energy consumption reaches the minimum in a finite time. The constraint condition that each robot moves only in its own area will also be considered.

Reference
[1] Angie S Farokh B B Yen I L 2011 10th International Symposium on Autonomous Decentralized Systems March 23–27, 2011 Tokyo, Hiroshima, Japan 147 10.1109/ISADS.2011.104
[2] Jaleel H Rahmani A Egerstedt M 2013 IEEE Trans. Autom. Control 58 534
[3] Wang G Q Luo H Hu X X 2018 Chin. Phys. 27 028901
[4] Hao X C Liu J S Xie L X Chen B Yan N 2018 Chin. Phys. 27 080102
[5] Alfieri A Bianco A Brandimarte P Chiasserini C F 2007 Eur. J. Oper. Res. 181 390
[6] Tokekar P Karnad N Isler V 2011 IEEE International Conference on Robotics and Automation May 9–13, 2011 Shanghai, China 1457 10.1109/ICRA.2011.5980374
[7] Setter T Egerstedt M 2017 IEEE Trans. Control Syst. Technol. 25 1257
[8] Nedic A Ozdaglar A 2009 IEEE Trans. Autom. Control 54 48
[9] Qiu Z R Liu S Xie L H 2016 Automatica 68 209
[10] Xi C G Khan U A 2017 IEEE Trans. Autom. Control 62 3986
[11] Yi P Hong Y G Liu F 2015 Syst. Control Lett. 83 45
[12] Lin P Ren W Farrel J A 2017 IEEE Trans. Autom. Control 62 2239
[13] Wang A J Liao X F Dong T 2018 IET Control Theory Appl. 12 1515
[14] Hale M T Nedic A Egerstedt M 2017 IEEE Trans. Autom. Control 62 4421
[15] Liu P Li H Q Dai X G 2018 Int. J. Syst. Sci. 49 1256
[16] Rahili S Ren W 2017 IEEE Trans. Autom. Control 62 1590
[17] Liu J Y Chen W S Dai H 2017 Int. J. Syst. Sci. 48 1836
[18] Guo Z J Chen G 2018 Int. J. Robust Nonlinear Control 28 4900
[19] Wang J H Xu Y L Zhang J Yang D D 2018 Chin. Phys. 27 040504
[20] Cao J Wu Z H Peng L 2016 Chin. Phys. 25 058902
[21] Miao G Y Cao J D Alsaedi A 2017 J. Frankl. Inst. 354 6956
[22] Qi B Cui B T Lou X Y 2014 Chin. Phys. 23 110501
[23] Q G Li H Q Xia D W 2017 Neurocomputing 235 255
[24] Kia S S Cortes J Martinez S 2015 Automatica 55 254
[25] Chen W S Ren W 2016 Automatica 65 90
[26] Liu J Y Chen W S Dai H 2016 Int. J. Control Autom. Syst. 14 1421
[27] Liu S Xie L H Quevedo D E 2018 IEEE Trans. Control Network Syst. 5 167
[28] Richert D Cortes J 2016 SIAM J. Control Optim. 54 1769
[29] Wang D Gupta V Wang W 2018 Neurocomputing 319 34
[30] Godsil C Royle G 2001 Algebraic Graph Theory New York Springer-Verlag 10.1007/978-1-4613-0163-9
[31] Lawton J Beard R Young B 2003 IEEE Trans. Robot. Autom. 19 933
[32] Chen X Hao F Ma B L 2017 IET Control Theory Appl. 11 890
[33] Horn R A Johnson C R 2003 Matrix Analysis Cambridge Cambridge University Press 10.1007/978-1-4612-0653-8
[34] Yan J X Yu H Xia X H 2018 Neurcomputing 296 100
[35] Bertsekas D P Nedic A Ozdaglar A E 2003 Convex Analysis and Optimization Belmont Athena Scientific