Manuscript submitted to doi:10.3934/xx.

xxxxxxx
AIMS’ Journals
Volume X, Number 0X, XX 200X pp. X–XX

MULTI-AGENT SYSTEMS FOR QUADCOPTERS

Richard Carney, Monique Chyba∗ , Chris Gray, and George Wilkens

arXiv:2101.00033v2 [math.DS] 16 Mar 2021

Department of Mathematics, University of Hawaii at Manoa

2565 McCarthy Mall
Honolulu, Hawaii 96822, USA

Corey Shanbrom
Department of Mathematics and Statistics
California State University, Sacramento
6000 J St., Sacramento, CA 95819, USA

Abstract. Unmanned Aerial Vehicles (UAVs) have been increasingly used

in the context of remote sensing missions such as target search and tracking,
mapping, or surveillance monitoring. In the first part of our paper we consider
agent dynamics, network topologies, and collective behaviors. The objective is
to enable multiple UAVs to collaborate toward a common goal, as one would
find in a remote sensing setting. An agreement protocol is carried out by the
multi-agents using local information, and without external user input. The
second part of the paper focuses on the equations of motion for a specific type
of UAV, the quadcopter, and expresses them as an affine nonlinear control
system. Finally, we illustrate our work with a simulation of an agreement
protocol for dynamically sound quadcopters augmenting the particle graph
theoretic approach with orientation and a proper dynamics for quadcopters.

1. Introduction. Unmanned Aerial Vehicles (UAVs) present unique opportunities

to investigate unsafe or otherwise unreachable locations. UAVs are already being
used in numerous agricultural applications ([20, 17]), search and rescue ([18, 33]),
and conservation biology ([34]). While there are still many open issues in the use
of UAVs, they also have much to offer when it comes to disaster preparation and
response ([19]).
In Spring 2018, motivated by the 2018 Kilauea volcano eruption, we conducted
work in collaboration with NASA Jet Propulsion Laboratory (JPL) to develop a
real-time data relay system for UAVs. A relay system augmented with an interactive
visual aid for data viewing was developed and implemented to exploit the live
stream, including image superposition between sequential flights to rapidly provide
information about changes to the situation under study. In September 2018, we
demonstrated the concept at the University of Hawai’i at Mānoa with live streaming
of images from Bellows Air Force Station and video from Fissure 8 on Hawai’i Island.
We then developed on-board capability for image analysis for real time situation
assessment ([12]). In parallel, we also started to study motion planning for a swarm

2020 Mathematics Subject Classification. Primary: 70E60; 93A14; Secondary: 70E55.

Key words and phrases. Quadcopter, rigid body dynamics, graph Laplacian, multi-agent coor-
dination, network consensus, weighted network, decentralized control.
∗ Corresponding author: M. Chyba.

1
2 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

of UAVs ([15]), which is the problem we focus on in this paper. See [2] for a video
demonstrating some of the work.
Implementing a coordinated group of UAVs, rather than a single vehicle, is nat-
urally more efficient for many mission types. These include tracking, surveillance,
and mapping ([14]). In this regard, there are numerous reasons, driven by the
variety of applications, why collective control of a group of UAVs is important.
Common objectives include synchronization, formation, flocking, consensus, and
rendezvous. Rendezvous is the primary focus here ([21]). Moreover, autonomy of
individual UAVs provides a flexibility and adaptability that can be beneficial in
many kinds of applications. Optimization techniques, as an approach to collective
control, have the general drawback of scaling poorly with the number of vehicles.
We shall focus on a common alternative: the so-called graph theoretic methods.
The relative simplicity of these methods makes them an attractive choice in terms
of both scalability and the desire for real-time control ([25, 27]). Moreover, they
apply quite naturally to situations where there is an existing network of commu-
nications between the UAVs, although the existence of such is also an additional
assumption being made. The benefits include control that is decentralized in that
it occurs over the network without any advanced planning, and certain objectives,
such as rendezvous, can be carried out with only relative information being shared
across the network.
In this paper we focus on path-planning of the most common type of UAV, quad-
copters, and we focus on the dynamics of this specific type of UAV in Sections 3 and
4. Section 2, however, applies to any multi-agent UAV path configuration since we
treat them as particle points. We also occasionally refer to any UAV by the com-
mon, though imprecise, more general term “drone”. These small versatile vehicles
have proved to be a major asset in the increasing demand for complex autonomous
robotics missions, particularly those involving survey and object detection. Our
end-goal is developing, for a swarm of quadcopters, the ability to autonomously
make decisions based on shared information.
The outline of the paper is as follows. In Section 2 we discuss the multi-agent
autonomous agreement. Agents are considered as particles and the methodology is
based on graph theory. Section 3 introduces the dynamics of quadcopters and their
navigation. In Section 4 we approximate our particle multi-agent trajectories by
implementable ones satisfying the quadcopter’s dynamic equations of motion.
It is a genuine pleasure to have our article appear in this special issue that
honors Professor Bloch. The depth and breadth of his prodigious body of work
inspires and enriches each of us. On those occasions when the authors feel like
explorers venturing into uncharted territory, we often find a flag marking Tony’s
prior visitation. Indeed, Tony’s contributions to the subject of geometric control of
mechanical systems are keenly felt in this paper. His recent work on multi-agent
formations with obstacle and inter-agent avoidance in [13] aims toward a similar
goal as ours. Although the mathematical approach taken differs from the one here,
some of his work in [29] does apply a graph theoretic methodology. In addition,
the rigid body dynamics that we highlight in the second half of the paper is core
to much of his research ([3, 5, 4, 30]). With sincere appreciation and gratitude, the
authors offer Professor Bloch their very best wishes.
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 3

2. Multi-Agent Autonomous Agreement. Suppose a mission with multiple

agents requires that at some point the vehicles come together at a single location,
for instance at the end of a mission. With batteries possibly running low, a main
objective is to conserve power and rely on short-range communication. Autonomous
agreement is one of the fundamental problems in multi-agent coordination, where
a collection of agents are to agree on a joint state value.
2.1. Agreement Protocol. Our multi-agent systems will be viewed as networks,
as described in this section. Our primary goal is to highlight the existence of an
intricate relationship between the convergence properties of the so-called agreement
protocol on one hand, and the structure of the underlying agent network connections
on the other.
Definition 2.1. We define a network of agents to be a finite collection of agents
possessing some inter-connectivity via relative information exchange links. The
network is modeled by a graph, denoted G(V, E), where V is the set of vertices and
E is the set of edges. The agents correspond to the vertices of G(V, E) and the
relative information exchange links correspond to the edges.
In the sequel we denote by V = {v1 , v2 , . . . , vn } the vertices corresponding to the
n agents, labeled 1, 2, . . . , n. The set of edges is a subset E ⊂ V × V ; each edge has
the form (vi , vj ) = vi vj . We assume vi vj ∈ E implies vj vi ∈ E and that vi vi ∈ / E.
We shall enhance this initial model to accommodate time-varying weighted net-
works, motivated by the desire to account for the strength of communications be-
tween agents, or possibly to take into account how the strength of communications
evolves in the rendezvous problem as distances between agents change. Interestingly,
suitable assumptions on the weights imply the agreement value is an invariant, and
only the trajectories leading to it differ.
The agreement protocol assumes the rate of change of each agent’s state is gov-
erned by a sum of relative states with respect to a subset of neighboring agents.
Assume the scalar function q i (t) denotes the state of agent i for i = 1, . . . , n, and
define q(t) = (q 1 (t), · · · , q n (t))T . It follows that
X
q̇ i (t) = − q i (t) − q j (t) ,


1 ≤ i ≤ n, (1)
j∈N (i)

where N (i) is the neighborhood of a vertex vi , i.e., the collection of adjacent vertices
N (i) = {vj ∈ V | vi vj ∈ E}. If we adopt the notion that adjacency is symmetric,
i.e., vi vj ∈ E ⇐⇒ vj vi ∈ E, then we say the network is undirected and we can
encapsulate the above system in a single matrix equation:
q̇(t) = −L(G)q(t), (2)
n×n
where L(G) ∈ R is called the graph Laplacian. The graph Laplacian is uniquely
determined by the structure of G(V, E).
Definition 2.2. We refer to Equation (2) as the agreement dynamics, the agreement
protocol, or the consensus protocol.
2.2. Consensus Joint Value. The graph Laplacian has desirable properties when
G(V, E) is a connected, undirected network. In the rest of this section we will always
assume G(V, E) is a connected, undirected graph unless specified otherwise.
Proposition 1. Assume G(V, E) is a connected, undirected graph. The graph
Laplacian L(G) is a symmetric positive semidefinite matrix with eigenvalues 0 =
4 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

λ1 < λ2 ≤ · · · ≤ λn . Moreover, the n-dimensional vector of all ones, ~1, is an

eigenvector corresponding to the zero eigenvalue λ1 , i.e., L(G)~1 = ~0 .
Proof. This is a standard result and can be found in [22, 27, 28].

Let Λ = diag(0, λ2 , . . . , λn ) and let U = [û1 û2 . . . ûn ] (the hat notation indicates
a unit vector) be an n × n matrix consisting of orthonormal eigenvectors of L(G)
corresponding to the eigenvalue ordering above, i.e., U T U = I and L(G) U = U Λ.
Since we can factor the Laplacian as L(G) = U ΛU T , the solution to Equation (2)
with initial condition q(0) = q0 is given by
T
q(t) = e−L(G)t q0 = e−U ΛU t q0 = U e−Λt U T q0 ,
which can be written as:
n
X
q(t) = e−λi t (ûTi q0 )ûi . (3)
i=1

Definition 2.3. Given a set of n multi-agents, its agreement set A ⊂ Rn is defined

as the subspace span{~1}. A consensus joint state value is an element of A, and the
common value of its (identical) components is called its agreed state value.
The next proposition shows that when the state of a network of multi-agents sat-
isfies the agreement protocol, the state converges to the consensus joint state value
whose agreed state value is obtained by averaging the state’s initial components.
Proposition 2. Let G(V, E) be a network of n agents whose state q(t) satisfies the
~T
agreement protocol q̇ = −L(G)q. Then limt→∞ q(t) = α~1, where α = 1 nq(0) ∈ R is
the average of the initial states. In addition, since λ2 is L(G)’s smallest nonzero
eigenvalue, its magnitude will determine the rate of convergence to α~1 ([28, 22]).
Proof. Since λ1 = 0 and λi > 0 for i ≥ 2 we have, from Equation (3), that q(t)
~T
converges to the consensus joint state value (ûT1 q0 )û1 = 1 nq0 ~1 = α~1 ∈ A as t → ∞,
~T
where 1 nq0 = α ∈ R is its agreed state value. Moreover, Equation (3) implies that
eλ2 t (q i (t) − α) converges, as t → ∞, for all 1 ≤ i ≤ n.
~1T q0
The point lim q(t) = ~1 = α~1 ∈ Rn is precisely the orthogonal projection
t→∞ n
of q0 onto the agreement subspace and therefore minimizes the quantity kq − q0 k
over all possible q ∈ A, where k·k is the standard Euclidean distance. Furthermore,
1T q(t)
~
n is a constant of motion since
!
d ~1T q(t) ~1T q̇(t) ~1T (−L(G)q(t)) q T (t)LT (G)~1
= = =− (4)
dt n n n n
q T (t)L(G)~1 q T (t)~0
=− =− =0 (5)
n n
~T
for arbitrary values of t. Hence the agreed state value α = 1 nq0 is completely
determined at t = 0.
If the state of each agent is a vector quantity rather than a scalar value, the
agreement protocols for the components can still be written in a single matrix
equation. Indeed, assume we desire agreement on r scalar values. This means that
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 5

to each agent i we associate a vector value q i (t) = (q1i (t), · · · , qri (t)) and we let Q
be the n × r matrix whose rows are given by the q i ,
 1
q1 · · · qr1

 q12 · · · qr2 
Q= . (6)
 
.. .. 
 .. . . 
q1n ··· qrn

and we can write the agreement dynamics as:

Q̇(t) = −L(G)Q(t). (7)

We introduce q i (0) = (q1i (0), · · · , qri (0)) the initial row vector for the state of agent
i and, more importantly, qk (0) = (qk1 (0), · · · , qkn (0))T the initial column vector for a
given state value over all agents. Since L(G) = U ΛU T , we have that the solution of
T
(7) is given by: Q(t) = e−L(G)t Q(0) = e−U ΛU t Q(0) = U e−Λt U T Q(0), which can
be written as:
" n n
#
X X
Q(t) = e−λi t ûTi q1 (0)ûi · · · e−λi t ûTi qr (0)ûi (8)
i=1 i=1

n
X
= [e−λ1 t ûT1 q1 (0)û1 · · · e−λ1 t ûT1 qr (0)û1 ] + [e−λi t ûTi q1 (0)ûi · · · e−λi t ûTi qr (0)ûi ]
i=2

" # n
~1T q1 (0) ~T
~1 · · · 1 qr (0) ~1 +
X
= [e−λi t ûTi q1 (0)ûi · · · e−λi t ûTi qr (0)ûi ]
n n i=2

h i Xn
= α1~1 · · · αr~1 + [e−λi t ûTi q1 (0)ûi · · · e−λi t ûTi qr (0)ûi ] (9)
i=2

~T
where 1 qnk (0) = αk ∈ R is the agreed state value for each scalar state, 1 ≤ k ≤ r.
Note that here we assume the states qk , measured by the agents, are shared over
the same network G(V, E). Hence the states share a common graph Laplacian,
L(G). In more general situations each state variable could have a distinct net-
work, leading to an agreement protocol of the form: ẋ(t) = −L(Gx )x(t), . . . , ż(t) =
−L(Gz )z(t).

2.3. Example: Consider a collection of n agents required to meet at a single loca-

tion, not given in advance, and the agents do not have access to their global posi-
tions. Rather, all they can measure is their relative distance with respect to their
neighbors and they have to agree on their spatial coordinates x, y and z. In this case,
r = 3 and q i (t) = (q1i (t), q2i (t), q3i (t))T = (xi (t), y i (t), z i (t))T ∈ R3 with Q ∈ Rn×3 as
an example of the agreement protocol in action. By executing the agreement pro-
tocol, the convergence to a unique rendezvous point (αx , αy , αz ) is guaranteed for
any arbitrary initial conditions: (x1 (0), y 1 (0), z 1 (0)), . . . , (xn (0), y n (0), z n (0)) (i.e.,
any initial spatial configuration for the n agents) so long as G(V, E) is a connected
graph.
6 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

2.4. Multi-agent Trajectories. In this section, êj denotes the standard unit vec-
tor of all 0’s with a single 1 in the jth component. We use u1 to denote the vector
~1, which is an eigenvector of L(G) with eigenvalue λ1 = 0. With this notation,
~1
û1 = √ .
n
Lemma 2.4. Assume agent j is connected to every other agent, and set u1 = ~1.
Then, u1 − nêj is an eigenvector for L(G) with corresponding eigenvalue λn = n:
L(G)(u1 − nêj ) = n(u1 − nêj ). (10)
Proof. We have L(G)(u1 − nêj ) = L(G)u1 − nL(G)êj = 0 − nL(G)j = −nL(G)j ,
where L(G)j is the jth column of the Laplacian L(G). But since G(V, E) has the jth
node with n − 1 connections we know L(G)j = (−1, −1, . . . , n − 1, . . . , −1, −1)T =
nêj − u1 , i.e., all −1’s except an n − 1 in the jth component. Hence −nL(G)j =
n(u1 − nêj ) as claimed.

Definition 2.5. We say that vertex vi is fully connected if there exists an edge
between vi and every other vertex vj ∈ V, j 6= i. Therefore a fully connected vertex
has n − 1 edges.
~
Recall that û1 = √1n and λ1 = 0. Furthermore, it is proven in [1] that the
eigenvalues of L(G) are less than or equal to n and the multiplicity of the eigenvalue
λ = n is equal to the number of fully connected vertices.
Theorem 2.6. Suppose vertex vj is fully connected and that q(t) satisfies the agree-
ment protocol q̇ = −L(G)q. Then q j (t) is a straight line.
n
X
Proof. We know that q i = nα is a constant. We also know that:
i=1
X
q̇ j = −n q j + q i = −n(q j − α). (11)
This simple ODE (Newton’s Law of Cooling) implies that
q j − α = (q j (0) − α)e−nt (12)
which is a straight line.
Note that the proof is the same regardless of whether q j is a scalar or row j of an
n × r state matrix Q(t), provided it satisfies the agreement protocol Q̇ = −L(G)Q.
2.4.1. Rendezvous Mission, Unweighted Network. In this section we illustrate some
of our prior results with a rendezvous mission between four agents. We let q i (t) =
(xi (t), y i (t), z i (t)) denote the position of agent i in R3 . We assume the agents to be
initially distributed as follows:
q 1 (0) = (4, 17, 24), q 2 (0) = (18, 10, 32), q 3 (0) = (15, 10, 26), q 4 (0) = (4, 2, 35).
In this simulation we assume the communication network between the agents is
fixed and unweighted. From our algorithm, the vector of agreed state values (the
agreed rendezvous position) is given by:
q ∗ = (10.25, 9.75, 29.25). (13)
In Figure 1 we illustrate how two different communication networks produce differ-
ent trajectories, even though the agreed rendezvous position is the same for each
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 7

network since it depends only on the initial spatial positions of the agents. In the
first scenario (solid curve), we assume the following: agent 1 is connected to agent 2;
agent 2 is connected to agents 1, 3 and 4; agent 3 is connected to agents 2, 4; agent
4 is connected to agents 2, 3. This is represented by the following graph Laplacian:
 
1 −1 0 0
 −1 3 −1 −1 
L1 =  0 −1 2 −1  .
 (14)
0 −1 −1 2
In the second scenario (dashed curve), we assume: agent 1 is connected to agents
3, 4; agent 2 is connected to agent 3; agent 3 is connected to agents 1, 2; agent 4 is
connected to agent 1:
 
2 0 −1 −1
 0 1 −1 0 
L2 =  −1 −1 2
. (15)
0 
−1 0 0 1
In the first scenario agent 2 moves along a straight line since it is connected to
all other agents. In Figure 2 we compare the trajectories for agent 1 for each

Figure 1. Rendezvous Missions with Unweighted Network. Dis-

plays two different communication network scenarios for a 4-agent
rendezvous mission. Agreement positions coincide but trajectories
differ.

spatial coordinate. The eigenvalues for scenario 1 are given by {0, 1, 3, 4} and the
eigenvalues for scenario 2 are given by {0, 0.586, 2, 3.414}.

2.5. Weighted Network. We can imagine a scenario where it is desirable to ac-

count for the communications strength between vehicles, or for how that strength
evolves with time. Interestingly, the agreed state value is an invariant, and only the
trajectories leading to it are affected [28]. Note that not all extensions are immedi-
ately straightforward, however, since the explicit solution may be difficult to write
8 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

Figure 2. Comparison of the x, y, z-motions for agent 1 for the

scenarios of Fig. 1 corresponding to rendezvous missions with Un-
weighted Network.

down. Not all properties can be expected to carry over; for example, the linear
trajectory proof relied on cancellation due to the integer eigenvalues.
We now assume that the edges are weighted, i.e., edges carry a numerical weight.
If the graph Gw is a weighted graph, then the off-diagonal entries of the Laplacian
L(Gw ) will not be limited to only 0 and −1. Instead, the −1’s corresponding to
edges in the graph would be replaced with −wij = −wji to account for the weights.
The diagonal entries will once again be the absolute value of the row or column sum
of the non-diagonal entries. Equations (1) and (2) therefore become:
X
q̇ i (t) = − wij q i (t) − q j (t)


(16)
j∈N (i)

q̇(t) = −L(Gw )q(t). (17)

~
The proof that the consensus joint state value is given by α = 1qn0 ~1 comes from
the same logic as before since L(Gw ) is a symmetric positive semidefinite matrix
with eigenvalues 0 = λ1 < λ2 ≤ · · · ≤ λn and L(Gw )~1 = ~0. This is a consequence
of the fact that weighting preserves the properties that Gw (V, E) is connected and
undirected. Moreover, each row and column of L(Gw ) still sums to zero regardless
of whether or not the edges of Gw (V, E) are weighted. Hence, the values of λ2 , . . . λn
~
and û2 . . . ûn , will change, but we still have λ1 = 0 and û1 = √1n . Finally, in full
generality:

Theorem 2.7. Consider the dynamical system:

n
X
q̇ i (t) = − wij (t)[q i (t) − q j (t)] (1 ≤ i ≤ n) (18)
j=1

with initial conditions q(0) = (q 1 (0), · · · q n (0))T .

For each i, j, assume wij : [0, ∞) → [0, ∞) and wij (t) = wji (t).

Define the Laplacian matrix L(Gw )(t) by

 Pn
j=1 w1j (t) P−w12 (t) ... −w1n (t)

 −w12 (t) n
w
j=1 2j (t) ... −w2n (t) 
L(Gw )(t) =  . (19)
 
.. .. .. ..
 . . . Pn .

−w1n (t) −w2n (t) ... j=1 wnj (t)
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 9

Then the system of first order equations can be rewritten as: q̇(t) = −L(Gw )(t)q(t).
If the rank of L(Gw )(t) is n − 1 for all t then we have:
 X n 
1
lim q(t) = q i (0) ~1.
t→∞ n i=1

Proof. By assumption, the matrix L(Gw ) is real and symmetric, therefore Hermit-
ian. Furthermore, it is diagonally dominant with non-negative diagonal entries.
Therefore it is positive semi-definite. As for the unweighted case, λ1 = 0 with cor-
responding eigenvector ~1. Since we assume the graph to be connected throughout
the motion, the rank of L is always n − 1 and the multiplicity of the zero eigenvalue
is one. We have for any time t: 0 = λ1 < λ2 (t) ≤ · · · ≤ λn (t) with a corresponding
set of orthonormal eigenvectors û1 = √1n ~1, û2 (t), . . . , ûn (t).
 
Pn
Let q ∗ = n1 i=1 q i (0) ~1. Even if the wij : [0, ∞) → [0, ∞) are time-varying,
we can show with initial conditions q(0) that the trajectories q(t) → q ∗ as t → ∞.
A useful observation is that
Z t
q̇(t) = −L(Gw )(t)q(t) ⇒ q(t) − q(0) = − L(Gw )(s)q(s)ds
0
Z t 
⇒ ~1T q(t) − q(0) = −~1T


L(Gw )(s)q(s)ds
0
Z t Z t Z t
=− ~1 L(s)q(s)ds = −
T ~0 q(s)ds = −
T
0ds = 0,
0 0 0
so that we have ~1T q(t) = ~1T q(0) for all t ∈ [0, ∞).
Let V : Rn → R be defined by:
1 1
q 7→ V (q) = kq − q ∗ k2 = (q − q ∗ )T (q − q ∗ ).
2 2
Clearly, we have:
1. V (q) ≥ 0,
2. V (q) = 0 ⇐⇒ q = q ∗ ,
3. V̇ (q) = −(q − q ∗ )T L(Gw )(q − q ∗ ) ≤ 0
since −v T L(Gw )v ≤ 0 for any v ∈ Rn since L(Gw ) is positive semi-definite. Clearly
q = q ∗ implies V̇ (q) = 0, but also V̇ (q) = 0 implies q = q ∗ . Let A = span{~1} =
N (L(Gw )) represent the null space of L(Gw ) and suppose (q − q ∗ )T L(Gw )(q − q ∗ ) =
0. This implies q − q ∗ ∈ A. But we know that ~1T q(t) = ~1T q(0), and therefore
q − q ∗ ⊥ A since
~T ~T
~1T (q − q ∗ ) = ~1T q − 1 q(0) ~1 = ~1T q − 1 q(0) ~1T ~1 = ~1T q − ~1T q(0) = 0. (20)


n n
Finally, q − q ∗ ∈ A and q − q ∗ ⊥ A imply q − q ∗ = 0 ⇒ q = q ∗ . We conclude:
V̇ (q) = −(q − q ∗ )T L(Gw )(q − q ∗ ) = 0 ⇐⇒ q = q ∗ . (21)
Therefore 1, 2, 3 and Equation (21) imply that V (q) is a non-negative function
which is strictly decreasing for all trajectories q(t) 6= q ∗ hence q(t) → q ∗ as t → ∞,
i.e.,
 X n 
1
lim q(t) = qi (0) ~1
t→∞ n i=1
as claimed.
10 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

2.5.1. Rendezvous Mission, Weighted Network. Here we imagine a scenario taking

into account the strength of communications between agents. As previously stated,
the agreement value is an invariant, and only the trajectories leading to it will differ
compared to not taking the strength of the signal into account. We assume four
agents to be initially distributed as follows:
q 1 (0) = (2, 16, 29), q 2 (0) = (14, 20, 31), q 3 (0) = (10, 1, 25), q 4 (0) = (10, 14, 36).
The agreement rendezvous position is given by:
q ∗ = (9, 12.75, 30.25). (22)
In Figure 3 we illustrate how a weighted network impacts the trajectories. For both
scenarios we assume: agent 1 is connected to agent 2; agent 2 is connected to agents
1, 3 and 4; agent 3 is connected to agents 2, 4; agent 4 is connected to agents 2, 3
(as in Section 2.4.1):  
1 −1 0 0
 −1 3 −1 −1 
 0 −1 2 −1  .
L1 =   (23)
0 −1 −1 2
In the second scenario (dashed curves), we assume a weighted network with weights
due to initial distance:
 
12.8062 −12.8062 0 0
 −12.8062 41.9036 −20.3224 −8.7750 
L2 =  . (24)
 0 −20.3224 37.3518 −17.0294 
0 −8.7750 −17.0294 25.8044
It can be observed that for the weighted network, even though agent 2 is connected
to all other agents, it does not move along a straight line anymore.
In Figure 4 we compare the trajectories for agent 1 for each of the coordinates
of motion. The eigenvalues for the unweighted network are given by {0, 1, 3, 4} and
the eigenvalues for the weighted network are given by {0, 13.060, 42.248, 62.558}.
2.5.2. Rendezvous Mission, Time-varying Weighted Network. The goal of this ren-
dezvous mission is to compare a fixed network determined by initial proximity versus
a time-varying weighted network, where the weights are the inter-agent distances
that vary in time, and also edges are added as the preset proximity threshold is
reached.
We assume four agents to be initially distributed as follows:
q 1 (0) = (16, 5, 36), q 2 (0) = (19, 19, 29), q 3 (0) = (12, 16, 33), q 4 (0) = (14, 1, 26).
The agreement rendezvous position is given by:
q ∗ = (15.25, 10.25, 31). (25)
In Figure 5 we illustrate how a weighted network impacts the trajectories. For the
fixed network (solid curves) we have: agent 1 is connected to agents 3, 4; agent 2 is
connected to agent 3; agent 3 is connected to agents 1, 3; agent 4 is connected to
agent 1:  
2 0 −1 −1
 0 1 −1 0 
L1 =  −1 −1 2
. (26)
0 
−1 0 0 1
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 11

Figure 3. Comparison between trajectories on rendezvous mis-

sions with unweighted and weighted networks. The solid curves
represents the trajectories for the unweighted network and the
dashed ones for the weighted network. The two scenarios converge
to the same agreement value.

Figure 4. Comparison of the x, y, z-motions for agent 1 for the

rendezvous missions displayed in Figure 3. Observe that each com-
ponent converges much more rapidly for the weighted network, cor-
responding to the difference in their λ2 values: 13.060 > 1.

In the time-varying scenario (dashed lines), the network changes throughout the
trajectories. Here we provide the initial Laplacian and the final one:
 
23.0375 0 −12.0830 −10.9545
 0 8.6023 −8.6023 0 
L2i = 
 −12.0830 −8.6023 20.6854
, (27)
0 
−10.9545 0 0 10.9545
 
0.1465 −0.0523 −0.0482 −0.0460
 −0.0523 0.1501 −0.0450 −0.0528 
L2f =  −0.0482 −0.0450 0.1457 −0.0525  .
 (28)
−0.0460 −0.0528 −0.0525 0.1513
Initially the connectivity network is the same as for the fixed network but with
weights, however as the agents get closer it changes, and all agents are connected in
12 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

the end. The trajectories corresponding to the time-varying network here would be
easier to approximate for a real quadcopter since they are “straighter” (the sudden
jump in information leads to straighter paths toward the rendezvous, however in
this case, the weights prevent the perfect cancellation that would lead to true lin-
earity.) In Figure 6 we compare the trajectories for agent 1 for each coordinate of

Figure 5. Rendezvous missions comparing and unweighted net-

work (solid curves) to a time-varying weighted one (dashed curves).
They both agree on the consensus joint value. The points where
the dashed curves diverge from the solid ones correspond to the
addition of edges.

motion. The eigenvalues for the unweighted network are given by {0, 1, 3, 4} and the
eigenvalues for the weighted network are initially given by {0, 6.234, 19.385, 37.655},
and near the rendezvous agreement are given by {0, 0.190, 0.196, 0.207}.

Figure 6. Comparison of the x, y, z-motions for agent 1 for the

rendezvous mission of Figure 5. The components of the trajectory
corresponding to the time-varying network clearly converge more
rapidly.
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 13

3. Dynamics of quadcopters. In order to successfully implement multi-agent

motion planning, we need dynamically sound equations of motion governing the
movement of each drone. To this end, in this section we will derive such equations
under certain realistic symmetry assumptions regarding the shape of the drone body.
Utilizing the geometric formalism of [10], we present these equations in both first-
order and second-order forms, and provide a differential geometric version in terms
of the Levi-Civita connection on our configuration space equipped with a natural
Riemannian metric. Finally, we give the equations explicitly in coordinates. Note
that the notation in this section is independent from that in the previous section.

3.1. Setup. In this subsection we provide the basic physical notation and standard
derivation of the equations of motion for a rigid body in the body-fixed frame. A
rigid body can model a number of vehicles in space or in a fluid, such as satellites
or underwater vehicles, which can be stabilized by spinning rotors ([5, 6]). Here,
we assume our drone is nothing more than a rigid body in a fluid; later we will
make certain additional assumptions explicit. We closely follow the derivations in
[16, 31].
The configuration space for a rigid body moving in a fluid ([3]) is
Q = SE(3) ∼
= R3 × SO(3).
Here we equip Q with coordinates (b, R) where b ∈ R3 represents the position
of the center of mass of the body in space and R ∈ R3×3 is a rotation matrix
representing the orientation of the body aligned with the principal axes of inertia.
This coordinate system corresponds to some inertial frame of reference.
We regularly use the notation of the hat map, the Lie algebra isomorphism
ˆ: (R3 , ×) → (so(3), [, ])
 
0 −y3 y2
given by ŷz = y × z, equivalently ŷ =  y3 0 −y1 .
−y2 y1 0
First, choose a non-inertial body-fixed frame and set the following notation:
v = translational velocity in body frame
Ω = angular velocity in body frame
p = translational momentum in inertial frame
π = angular momentum in inertial frame
P = translational momentum in body frame
Π = angular momentum in body frame.
We seek to derive the dynamic equations of motion for the system in the body
frame. Here this will take the form of a system of ODEs in the variables (b, R, v, Ω) ∈
T Q. Our kinematic equations are:
ḃ = Rv (29)
Ṙ = RΩ̂. (30)
We transform momenta between the inertial and body frames according to:
p = RP (31)
π = RΠ + b̂p (32)
14 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

Begin by differentiating equations (31) and (32):

ṗ = ṘP + RṖ (33)
π̇ = ṘΠ + RΠ̇ + ḃ × p + b × ṗ. (34)
To rewrite these equations in a single frame, we must introduce additional quanti-
ties. In the inertial frame, let
fi = external forces, i = 1, . . . , k
τi = external torques, i = 1, . . . , l,
so that
k
X
EF = Rt fi = total external force in body frame (35)
i=1
l
X
ET = Rt τi = total external torque in body frame. (36)
i=1

Now in the inertial frame, the dynamics are solely due to external forces and torques:
k
X
ṗ = fi (37)
i=1
k
X l
X
π̇ = x̂i fi + τi , (38)
i=1 i=1

where xi is the vector from inertial frame origin to line of action of fi .

We then solve (33), (34) for Ṗ and Π̇ and simplify using equations (29), (30),
(31), (35), (36), (37), (38):
Ṗ = P̂ Ω + EF (39)
X
Π̇ = Π̂Ω + P̂ v + ET + Rt (xi − b) × fi . (40)
i

This gives the evolution of the momenta in the body frame, but the equations mix
momenta and velocities. In order to arrive at equations in terms of solely momenta
or velocities, we must find explicit formulas relating the two. This can be acheived
by the Legendre transform
∂T ∂T
P = , Π= .
∂v ∂Ω
To make this explicit, we introduce the kinetic energies
 t   
1 v mI3 −mr̂CG v
Tbody = (41)
2 Ω mr̂CG Jb Ω
 t 
Mf Cft
 
1 v v
Tf luid = (42)
2 Ω Cf Jf Ω
T = Tbody + Tf luid
 t 
mI3 + Mf −mr̂CG + Cft
 
1 v v
= (43)
2 Ω mr̂ CG + C f J b + J f Ω
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 15

where m is the mass of the body, rCG is vector from center of gravity to body frame
origin, Jb is the body inertia tensor, Jf is the added mass inertia tensor, Mf is the
added mass, Cf is the added cross terms, and I3 is the 3 × 3 identity matrix.
Then by differentiating the total kinetic energy (43), we find
∂T
P = = (mI3 + Mf )v + (−mr̂CG + Cft )Ω (44)
∂v
∂T
Π= = (mr̂CG + Cf )v + (Jb + Jf )Ω. (45)
∂Ω
These equations transform velocities into momenta.
We now assume the body has an especially nice shape, which is realistic for most
drones.
Assumption 1. Assume the body has three planes of symmetry and the prin-
ciple axes of inertial coincide with the body frame axes. This implies Jb , Jf , and
Mf are all diagonal and Cf = 0. When we consider gravity in Section 3.2, we also
assume that this symmetry extends to the mass density of the body.
Assumption 2. Assume the center of gravity of the body coincides with the
origin of the body frame. This implies r̂CG = 0.
With these two assumptions, equations (44) and (45) reduce to
P = Mv (46)
Π = JΩ, (47)
where M = mI3 +Mf is the total mass of the system and J = Jb +Jf is the moment
of inertia of the system.
Now differentiate (46) and (47) and simplify using (39), (40), (46) and (47) to
obtain the derivatives of the velocities expressed solely in terms of positions and
velocities, without momenta:
mv̇ = M v × Ω + EF (48)
X
t
J Ω̇ = JΩ × Ω + M v × v + ET + R (xi − b) × fi . (49)
i

Finally, since for drones our fluid is air, we make the following simplifying as-
sumption.
Assumption 3. Assume the added fluid is negligible. This implies Mf = Jf = 0.
Then J = Jb , and since M = mI3 is a scalar matrix, the fictitious force M v × v = 0.
Combining this with (29), (30), (48), (49), we obtain the complete equations
of motion for a rigid drone in air as a first order system on T Q coordinatized by
(b, R, v, Ω):
ḃ = Rv (50)
Ṙ = RΩ̂ (51)
mv̇ = mv × Ω + EF (52)
k
X
J Ω̇ = JΩ × Ω + ET + Rt (xi − b) × fi . (53)
i=1

3.2. External forces and torques. In this section we explicitly work out the
external terms EF and ET for a drone in the body frame. We have four rotors
producing thrust in a quadcopter, with the ith rotor possessing angular velocity ωi
16 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

producing thrust ti = Kr ωi2 , where Kr is the thrust coefficient ([26]). Here we are
implicitly treating the ωi as our controls.
For the external forces, we follow [32]. The force due to drag is
f1 = −diag(v1 |v1 |, v2 |v2 |, v3 |v3 |)CD ,
where vj is the jth component of the linear velocity v and CD is the vector of
translational drag coefficients in the body frame. The force due to gravity is
f2 = −mgRt eI3 ,
where g is the gravitational constant and the third inertial frame basis vector eI3 rep-
resents the opposite direction of gravitational attraction. Lastly, the force produced
by thrust is
X4
f3 = eB
3 ti ,
i=1

where the third body frame basis vector eB 3 represents the direction in which each
of the motors produce thrust. Thus EF = f1 + f2 + f3 .
Ostensibly, each of these forces induces a torque, represented by the last term
t
Pk
R i=1 (xi − b) × fi in (53). However, for f1 and f2 , no torque is induced as the
line of force passes through the center of mass of the body due to our symmetry
assumptions. The force f3 , however, represents the sum of four individual forces
located at each of the four rotors. Each of these forces induces a torque, and their
sum gives the net torque generated by the rotors:
Kr d(ω32 − ω12 )
 
2 2
τf =  K r d(ω4 − ω2 ) ,
P 4 i+1 2
Kd i=1 (−1) ωi
where Kd is the propeller drag coefficient and d is the distance from the drone center
of mass to the rotation axis of each rotor.
The external torques are computed as in [8, 32]. The torque in the body frame
generated by drag is
τ1 = −diag(Ω1 |Ω1 |, Ω2 |Ω2 |, Ω3 |Ω3 |)Cτ ,
where Ωj is the jth component of the angular velocity Ω and Cτ is the vector of
rotational drag coefficients in the body frame. The torque due to gyroscopic effects
is
X4
τ2 = Ω × (−1)i+1 Jr (0, 0, ωi )t ,
i=1
where Jr is the moment of inertia for a rotor. Assuming the rotor has nontrivial
inertia about only the vertical axis, Jr has only one nonzero entry which we denote
by J¯r , so that Jr = diag(0, 0, J¯r ). We then have ET = τ1 + τ2 .
We can now give our updated equations of motion for a quadcopter:
ḃ = Rv (54)
Ṙ = RΩ̂ (55)
mv̇ = mv × Ω + f1 + f2 + f3 (56)
J Ω̇ = JΩ × Ω + τf + τ1 + τ2 . (57)
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 17

3.3. Coordinate expression. Here we work in coordinates η = (b1 , b2 , b3 , φ, θ, ψ)

on Q = SE(3) representing the position and orientation of the drone in the in-
ertial frame. The positions b1 , b2 , b3 are the standard coordinates for R3 . The
angles φ, θ, ψ are Tait-Bryan angles, known in aeronautics as roll, pitch, and yaw
respectively, and sometimes referred to as Euler angles (although not the proper
or classical kind). The rotation angles are given intrinsically by the sequence
z − y 0 − x00 or extrinsically by the sequence x − y − z. Here we take (φ, θ, ψ) ∈
(−π, π) × (−π/2, π/2) × (−π, π). These coordinates on Q induce natural coordi-
nates for the velocities (v, Ω) = (v1 , v2 , v3 , Ω1 , Ω2 , Ω3 ).
Now let
 
Cψ Cθ Cψ Sθ Sφ − Sψ Cφ Cψ Sθ Cφ + Sψ Sφ
R(η) =  Sψ Cθ Sψ Sθ Sφ + Cψ Cφ Sψ Sθ Cφ − Cψ Sφ  ,
−Sθ Cθ Sφ Cθ Cφ

and
 
1 Sφ Tθ Cφ Tθ
Θ(η) = 0 Cφ −Sφ  ,
Sφ Cφ
0 Cθ Cθ

where Cx = cos(x), Sx = sin(x) and Tx = tan(x). We can write R as the composi-

tion of three pure rotations as follows. Let Ri,α denote rotation by α radians about
the ith principal axis of inertia. Explicitly,

 
1 0 0
R1,φ = 0 Cφ −Sφ  ,
0 Sφ Cφ
 
Cθ 0 Sθ
R2,θ =  0 1 0 ,
−Sθ 0 Cθ
 
Cψ −Sψ 0
R3,ψ =  Sψ Cψ 0 .
0 0 1

Then R = R3,ψ R2,θ R1,φ .

Now R transforms linear velocity in the body frame to linear velocity in the
inertial frame, and Θ transforms the body-fixed angular velocity (Ω1 , Ω2 , Ω3 ) into
the Euler rate vector (φ̇, θ̇, ψ̇). Thus our kinematic equations (54) and (55) take the
form
  
R(η) 0 v
η̇ = .
0 Θ(η) Ω

Substituting R into equations (56) and (57) and expanding yields the following.
18 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

Lemma 3.1. In coordinates developed above, the equations of motion for a quad-
copter take the form
ḃ1 = v1 Cψ Cθ + v2 (Cψ Sθ Sφ − Sψ Cφ ) + v3 (Cψ Sθ Cφ + Sψ Sφ ) (58)
ḃ2 = v1 Sψ Cθ + v2 (Sψ Sθ Sφ + Cψ Cφ ) + v3 (Sψ Sθ Cφ − Cψ Sφ ) (59)
ḃ3 = −v1 Sθ + v2 Cθ Sφ + v3 Cθ Cφ (60)
φ̇ = Ω1 + Ω2 Sφ Tθ + Ω3 Cφ Tθ (61)
θ̇ = Ω2 Cφ − Ω3 Sφ (62)
Sφ Cφ
ψ̇ = Ω2 + Ω3 (63)
Cθ Cθ
1
v̇1 = v2 Ω3 − v3 Ω2 − v1 |v1 |CD1 + gSθ (64)
m
1
v̇2 = v3 Ω1 − v1 Ω3 − v2 |v2 |CD2 − gCθ Sφ (65)
m !
4
1 X
v̇3 = v 1 Ω 2 − v 2 Ω1 + ti − v3 |v3 |CD3 − gCθ Cφ (66)
m i=1
" 4
#
1 X
Ω̇1 = (J2 − J3 )Ω2 Ω3 + J¯r (−1)i+1 Ω2 ωi + Kr d(ω32 − ω12 ) − Ω1 |Ω1 |Cτ1
J1 i=1
(67)
" 4
#
1 X
Ω̇2 = (J3 − J1 )Ω1 Ω3 − J¯r (−1)i+1 Ω1 ωi + Kr d(ω42 − ω22 ) − Ω2 |Ω2 |Cτ2
J2 i=1
(68)
" 4
#
1 X
Ω̇3 = (J1 − J2 )Ω1 Ω2 + Kd (−1)i+1 ωi2 − Ω3 |Ω3 |Cτ3 . (69)
J3 i=1

3.4. Geometric formulation. Here we will equip Q with a natural metric allowing
us to characterize the equations of motion as a second order system on Q. We then
lift the dynamics to a first order system on T Q.
Under our assumptions, the kinetic energy (43) can be encoded as the matrix
 
mI3 0
G=
0 J
where m and J denote the mass and inertia tensor of the drone, respectively. This
is a (diagonal) positive definite symmetric matrix and induces a Riemannian metric
on Q, which in turn yields the Levi-Civita connection ∇. This allows us to define the
acceleration of a curve in Q as ∇γ̇ γ̇, where a curve γ(t) = (b(t), R(t)) has velocity
γ̇ = (b(t), R(t), v(t), Ω(t)). Explicitly, we compute
 
v̇ + Ω × v
∇γ̇ γ̇ = .
Ω̇ + J −1 (Ω × JΩ)
This connection evidently contains the fictitious forces, and is independent of choice
of coordinates. The geodesics for ∇ (curves with zero acceleration) represent mo-
tions of a drone subject to no external forces.
Now in general, a Riemannian metric G on a manifold Q induces the “musical
isomorphisms” G] : T ∗ Q → T Q and G[ : T Q → T ∗ Q. Newton’s equation can be
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 19

expressed geometrically as F (γ) = G[ ∇γ̇ γ̇ where F is some force. In our setting,

this produces the geometric version of our equations of motion:
∇γ̇ γ̇ = G] F (γ) (70)
where
m−1 I3
 
0
G] =
0 J −1
and  
f1 + f2 + f3
F (b, R) = .
τf + τ1 + τ2
Equation (70) represents a second order system on Q. We now express this as
a first order system on T Q. To that end, we invoke the vertical lift vlftw : Tq Q →
Tw T Q where w ∈ Tq Q, and the geodesic spray S : T Q → T T Q ([23, 16]). For our
system, we can explicitly compute
 
v
 Ω 
S(b, R, v, Ω) =  .
 v×Ω 
J −1 (JΩ × Ω)
Here the components are expressed relative to the standard left-invariant basis of
vector fields {Yk } on T Q = T SE(3), where at a point (b, R) ∈ SE(3) we have
Yi = (Rei , 0) for i = 1, 2, 3 and Yj+3 = (0, Rêj ) for j = 1, 2, 3.
Then our equations of motion can be expressed by the first order system
Γ̇ = X(Γ) (71)
where Γ is a curve in T Q and X is the vector field on T Q given by
X = S + vlft(G] F ). (72)
Next, we express this system as an affine nonlinear control system as in [3]. We
can rewrite equation (71) as
   
v 0
 Ω   0 
Γ̇ = 
v × Ω + m−1 f1 + m−1 f2  + 
, (73)
m−1 f3
 

J −1 (JΩ × Ω) + J −1 τ1 J −1 τf + J −1 τ2
where the first summand is the drift, and the second summand contains all the
controls ωi . Explicitly, we have
4
X
m−1 f3 = m−1 Kr ωi2 eB
3
i=1
Kr d(ω32 − ω12 )
   
X4 0
J −1 τf + J −1 τ2 = J −1  KP4
2 2
r d(ω4 − ω2 )  + J −1 Ω × (−1)i+1
J r
 0 .
Kd i=1 (−1)i+1 ωi2 i=1 ωi
Note that the torque due to gyroscopic effects is linear in ωi , while the force and
torque due to thrust are linear in ωi2 . Consequently we cannot write this system as
an affine control system in independent controls. As is common in the literature (e.g.
[26, 32]) we now simplify by ignoring the gyroscopic effects in τ2 . This simplification
does not affect the simple motion planning in Section 4.1, but the gyroscopic effects
are present in the simulations of Section 4.2. Consider the controls
ui = ωi2 for i = 1, 2, 3, 4.
20 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

Proposition 3. Ignoring gyroscopic effects, the quadcopter can be expressed as an

affine nonlinear control system as:
4
X
Γ̇ = f (Γ) + gi (Γ)ui (74)
i=1

where the drift vector field is

 
v
 Ω 
f (Γ) = 
v × Ω + m−1 f1 + m−1 f2 


J −1 (JΩ × Ω) + J −1 τ1
and the control vector fields are
   
0 0
 0   0 
g1 =  (0, 0, m−1 Kr )t  , g2 =  (0, 0, m−1 Kr )t  ,
  

(−Kr d, 0, Kd )t (0, −Kr d, −Kd )t

   
0 0
 0   0 
g3 = (0, 0, m−1 Kr )t  , g4 =  (0, 0, m−1 Kr )t  .
  

(Kr d, 0, Kd )t (0, Kr d, −Kd )t

4. Quadcopter multi-agent dynamics.

4.1. Single-agent motion planning. We first explore some simple motions for a
single quadcopter. We use the coordinates and notation from Section 3.
Lemma 4.1. Suppose a quadcopter undergoes yaw only, with no translational or
other rotational motion. Then
r
mg
ω1 = ω3 = − ω22 and ω4 = ω2 ,
2Kr
where ω2 is a free parameter.
Proof. The desired motion keeps all variables constant except Ω3 and ψ. This
imposes ω1 = ω3 and ω2 = ω4 as well as
ḃ = v1 = v2 = θ = φ = Ω1 = Ω2 = 0
for the duration of the motion.
The equations of motion Pin Lemma 3.1 simplify considerably; in particular, we
1 4
have ḃ3 = v3 and v̇3 = m i=1 ti − g. Assuming ḃ3 = 0 gives the expected result
P4 2
for hovering: qti = mg. Using ω1 = ω3 and ω2 = ω4 with ti = Kr ωi allows us
i=1
mg
to solve for ω1 = 2Kr − ω22 .

Lemma 4.2. Suppose a quadcopter moves only in the direction of the first body
axis. Then
s 
1 2
6a − 8Kr ω4 1
ω2 =  − ω4  and ω1 = ω3 = (ω2 + ω4 ),
3 Kr 2
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 21

where
 
1
a=m (v̇1 Sθ + v1 Cθ Ω2 + v3 Sθ Ω2 ) − v1 Ω2 + gCθ + v3 |v3 |CD3
Cθ
and ω4 is a free parameter.
Proof. We desire to move only along the body x-axis with no motion in the body
y or z direction, requiring the body to pitch without any yaw or roll. This imposes
ω1 = ω3 as well as
ḃ3 = φ = V2 = Ω3 = Ω1 = 0
for the duration of the motion.
The equations of motion in Lemma 3.1 simplify considerably; in particular, we
have θ̇ = Ω2 and ḃ3 = −V1 Sθ + V3 Cθ . Using ω1 = ω3 and setting Ω̇1 = 0 forces
ω1 = ω3 = 21 (ω2 + ω4 ). Setting b̈3 = 0 and using θ̇ = Ω2 gives
1
v̇3 = (v̇1 Sθ + v1 Cθ Ω2 + v3 Sθ Ω2 ).
Cθ
Setting this equal to the equation of motion for v̇3 , we find
4  
X 1
ti = m (v̇1 Sθ + v1 Cθ Ω2 + v3 Sθ Ω2 ) − v1 Ω2 + gCθ + v3 |v3 |CD3 .
i=1
Cθ
Denote the right side of this equation by a. Finally, using ti = Kr ωi2 , we can solve
for s 
1  6a − 8Kr ω42
ω2 = − ω4  .
3 Kr

Note that in Lemma 4.2 we could alternatively treat ω4 as the free parameter
and solve for an identical expression for ω4 in terms of ω2 . We have similar choices
for the yaw motion in Lemma 4.1.
Moreover, a nearly identical calculation gives path planning along the second
body axis. Treating ω3 as the free parameter, we can move only in the body y
direction using
s 
1  6a − 8Kr ω32 1
ω1 = − ω3  and ω2 = (ω1 + ω3 ).
3 Kr 2

In order to move in another direction in the body xy-plane (with no motion in the
z direction), we can simply combine the motions in Lemmas 4.1 and 4.2.
4.2. Simulations.
4.2.1. Single agent. We demonstrate the simple motions from Section 4.1 in Fig-
ures 7 and 8. In the coordinates of Section 3.3, the drone begins at (0, 0, 0, 0, 0, 0) ∈
SE(3) with motors producing the necessary amount of thrust to remain at a hov-
ering equilibrium. Over the first 12 seconds, the drone maintains orientation while
rising vertically in the body (and inertial) z-direction, ending at another hover.
Over the next four seconds, the drone maintains position while rotating π/2 radi-
ans counterclockwise about the body z-axis as described in Lemma 4.1. Finally, for
the last four seconds the drone executes the motion in Lemma 4.2, flying directly
along the body x-axis. Figure 7 shows the path of the drone in space as well as
22 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

Figure 7. A quadcopter rising straight up, then yawing while hov-

ering, then flying straight along the body x-axis. Top: position in
space. Bottom: orientation angles over time.

the orientation angles over time. Figure 8 shows the controls used to produce this
motion.

4.2.2. Multi-agent. Here we combine most of the work completed above in order
to simulate an agreement protocol for dynamically sound quadcopters. Consider
three drones on a rendezvous mission with given initial conditions. We assume
each drone begins hovering with all translational and angular velocities zero. This
forces each initial pitch and roll to also be zero, but we choose different initial yaws.
For simplicity, we assume all three drones start in the z = 0 plane (which need not
correspond to the ground) but choose different initial x, y positions. This simplifying
assumption does not sacrifice much generality: if the three drones started at three
different nonzero altitudes, we would simply find the average initial altitude and
execute simple motions bringing each drone up or down to this height.
In the coordinates (b1 , b2 , b3 , φ, θ, ψ) ∈ SE(3) from Section 3.3, we choose our
three drones to have initial configurations (0, 0, 0, 0, 0, 0), (0, 9, 0, 0, 0, −π/4), and
(15, 9, 0, 0, 0, π/2).
For this example we will assume a complete network G = K3 , so by Theorem 2.6
we have that all trajectories are straight lines. In particular, each drone will traverse
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 23

Figure 8. The controls used to produce the motion in Figure 7.

Top: angular velocities of each of the four motors over time. Bot-
tom: total thrust over time.

the line segment connecting their initial position in R3 to the rendezvous position,
which is simply the component-wise average of the initial conditions: (5, 6, 0). Figure
9 shows these trajectories.
There are, however, two different methods for parametrizing these line segments
and hence planning the drones’ flights. First, we could use the methods from Section
2, obtaining trajectories which exponentially approach the rendezvous position. In
this setting, the drones’ orientations are ignored completely, as are all dynamics,
giving a simple yet unrealistic path.
A main objective of this project was to investigate a second method: implement
the path-planning from Section 4.1 to find controls bringing the drones along the
desired line segments in a manner that respects the realistic dynamics from Section
3. This requires each drone beginning its flight by executing a pure yaw rotation
to orient its body x-axis toward the rendezvous position, then fly straight in this
direction. See Figure 10 to see the yaw for each drone over time; note that the total
flight times differ for each drone. Also note that for each drone, the roll over time is
identically zero, but the pitch is non-zero in order to produce lateral displacement
(compare with Figure 7).
24 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

Figure 9. Three drones start at position (0, 0, 0), (0, 9, 0) and

(15, 9, 0) with initial yaw angles 0, −π/4, and π/2. Their trajecto-
ries to the rendezvous position are shown.

Figure 10. The yaw ψ over time for each drone. Total flight times
differ for each drone.

In order to create the desired trajectories in accordance with our equations of

motion, we utilize Lemmas 4.1 and 4.2. The resulting controls – the four motor
speeds for each of the three drones – are shown in Figure 11, as is the total thrust
for each drone. The parameters used in this simulation were taken from [7], and
appear in Table 1.
Finally, in Figure 12 we compare the paths generated by the two methods above.
The solid curves represent the simple but unrealistic paths derived in Section 2. The
dashed curves represent the dynamically sound paths in Figures 10 and 11. Both
methods trace the same curves shown in Figure 9, but differ in their parametriza-
tions. The top plot shows the x-coordinate over time; the solid curves for drones
1 and 2 coincide as both start with an initial x-value of 0. The bottom plot shows
the y-coordinate over time; the solid curves for drones 2 and 3 coincide as both
start with an initial y-value of 9. Note that the solid curves approach (but do not
reach) the rendezvous position much faster than the dashed curves as they are un-
constrained by orientation or realistic acceleration. To move from the simulation
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 25

Figure 11. The controls used to produce the motions in Figures

9 and 10. We show the four motor speeds and total thrust for each
drone as functions of time.
26 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

Table 1. Parameters used in simulations

Constant Symbol Value

drone mass m 0.468
drone inertia J diag((3.8278, 3.8288, 7.6566) · 10−3 )
rotor inertia Jr diag(0, 0, 2.8385 · 10−5 )
distance to rotor d 0.25
thrust coefficient Kr 2.9842 · 10−5
translational drag CD (5.5670, 5.5670, 6.3540) · 10−4
rotational drag Cτ (5.5670, 5.5670, 6.3540) · 10−4
propeller drag Kd 3.2320 · 10−7

Figure 12. Comparison of two methods of motion planning.

Dashed curves include dynamics; solid curves do not. Left: x-
coordinate over time. Right: y-coordinate over time.

environment to a more realistic experiment with actual drones would require a more
careful consideration of the observer design as in [9].

5. Conclusion. The primary goals of the paper were twofold. First, we intro-
duce the agreement protocol from a graph theoretic viewpoint and apply it to a
rendezvous mission for a set of particle agents without orientation. Second, we de-
rive the equations of motion for a quadcopter to take into account its orientation
and dynamics. We then apply the quadcopter dynamics to a rendezvous mission
and compare the trajectories with the ones from the network multi-agents algo-
rithm. This is a first step towards developing more complex multi-agent systems
for quadcopters. A next step will be to analyze the decoupling vector fields for
the quadcopter from its affine connection control system formulation and use these
trajectories to approximate geometric trajectories obtained from the multi-agents
network approach. We focused here on the agreement protocol, but ongoing work
shows that it can be extended to formation flying for multi-agents and more com-
plex missions, such as search and rescue with a large team of quadcopters including
leaders and followers. We are also planning to run real drone flights to demonstrate
the applicability of our techniques.
 MULTI-AGENT SYSTEMS FOR QUADCOPTERS 27

REFERENCES

[1] W. Anderson Jr. and T. Morley, Eigenvalues of the Laplacian of a graph, Linear and Multi-
linear Algebra, 18 (1985), 141–145.
[2] AUVs on-board capability for image analysis for real time situation assessment, https://
www.youtube.com/watch?v=qo0VWWvTgdU
[3] A. Bloch, Nonholonomic Mechanics and Control, 2nd edition, Springer-Verlag, New York,
2015.
[4] A. Bloch, I. Hussein, M. Leok, and A. Sanyal, Geometric structure-preserving optimal control
of the rigid body, Journal of Dynamical and Control Systems, 15 (2009), 307–330.
[5] A. Bloch, P. Krishnaprasad, J. Marsden, and G. Sanchez De Alvarez, Stabilization of rigid
body dynamics by internal and external torques, Automatica, 28 (1992), 745–756.
[6] A. Bloch, N. Leonard, and J. Marsden, Controlled Lagrangians and the stabilization of me-
chanical systems I: The first matching theorem, IEEE Trans. Automat. Control, 45 (2000),
2253–2270.
[7] H. Bouadi, M. Bouchoucha, and M. Tadjine, Sliding mode control based on backstepping
approach for an UAV type-quadrotor, International Journal of Mechanical and Mechatronics
Engineering, 1 (2007), 39–44.
[8] H. Bouadi and M. Tadjine, Nonlinear observer design and sliding mode control of four rotors
helicopter, International Journal of Aerospace and Mechanical Engineering, 1 (2007), 354–
359.
[9] F. Bullo, J. Cortes, and S. Martinez, Distributed Control of Robotic Networks, Princeton
University Press, 2009.
[10] F. Bullo and A. Lewis, Geometric Control of Mechanical Systems, Springer-Verlag, New York,
2004.
[11] Y. Cao, W. Yu, W. Ren and G. Chen, An overview of recent progress in the study of dis-
tributed multi-agent coordination, IEEE Transactions on Industrial Informatics, 9 (2013),
427–438.
[12] R. Carney, M. Chyba et al., Multi-agents path planning for a swarm of unmanned aerial
vehicles, 2020 IEEE International Geoscience and Remote Sensing Symposium, (2020).
[13] R. Chandrasekaran, L. Colombo, M. Camarinha, R. Banavar, and A. Bloch, Variational colli-
sion and obstacle avoidance of multi-agent systems on Riemannian manifolds, 2020 European
Control Conference, (2020), 1689–1694.
[14] S. Chung, A. Paranjape, P. Dames, S. Shen, and V. Kumar, A survey on aerial swarm robotics,
IEEE Transactions on Robotics, 34 (2018), 837–855.
[15] M. Chyba, R. Carney, C. Gray, and Z. Trimble, Increasing small unmanned aerial system
real-time autonomy, 2020 IEEE International Geoscience and Remote Sensing Symposium,
(2020).
[16] M. Chyba, T. Haberkorn, R. Smith, and G. Wilkens, A geometric analysis of trajectory design
for underwater vehicles, Discrete Contin. Dyn. Syst. Ser. B, 11 (2009), 233–262.
[17] F. Costa, J. Ueyama, T. Braun, G. Pessin, F. Osório and P. Vargas, The use of unmanned
aerial vehicles and wireless sensor network in agricultural applications, IEEE International
Symposium on Geoscience and Remote Sensing, (2012).
[18] P. Doherty and P. Rudol, A UAV search and rescue scenario with human body detection and
geolocalization, Australian Conference on Artificial Intelligence, (2007).
[19] M. Erdelj and E. Natalizio, UAV-assisted disaster management: applications and open issues,
2016 International Conference on Computing, Networking and Communications (ICNC),
(2016).
[20] J. Kim, S. Kim, C. Ju and H. Son, Unmanned aerial vehicles in agriculture: a review of
perspective of platform, control, and applications, IEEE Access, 7 (2019).
[21] S. Knorn, Z. Chen and R. Middleton, Collective control of multiagent systems, IEEE Trans-
actions on Control of Network Systems, 3 (2016), 334–347.
[22] G. Lafferriere, A. Williams, J. Caughman, and J. Veerman, Decentralized control of vehicle
formations, Systems & Control Letters, 54 (2005), 899–910.
[23] A. Lewis and R. Murray, Configuration controllability of simple mechanical control systems,
SIAM Journal on Control and Optimization, 35 (1997), 766–790.
[24] T. Luukkonen, Modelling and control of quadcopter, Independent research project, Aalto
University in Espoo, Finland, 2011.
28 CARNEY, CHYBA, GRAY, SHANBROM AND WILKENS

[25] M. Mesbahi and M. Egerstedt, Graph Theoretic Methods in Multiagent Networks, Princeton
University Press, 2010.
[26] M. Mueller and T. D’Andrea, Stability and control of a quadrocopter despite the complete
loss of one, two, or three propellers, 2014 IEEE international conference on robotics and
automation (ICRA), (2014), 45–52.
[27] I. Okoloko, Path planning for multiple spacecraft using consensus with LMI avoidance con-
straints, IEEE Aerospace Conference, (2012), 1–8.
[28] R. Olfati-Saber, J. Fax, and R. Murray, Consensus and cooperation in networked multi-agent
systems, Proceedings of the IEEE, 95 (2007), 215–233.
[29] P. Rooney, A. Bloch, and C. Rangan, Trees, forests, and stationary states of quantum Lindblad
systems, arXiv:1810.11144, (2018).
[30] A. Sanyal, A. Bloch, and N. McClamroch, Dynamics of multibody systems in planar motion
in a central gravitational field, Dynamical Systems,19 (2004),303–343.
[31] R. Smith, M. Chyba, G. Wilkens, and C. Catone A geometrical approach to the motion
planning problem for a submerged rigid body, International Journal of Control, 82 (2009),
1641–1656.
[32] V. Stepanyan and K. Krishnakumar, Estimation, navigation and control of multi-rotor drones
in an urban wind field, AIAA Information Systems-AIAA Infotech @ Aerospace, (2017).
[33] S. Waharte and N. Trigoni, Supporting search and rescue operations with UAVs, 2010 Inter-
national Conference on Emerging Security Technologies, (2010).
[34] S. Wich and L. Koh, Conservation Drones: Mapping and Monitoring Biodiversity, Oxford
University Press, New York, 2018.

Received xxxx 20xx; revised xxxx 20xx.

E-mail address: rscarney@hawaii.edu
E-mail address: chyba@hawaii.edu
E-mail address: cgray8@hawaii.edu
E-mail address: grw@math.hawaii.edu
E-mail address: corey.shanbrom@csus.edu