Synchronization of multi-agent systems (MASs) is an important component of many cooperative control problems, such as rendezvous , formation control , flocking control , containment control  and sensor networks . Synchronization problems can be roughly categorized into leaderless and leader-following. In the leaderless synchronization, which is closely related to the consensus problem, the agents aim to reach to a static or dynamic agreement on a common value [6, 7, 8]. On the other hand, in the leader-following synchronization, agents aim to make the agreement on the states generated by a leader. In this paper, we focus on the design of a distributed leader-following algorithm when the only information available about the leader is its sampled state, which is only available to a subset of followers.
Literature review: The leader-following algorithms for single integrator and double integrator dynamics are presented in , and for homogeneous LTI systems are proposed in  and . For systems constituted of heterogeneous LTI followers,  and  propose the algorithms to synchronize with a passive zero-input LTI leader.  and  develop the controls for the single and double integral system, respectively, to track an active leader (active leader is a leader that has a control input). But their works assume the leader’s control input is available to all the followers.  and  propose a leader-following algorithm respectively for homogeneous LTI and heterogeneous nonlinear MASs in which the unknown input of the leader is bounded and is not available to any follower. But the control inputs in  and  have the sliding mode structure and suffer from the well-known undesirable chattering behavior. We recall that from a practical perspective, chattering is undesirable and leads to excessive control energy expenditure .  is the recent result for the leader-following problem, which is based on the result of 
and develops a distributed observer to estimate the leader’s state for each follower. Then, the output synchronization of heterogeneous leader-follower linear systems is achieved by optimal local tracking of the output of the observer. We note that in both and , the active leader is restricted to be linear and have limited input. The work reviewed so far are all converge to leader following in an asymptotic manner, i.e., the settling time to reach an agreement is infinity. For fast convergence, , and  propose the finite-time synchronization algorithms for single and double integral MASs, where the upper bound of the settling time explicitly depends on the initial state of the MAS. Therefore, to use these algorithms, the centralized knowledge of the initial state of the MAS is essential to estimate the settling time.  and  propose the fixed-time synchronization algorithms, where the settling time is bounded and independent of the initial state of the MAS. However, for both these finite and fixed-time algorithms, the settling time is upper bounded by a conservative estimation.  introduces the specified-time synchronization control for the leaderless MASs in which one can determine the settling time exactly in advance. Specified-time synchronization can be useful to the applications that require precise acting time, such as target attack at a specified time.
Statement of contributions: In this paper, we consider a leader-following problem in which the only information available about the leader is its instantaneous sampled state that is known only to a subset of a group of heterogeneous LTI followers at the sampling times. We make no assumptions about the input of the leader or the structural form of its dynamics. That is, the state of the leader is perceived by the followers as an exogenous signal. The sampled states of the leader can be the states of a physical system (e.g., in a pursuit-evasion problem) or a set of desired reference states of a virtual leader (e.g., in a waypoint tracking problem). Given the limited information about the leader, we seek a practical solution that enables the followers to arrive at the sampled state of the leader before the next sampling time. That is, we design a distributed algorithm that steers a group of heterogeneous LTI followers to be at the sampled states of the leader at finite time just before the next sampled state is obtained. We note that practical one step lagged tracking has also been used in [26, 27, 28] for a set of dynamic average consensus algorithms with asymptotic tracking behavior. Our solution is inspired by the minimum energy controller design  in the classical optimal control theory, and is proposed for problems where the interaction topology of the followers plus the leader is an acyclic digraph with the leader as the global sink. Directed acyclic interaction topology can be interpreted as the agents only obtaining information from those in front of them (see, [30, 31] for algorithms designed over acyclic graphs). Our algorithm also allows the followers to track the sampled state of the leader with a locally chosen offset, which can be time-varying. This offset, when the followers are mobile agents and their whole state or part of it is the position vector, can be used to enable the followers to form a transnational invariant formation  about the sampled state of the leader. For a special class of non-homogeneous LTI MAS, we show that our results can be extended to solve a leader-following problem where we want only an output of the followers to follow the leader’s sampled state. Finally, we show that if the followers are homogeneous, our algorithm not only results in a leader following behavior, but also it makes the states and inputs of the followers fully synchronized after the first sampling epoch. We demonstrate our leader-following results via three numerical examples. In the first example, we show the application of our leader-following algorithm in following a nonlinear mass-spring-damper leader under a specific formation structure for a group of heterogeneous linear mass-spring-damper systems. In the second example, we demonstrate the use of our algorithm for reference state tracking via a group of second order integrator followers with bounded control. The result shows the synchronization of the homogeneous followers is realized. Moreover, using the intrinsic properties of our leader following algorithm, we show that the arrival times at the reference states can be specified in such a way that the inputs of the followers stay within the saturation bounds. Our last example demonstrates an output-tracking scenario for a group of aircraft.
Organization: The rest of this parer is organized as follows. Section II gathers basic notation and graph-theoretic notions. Section III gives our problem definition and objective statement. Section IV proposes our distributed leader-following algorithm. In Section V, three applications are demonstrated. Section VI concludes the results of this paper. Appendix A contains the proof of our main result, Theorem IV.1. Finally, Appendix B presents an auxiliary result, which is invoke to support the feasibility of the sampling time design in our second numerical example.
Notation: We let , , , , and denote the set of real, positive real, non-negative real, integer, and non-negative integer numbers, respectively. The transpose of a matrix is .
Graph theoretic notations and definitions: Here we review our graph related notations and relevant definitions and concepts from graph theory following . A digraph, is a triplet , where is the node set and is the edge set, and is the adjacency matrix of the graph defined according to if and , otherwise. An edge from to means that agent can send information to agent . Here, is called an in-neighbor of and is called an out-neighbor of . A directed path is a sequence of nodes connected by edges. A directed path that starts and ends at the same node and all other nodes on the path are distinct is called a cycle. A digraph without cycles is called directed acyclic graph. The out-degree of a node is . The out-degree matrix of a graph is . We denote the set of in-neighbors of an agent by and the out-neighbors of agent by . A node is called a global sink of if it outdegree and for every node there is at least a path from to .
Iii Problem definition
We consider a group of heterogeneous MAS whose dynamics is described by
where is the state vector and is the control vector. Throughout the paper we assume that the agents’ dynamics (1) is controllable, i.e., for is controllable. These agents (referred hereafter as followers) aim to follow a dynamic signal with possibly a locally chosen offset. This signal can be a dynamic reference signal of a virtual leader or the state of an active physical leader with (possibly) a nonlinear dynamics, e.g., . The dynamical model and the input of the leader is not known to the followers. The interaction topology between the followers is described by a acyclic digraph, denoted by . Only a subset of followers in , denoted by , has access to at the sampling times , . Throughout the paper we assume that for any with . We let be the digraph consisted of the leader and and the directed edges connecting to the leader. In what follows, we assume that the leader is the global sink of , so that its information reaches all the agents in an explicit or implicit manner (see Fig. 1 for an example). We let be the set of the out-neighbors of agent in graph ; we note the . Finally, we call the followers homogeneous if , for .
Give that we only have a limited information about the leader (only the sampled states of the leader is available), we seek a practical solution that enables the followers to arrive at the sampled state of the leader before the next sampling time. Therefore, our objective in this paper is to design a distributed control rule for the input vector of each follower such that
That is, the follower can steer itself to be in offset with respect to the state of the leader in time before the next sampling time . We note that the set of offsets , when it is related to the position offsets of the agents, defines the formation of the followers around the leader. Here, the term formation refers to transnational invariant formation [32, Section 6.1.1]. For scenarios where the objective is to synchronize to the state of the leader, is set to zero for all . To form the offset, we assume that at each sampling time , follower knows for ; either the follower is given with respect to its out-neighbor or constructs it locally after agent sends its to agent . We note that if the leader is a global sink of , given and a set of , and , we can show that the state offset for follower with respect to the leader is unique.
Iv Main result
In this section, we develop a novel distributed solution to solve the leader-following problem stated in Section III. To present this result, we recall that
is the controllability Gramian of for any finite time . Since is controllable, is full rank and invertible at each time . We start by using a classical optimal control result to make the following statement.
Consider a leader-following with an offset problem where each follower’s dynamics is given by (1) with controllable. Suppose is a follower in that has access to of the leader at each sampling time , , i.e., . Also, is the desired state offset with respect to . Starting at an initial condition with , for any let
for , where where , and
Then, for every we have for all . Moreover, at each time , the control input of satisfies
The proof follows from the classical finite time minimum energy optimal control design [29, page 138]. ∎
Lemma IV.1 essentially states that any follower that samples the leader, in the inter-sampling time interval can use the classical minimum energy control to steer towards the latest sampled state of the leader (with an offset if specified). Next, we show that this idea can be extended to a distributed setting in which only a subset of the followers have access to the leader’s sampled state. To present our results we first introduce some notations. We denote the adjacency matrix and out-degree matrix of the followers’ interaction topology , respectively, by and . We let
be the indicator operator that defines the state of connectivity of follower to the leader. For , we also define
We notice that , where is the controllability Gramian (3). Therefore at each finite time , by virtue of controllability of , is invertible. Moreover, note that using the classical control results we can show that can be computed numerically from where and for are obtained from
With the proper notations at hand, we present our distributed solution to solve our leader-following problem of interest as follows.
Theorem IV.1 (A leader-following algorithm for a group of heterogeneous LTI followers).
Consider a leader-following problem where the followers’ dynamics are given by (1). Suppose the leader’s time-varying state is . Let the network topology be an acyclic digraph with leader, node , as the global sink. Suppose every follower has access to at each sampling time , . Let and be the desired state offset (formation) with reference to and , respectively. Starting at an initial condition with , let for
where is given in (8a), , and . Then, the followings hold for and :
, moreover, and ;
the trajectory of every follower is
the control input of every agent is equal to (4).
Remark IV.1 (Robustness to state perturbations).
We observe that the leader-following algorithm of Theorem IV.1 has robustness to state perturbations similar to the well-known Model Predictive Control (MPC). Even though the controller implemented in each epoch is an open-loop control, since every follower exerts its state at time as initial condition to the controller, the algorithm can account for the slight perturbations in the follower final state at the end of each epoch.
Remark IV.2 (Implementation of control law (iv.1)).
To implement (IV.1), we note that the component of (IV.1) that multiplies is computed using the local variables of follower and the sampled state of the leader if is non-zero, i.e., . The component of (IV.1) that multiplies is computed using the local variables of follower and variables of its out-neighbors if . To compute this term, if the follower knows of its out-neighbor (which is the case e.g., when the group is homogeneous), it can implement control (IV.1) by obtaining the state of its out-neighbor and computing and locally. Otherwise, each follower needs to obtain and of its each out-neighbor . We note here that since the interval is open from the left and the dynamics of all the followers is controllable, is well defined. However, for from the right, goes to infinity. But, since goes to zero as from the right, the product goes to zero as from the right. The “high-gain-challenge” observed here is often a common feature in any approach that is geared towards regulation in prescribed finite time. For example, finite-horizon optimal controls with a terminal constraint inevitably yield gains that go to infinity (see e.g., [34, 35, 36]). In practice, the multiplication of very large and very small values can create numerical problems. To address the problem, one way proposed in the literature is equivalent to employ a deadzone on at the beginning of each time interval. Another approach is equivalent to using a lager interval , where is small positive number, to compute such that for is no longer goes unbounded when goes to from the right. These approaches of course result in somewhat sacrifices in the accuracy at each arrival value at at the end of time interval . However, as discussed in Remark 4.1, the errors will not accumulate. As interestingly discussed in , the high-gain-challenge in the finite-time control can be contrasted with non-smooth feedback in the sliding mode control where the gain approaches infinity near sliding surface  (but the total control input is zero). However, unlike the sliding mode control, where the practical implementation of the high-gain leads to persistent chattering on the sliding surface , in our case the concern arises only at the start of each transition from one sampling time to the other. The aforementioned practical measures to handle the high-gain-challenge in our setting indeed can be compared to the boundary layer approach  in the sliding mode control to eliminate chattering.
Remark IV.3 (Time-varying MAS dynamics and network topology).
From the proof of Theorem IV.1, we can see that the followers dynamics can be allowed to be time-varying but piece-wise constant over each time interval , i.e., and , for . Similarly the network topology can be allowed to be time-varying as long as between the topology is fixed and satisfies the connectivity condition of Theorem IV.1.
Remark IV.4 (Minimum energy control in ).
From statement (c) of Theorem IV.1 it follows that at each time interval , , the control input of each follower is the minimum energy controller that transfers the follower from its current state to their desired state .
Remark IV.5 (Tracking a priori known desired states at exact sampling time and design of arrival times).
We note that if the leader is virtual and the sampled states are some desired states that are known a priori to with desired arrival time in , the agents can arrive at the desired state of the leader at the desired arrival time. Furthermore, for the homogeneous followers, in cases that the arrival times is not specified one of the followers in (we refer to it as super node that knows the initial state of all the other followers) can design the arrival times to meet other optimality conditions or to avoid violating constraints such as input saturation. In case of input saturation, the fact that by virtue of statement (c) of Theorem IV.1 the form of input vector of the followers are known to be (4) can be instrumental to the super node in design of arrival times. Our second demonstrative example in Section V offers the details.
Remark IV.6 (Extension of results to output tracking for a special class of MAS).
The design methodology of the state offset (formation) algorithm of Theorem IV.1 can be used in output tracking for a special class of MAS. Let the network topology be as described in Theorem IV.1 and the system dynamics of the followers be (1) where and (the state and input dimensions of the followers are not necessarily the same). Let the objective be that the output , , of each follower should satisfy
If is full row rank, we can use the control , , to write the output dynamics of each follower as . Then the method of Theorem IV.1 can be used to design , which can then be used to obtain the appropriate that will make the followers meet (11).
Finally, we note that if the followers are homogeneous, the followers can achieve full synchronization in the sense stated in the result below.
Corollary IV.1 (Full synchronization for homogeneous followers).
Let the state offset be constant i.e., for all or (equivalently for ), and assume that the followers are homogeneous. Then, it follows from statements (b) and (c) of Theorem IV.1 that the followers’ trajectories and inputs satisfy for and for , for every . One can easily verify this point by shifting the state coordinate with . Moreover, if the agents are initially in the specified offset i.e., for all , then these qualities also hold for .
Assume that there exists , for and a controllable pair known to all followers, such that using , , makes the followers dynamic homogeneous, i.e., , and , . Then, it is also possible to achieve full state synchronization by implementing (IV.1) to for heterogeneous followers. One sufficient condition for the existence of and , , is that of each follower is full row rank. Then, , and can be any controllable pair.
V Demonstrative examples
In this section, we demonstrate our results via numerical examples.
V-a A nonlinear-leader following problem for a group of heterogeneous followers
Consider a group of mass-spring-damper system (followers)
where is the state vector with and representing the displacement and velocity of the mass, , and are spring constant, damping constant and mass, respectively, and is the input force. The system’s parameters for are , , , , , , and , respectively. The leader denoted by is a nonlinear mass-spring-damper system
where the input is unknown to the followers and the system parameters . The interaction topology of the systems is shown in Fig. 1. Followers , and obtain the state of the leader with a sampling rate of per second, i.e., second, . The followers start at , , , , , , in a formation with uniform distance to the previous number of the follower. The objective is for the followers to track the sampled state of the leader while preserving the initial formation of the systems at every sampling time . The follower only knows the local formation, i.e., for . For example, follower 3 knows , , and .
The result of implementing the algorithm of Theorem IV.1 is shown in Fig. 2. The ‘’ represents the sampled leader’s states and ‘’ shows the followers track the sampled state in the desired formation at the next sampled time. In this example interestingly in the transition times similar to what is expected from homogeneous followers the state and input of all the followers are almost offset-synchronized. However, this property is not necessarily true in general for heterogeneous followers.
V-B Reference state tracking for a group of second integrator dynamics with bounded inputs
We consider a group of followers with second order integrator dynamics
for . The interaction topology of these followers is shown in Fig. 3, where, agent is the virtual leader that is defined more precisely below. Starting at initial conditions , , , , , , the leader-following mission for this team is to traverse through the sequence of desired states , which for privacy reason are only known to follower . The objective is to meet the sequence of desired states without violating any of the followers’ control bounds. In this problem setting, follower is the super node that knows the initial starting state of all the followers in the team and has computational power to compute the arrival times as follows to meet the team’s objective. First, we note that by virtue of statement (c) of Theorem IV.1 the form of input vector of the followers are known to be (4). Since follower 1 knows for , follower can evaluate of all the followers. Starting with , follower computes the arrival time at desired state from the process below
where with . Then, the arrival time so that the followers input do not saturate over is set to . (II) Due to Corollary IV.1, after first epoch, the followers inputs are equal to each other. Then, the remaining arrival time , are computed from the optimization problem
where with , for . The solution for this set of sequential optimal control problem is , , and seconds. At the end of process, follower broadcasts the times to the network. Broadcasting the reference states is not allowed due to privacy reasons. We note that the desired arrival times can be done offline by the system operator. To match the notation in (IV.1), at the implantation stage, we set , , and , , where . Figures 4 shows that all the followers meet the desired reference state of the virtual leader at the specified arrival times without delay (the ‘’ marks the reference states). Figure 4 also shows the control history of the agents. As seen, the control inputs respect the saturation bounds or . We can also observe that the followers’ states and inputs, as predicted in Corollary IV.1, are all synchronized after the first epoch. We should mention that by virtue of Lemma B.1 in Appendix B, every , , designed as described above is guaranteed to be a finite value.
V-C Output tracking for a group of aircraft
We consider a group of aircraft whose short-period dynamics is given by (taken from [38, Example 10.1])
where , and are are respectively, angle of attack, pitch rate and elevator angle of aircraft . The interaction topology of these aircraft is shown in Fig. 3, where, agent is the leader. For this system , therefore the condition of Remark IV.6 is satisfied and we can design a distributed algorithm to synchronize the pitch rate of the follower aircraft to the pitch rate of the leader aircraft when only sampled pitch rate of the leader at every 0.1 seconds is available to the follower aircraft . Figure 5 demonstrates the results.
In this paper, we have proposed a distributed leader-following algorithm for heterogeneous multi-agent systems with an active leader with unknown input. We have proved that our distributed leader-following algorithm for the linear followers steers the group to be at the sampled states of the leader at the specified arrival times. We showed that the control input of each follower agent between the sampling times is a minimum energy control. We also showed that after the first sampling epoch, the states of all the homogeneous follower agents are synchronized with each other. We demonstrated our results via leader-following problems of mass-spring-damper systems, mobile agents with second order integrator dynamics, and a group of aircraft. Future work will focus on extending our results to output following problem.
-  J. Lin, A. S. Morse, and B. D. O. Anderson, “The multi-agent rendezvous problem,” 42nd IEEE International Conference on Decision and Control, vol. 2, pp. 1508–1513, 2003.
-  W. Ren, “Consensus strategies for cooperative control of vehicle formations,” IET Control Theory Applications, vol. 1, no. 2, pp. 505–512, 2007.
-  H. Su, X. Wang, and Z. Lin, “Flocking of multi-agents with a virtual leader,” IEEE Transactions on Automatic Control, vol. 54, no. 2, pp. 293–307, 2009.
-  Y. Cao, W. Ren, and M. Egerstedt, “Distributed containment control with multiple stationary or dynamic leaders in fixed and switching directed networks,” Automatica, vol. 48, no. 8, pp. 1586 – 1597, 2012.
-  R. Olfati-Saber and J. S. Shamma, “Consensus filters for sensor networks and distributed sensor fusion,” Proceedings of the 44th IEEE Conference on Decision and Control, pp. 6698–6703, 2005.
-  R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of agents with switching topology and time-delays,” IEEE Transactions on Automatic Control, vol. 49, pp. 1520–1533, Sep. 2004.
-  D. V. Dimarogonas and K. J. Kyriakopoulos, “On the rendezvous problem for multiple nonholonomic agents,” IEEE Transactions on Automatic Control, vol. 52, no. 5, pp. 916–922, 2007.
-  S. S. Kia, B. V. Scoy, J. Cortés, R. A. Freeman, K. M. Lynch, and S. Martínez, “Tutorial on dynamic average consensus: The problem, its applications, and the algorithms,” IEEE Control Systems Magazine, vol. 39, no. 3, pp. 40–72, 2019.
-  W. Ren and R. W. B. Syrmos, Distributed Consensus in Multi-vehicle Cooperative Control. Springer London, 2008.
-  W. Ni and D. Cheng, “Leader-following consensus of multi-agent systems under fixed and switching topologies,” Systems & Control Letters, vol. 59, pp. 209–217, 2010.
-  H. Zhang, F. L. Lewis, and A. Das, “Optimal design for synchronization of cooperative systems: State feedback, observer and output feedback,” IEEE Transactions on Automatic Control, vol. 56, no. 8, pp. 1948–1952, 2011.
-  X. Wang and Y. Hong, “A distributed control approach to a robust output regulation problem for multi-agent linear systems,” IEEE Transactions on Automatic Control, vol. 55, no. 12, pp. 2891–2895, 2010.
-  Y. Su and J. Huang, “Cooperative output regulation of linear multi-agent systems,” IEEE Transactions on Automatic Control, vol. 57, no. 4, pp. 1062–1066, 2012.
-  Y. Hong, J. Hu, and L. Gao, “Tracking control for multi-agent consensus with an active leader and variable topology,” Automatica, vol. 42, pp. 1177–1182, 2006.
-  Y. Hong, G. Chen, and L. Bushnell, “Distributed observers design for leader-following control of multi-agent networks,” Automatica, vol. 44, pp. 846–850, 2008.
-  Z. Li and W. Ren, “Distributed tracking control for linear multiagent systems with a leader of bounded unknown input,” IEEE Transactions on Automatic Control, vol. 58, no. 2, pp. 518–544, 2013.
-  Y. Tang, Y. Hong, and X. Wang, “Distributed output regulation for a class of nonlinear multi-agent systems with unknown-input leaders,” Automatica, vol. 62, pp. 154–160, 2015.
-  J.-J. E. Slotine and W. Li, Applied Nonlinear Control. Prentice Hall, 1991.
Y. Yang and H. Modares, “Leader-follower output synchronization of linear
heterogeneous systems with active leader,”
IEEE Transactions on Neural Networks and Learning Systems, vol. 29, no. 6, pp. 2139–2153, 2018.
-  L. Wang and F. Xiao, “Finite-time consensus problems for networks of dynamic agents,” IEEE Transactions on Automatic Control, vol. 55, no. 4, pp. 950–955, 2010.
-  S. Li, H. Du, and X. Lin, “Finite-time consensus algorithm for multi-agent systems,”