Towards A Swarm of Agile Micro Quadrotors
Alex Kushleyev, Daniel Mellinger, Vijay Kumar
GRASP Lab, University of Pennsylvania
Abstract—We describe a prototype 73 gram, 21 cm diameter
micro quadrotor with onboard attitude estimation and control
that operates autonomously with an external localization system.
We argue that the reduction in size leads to agility and the
ability to operate in tight formations and provide experimental
arguments in support of this claim. The robot is shown to be
capable of 1850
◦
/sec roll and pitch, performs a 360
◦
flip in 0.4
seconds and exhibits a lateral step response of 1 body length
in 1 second. We describe the architecture and algorithms to
coordinate a team of quadrotors, organize them into groups and
fly through known three-dimensional environments. We provide
experimental results for a team of 20 micro quadrotors.
I. INTRODUCTION
The last decade has seen rapid progress in micro aerial
robots, autonomous aerial vehicles that are smaller than 1
meter in scale and 1 kg or less in mass. Winged aircrafts
can range from fixed-wing vehicles [14] to flapping-wing
vehicles [6], the latter mostly inspired by insect flight. Rotor
crafts, including helicopters, coaxial rotor crafts [9], ducted
fans[22], quadrotors [10] and hexarotors, have proved to be
more mature [15] with quadrotors being the most commonly
used aerial platform in robotics research labs. In this class,
the Hummingbird quadrotor sold by Ascending Technologies,
GmbH [2], with a tip-to-tip wingspan of 55 cm, a height of
8 cm, mass of about 500 grams including a Lithium Polymer
battery and consuming about 75 Watts is a remarkably capable
and robust platform as shown in [16, 17].
Of course micro aerial robots have a fundamental payload
limitation that is difficult to overcome in many practical
applications. However larger payloads can be manipulated
and transported by multiple UAVs either using grippers or
cables [20]. Applications such as surveillance or search and
rescue that require coverage of large areas or imagery from
multiple sensors can be addressed by coordinating multiple
UAVs, each with different sensors.
Our interest in this paper is scaling down the quadrotor
platform to develop a truly small micro UAV. The most
important and obvious benefit of scaling down in size is
the ability of the quadrotor to operate in tightly constrained
environments in tight formations. While the payload capacity
of the quadrotor falls dramatically, it is possible to deploy
multiple quadrotors that cooperate to overcome this limitation.
Again, the small size benefits us because smaller vehicles
can operate in closer proximity than large vehicles. Another
interesting benefit of scaling down is agility. As argued later
and illustrated with experimental results, smaller quadrotors
exhibit higher accelerations allowing more rapid adaptation to
disturbances and higher stability.
Fig. 1. A formation of 20 micro quadrotors in flight. See video at
http://youtu.be/50Fdi7712KQ
II. AGILITY OF MICRO QUADROTORS
It is useful to develop a simple physics model to analyze a
quadrotor’s ability to produce linear and angular accelerations
from a hover state. If the characteristic length is L, the rotor
radius R scales linearly with L. The mass scales as L
3
and the
moments of inertia as L
5
. On the other hand the lift or thrust,
F , and drag, D, from the rotors scale with the cross-sectional
area and the square of the blade-tip velocity, v. If the angular
speed of the blades is defined by ω =
v
L
, F ∼ ω
2
L
4
and D ∼
ω
2
L
4
. The linear acceleration a scales as a ∼
ω
2
L
4
L
3
= ω
2
L.
Thrusts from the rotors produce a moment with a moment arm
L. Thus the angular acceleration α ∼
ω
2
L
5
L
5
= ω
2
.
The rotor speed, ω also scales with length since smaller
motors produce less torque which limits their peak speed
because of the drag resistance that also scales the same
way as lift. There are two commonly accepted approaches
to study scaling in aerial vehicles [28]. Mach scaling is used
for compressible flows and essentially assumes that the tip
velocities are constant leading to ω ∼
1
R
. Froude scaling is
used for incompressible flows and assumes that for similar
aircraft configurations, the Froude number,
v
2
Lg
, is constant.
Here g is the acceleration due to gravity. This yields ω ∼
1
√
R
.
However, neither Froude or Mach number similitudes take
motor characteristics nor battery properties into account. While
motor torque increases with length, the operating speed for the
rotors is determined by matching the torque-speed character-
istics of the motor to the drag versus speed characteristics
of the propellors. Further, the motor torque depends on the
ability of the battery to source the required current. All these
variables are tightly coupled for smaller designs since there
are fewer choices available at smaller length scales. Finally,
the assumption that propeller blades are rigid may be wrong
and the performance of the blades can be very different at
smaller scales, the quadratic scaling of the lift with speed may
Fig. 2. A prototype micro quadrotor.
not be accurate. Nevertheless these two cases are meaningful
since they provide some insight into the physics underlying
the maneuverability of the craft.
Froude scaling suggests that the acceleration is independent
of length while the angular acceleration α ∼ L
−1
. On the
other hand, Mach scaling leads to the conclusion that a ∼ L
while α ∼ L
−2
. Since quadrotors must rotate (exhibit angular
accelerations) in order to translate, smaller quadrotors are
much more agile.
There are two design points that are illustrative of the
quadrotor configuration. The Pelican quadrotor from Ascend-
ing Technologies [2] equipped with sensors (approx. 2 kg gross
weight, 0.75 m diameter, and 5400 rpm nominal rotor speed
at hover), consumes approximately 400 W of power [25]. The
Hummingbird quadrotor from Ascending Technologies (500
grams gross weight, approximately 0.5 m diameter, and 5000
rpm nominal rotor speed at hover) without additional sensors
consumes about 75 W. In this paper, we outline a design for
a quadrotor which is approximately 40% of the size of the
Hummingbird, 15% of its mass, and consuming approximately
20% of the power for hovering.
III. THE MICRO QUADROTOR
A. The Vehicle
The prototype quadrotor is shown in Figure 2. Its booms
are made of carbon fiber rods which are sandwiched between
a custom motor controller board on the bottom and the main
controller board on the top. To produce lift the vehicle uses
four fixed-pitch propellers with diameters of 8 cm. The vehicle
propeller-tip-to-propeller-tip distance is 21 cm and its weight
without a battery is 50 grams. The hover time is approximately
11 minutes with a 2-cell 400 mAh Li-Po battery that weighs
23 grams.
B. Electronics
Despite its small size this vehicle contains a full suite of
onboard sensors. An ARM Cortex-M3 processor, running at
72 MHz, serves as the main processor. The vehicle contains
a 3-axis magnetometer, a 3-axis accelerometer, a 2-axis 2000
deg/sec rate gyro for the roll and pitch axes, and a single-
axis 500 deg/sec rate gyro for the yaw axis. The vehicle also
contains a barometer which can be used to sense a change in
−0.1 0 0.1 0.2 0.3
0
10
20
30
40
50
Pitch Angle (deg)
Time (sec)
(a) Pitch angle step input re-
sponse
(b) Data for the flipping maneuver
Fig. 3. Attitude controller performance data
altitude. For communication the vehicle contains two Zigbee
transceivers that can operate at either 900 MHz or 2.4 GHz.
C. Software Infrastructure
The work in this paper uses a Vicon motion capture system
[5] to sense the position of each vehicle at 100 Hz. This
data is streamed over a gigabit ethernet network to a desktop
base station. High-level control and planning is done in
MATLAB on the base station which sends commands to each
quadrotor at 100 Hz. The software for controlling a large
team of quadrotors is described later in Sec. V (see Fig. 7).
Low-level estimation and control loops run on the onboard
microprocessor at a rate of 600 Hz.
Each quadrotor has two independent radio transceivers,
operating at 900 MHz and 2.4 GHz. The base station sends,
via custom radio modules, the desired commands, containing
orientation, thrust, angular rates and attitude controller gains
to the individual quadrotors. The onboard rate gyros and
accelerometer are used to estimate the orientation and angular
velocity of the craft. The main microprocessor runs the attitude
controller described in Sec. IV and sends the desired propeller
speeds to each of the four motor controllers at full rate
(600Hz).
D. Performance
Some performance data for the onboard attitude controller
in Fig. 3. The small moments of inertia of the vehicle enable
the vehicle to create large angular accelerations. As shown in
Fig. 4(a) the attiude control is designed to be approximately
critically damped with a settling time of less than 0.2 seconds.
Note that this is twice as fast as the settling time for the attitude
controller for the AscTec Hummingbird reported in [18]. Data
for a flip is presented 3(b). Here the vehicle completes a
complete flip about its y axis in about 0.4 seconds and reaches
a maximum angular velocity of 1850 deg/sec.
The position controller described in Sec. IV uses the roll
and pitch angles to control the x and y position of the vehicle.
For this reason, a stiff attitude controller is a required for stiff
position control. Response to step inputs in the lateral and
vertical directions are shown in Fig. 4(b). For the hovering
performance data shown in Fig. 4 the standard deviations of
the error for x and y are about 0.75 cm and about 0.2 cm for
z.
(a) Position error (b) Position step input re-
sponse
Fig. 4. The red, green, and blue lines in (a) represent the x, y, and z errors
while hovering. (b) shows the step response for the position controller in x
(top) and z (bottom).
Fig. 5. The reference frames and propeller numbering convention.
IV. DYNAMICS AND CONTROL
The dynamic model and control for the micro quadrotor
is based on the approach in [17]. As shown in Figure 5, we
consider a body-fixed frame B aligned with the principal axes
of the quadrotor (unit vectors b
i
) and an inertial frame A with
unit vectors a
i
. B is described in A by a position vector r to
the center of mass C and a rotation matrix R. In order to avoid
singularities associated with parameterization, we use the full
rotation matrix to describe orientations. The angular velocity
of the quadrotor in the body frame, ω, is given by ˆω = R
T
˙
R,
where ˆ denotes the skew-symmetric matrix form of the vector.
As shown in Fig. 5, the four rotors are numbered 1-4, with
odd numbered rotors having a pitch that is opposite to the
even numbered rotors. The angular speed of the rotor is ω
i
.
The resulting lift, F
i
, and the reaction moment, M
i
, are given
by:
F
i
= k
F
ω
2
i
, M
i
= k
M
ω
2
i
.
where the constants k
F
and k
M
are empirically determined.
For our micro quadrotor, the motor dynamics have a time
constant less than 10 msec and are much faster than the
time scale of rigid body dynamics and aerodynamics. Thus
we neglect the dynamics and assume F
i
and M
i
can be
instantaneously changed. Therefore the control input to the
system, u, consists of the net thrust in the b
3
direction,
u
1
= Σ
4
i=1
F
i
, and the moments in B, [u
2
, u
3
, u
4
]
T
, given
by:
u =
k
F
k
F
k
F
k
F
0 k
F
L 0 −k
F
L
−k
F
L 0 k
F
L 0
k
M
−k
M
k
M
−k
M
ω
2
1
ω
2
2
ω
2
3
ω
2
4
, (1)
where L is the distance from the axis of rotation of the
propellers to the center of the quadrotor.
The Newton-Euler equations of motion are given by:
m
¨
r = −mga
3
+ u
1
b
3
(2)
˙ω = I
−1
−ω × Iω +
u
2
u
3
u
4
(3)
where I is the moment of inertia matrix along b
i
.
We specify the desired trajectory using a time-parameterized
position vector and yaw angle. Given a trajectory, σ(t) :
[0, t
f
] → R
3
× SO(2), the controller derives the input u
1
based on position and velocity errors:
u
1
= (−K
p
e
p
− K
v
e
v
+ mga
3
) · b
3
(4)
where e
p
= r − r
T
and e
v
=
˙
r −
˙
r
T
. The other three inputs
are determined by computing the desired rotation matrix. We
want to align the thrust vector u
1
b
3
with (−K
p
e
p
− K
v
e
v
+
mga
3
) in (4). Second, we want the yaw angle to follow the
specified yaw ψ
T
(t). From these two pieces of information
we can compute R
des
and the error in rotation according to:
e
R
=
1
2
(R
T
des
R − R
T
R
des
)
∨
where
∨
represents the vee map which takes elements of
so(3) to R
3
. The desired angular velocity is computed by
differentiating the expression for R and the desired moments
can be expressed as a function of the orientation error, e
R
,
and the angular velocity error, e
ω
:
[u
2
, u
3
, u
4
]
T
= −K
R
e
R
− K
ω
e
ω
, (5)
where K
R
and K
ω
are diagonal gain matrices. Finally we
compute the desired rotor speeds to achieve the desired u by
inverting (1).
V. CONTROL AND PLANNING FOR GROUPS
A. Architecture
We are primarily interested in the challenge of coordinating
a large team of quadrotors. To manage the complexity that
results from growth of the state space dimensionality and
limit the combinatorial explosion arising from interactions
between labeled vehicles, we consider a team architecture
in which the team is organized into labeled groups, each
with labeled vehicles. Formally, we can define a group of
agents as a collection of agents which work simultaneously to
complete a single task. Two or more groups act in a team to
complete a task which requires completing multiple parallel
subtasks [7]. We assume that vehicles within a group can
communicate at high data rates with low latencies while the
communication requirements for coordination across groups
are much less stringent. Most importantly, vehicles within a
group are labeled. The small group size allows us to design
controllers and planners that provide global guarantees on
!"#$%&'&
!"#$%&(&
!"#$%&!"
Fig. 6. The team of quadrotors is organized into m groups. While vehicles
within the group are tightly coordinated and centralized control and planning
is possible, the inter-group coordination need not be centralized.
shapes, communication topology, and relative positions of
individual, agile robots.
Our approach is in contrast to truly decentralized approaches
which are necessary in swarms with hundreds and thousands
of agents [21]. While models of leaderless aggregation and
swarming with aerial robots are discussed in the robotics
community [11, 26, 19], here the challenge of enumerating
labelled interactions between robots is circumvented by con-
trolling such aggregate descriptors of formation as statistical
distributions. These methods cannot provide guarantees on
shape or topology. Reciprocal collision avoidance algorithms
[27] have the potential to navigate robots to goal destinations
but no guarantees are available for transient performance and
no proof of convergence is available.
On the other hand, the problem of designing decentralized
controllers for trajectory tracking for three dimensional rigid
structures is now fairly well understood[12, 13, 8], although
few experimental results are available for aerial robots. Our
framework allows the maintenance of such rigid structures in
groups.
B. Formation Flight
Flying in formation reduces the complexity of generating
trajectories for a large team of vehicles to generating a trajec-
tory for a single entity. If the controllers are well-designed,
there is no need to explicitly incorporate collision avoidance
between vehicles. The position error for quadrotor q at time t
can be written as
e
pq
(t) = e
f
(t) + e
lq
(t) (6)
where e
f
(t) is the formation error describing the error of
position of the group from the prescribed trajectory, and e
lq
(t)
is the local error of quadrotor q within the formation of the
group. As we will show in Sec. VI the local error is typically
quite small even for aggressive trajectories even though the
formation error can be quite large.
A major disadvantage of formation flight is that the rigid
formation can only fit through large gaps. This can be ad-
dressed by changing the shape of the formation of the team
or dividing the team into smaller groups, allowing each group
to negotiate the gap independently.
C. Time-Separated Trajectory Following
Another way to reduce the complexity of the trajectory
generation problem is to require all vehicles to follow the same
team trajectory but be separated by some time increment. Here
we let the trajectory for quadrotor q be defined as
r
T q
(t) = r
T T
(t + ∆t
q
) (7)
where r
T T
is the team trajectory and ∆t
q
is the time shift for
quadrotor q from some common clock, t. If the team trajectory
does not intersect or come within an unsafe distance of itself
then vehicles simply need to follow each other at a safe
time separation. Large numbers of vehicles can follow team
trajectories that intersect themselves if the time separations,
∆t
q
, are chosen so that no two vehicles are at any of the
intersection points at the same time. An experiment for an
intersecting team trajectory is shown in Sec. VI.
D. Trajectory Generation with MIQPs
Here we describe a method for generating smooth, safe
trajectories through known 3-D environments satisfying spec-
ifications on intermediate waypoints for multiple vehicles. In-
teger constraints are used to enforce collision constraints with
obstacles and other vehicles and also to optimally assign goal
positions. This method draws from the extensive literature on
mixed-integer linear programs (MILPs) and their application
to trajectory planning from Schouwenaars et al. [23, 24].
1) Basic Method: As described in [17] an optimization
program can be used to generate trajectories that smoothly
transition through n
w
desired waypoints at specified times,
t
w
. The optimization program to solve this problem while
minimizing the integral of the k
r
th derivative of position
squared for n
q
quadrotors is shown below.
min
P
n
q
q =1
R
t
n
w
t
0
d
k
r
r
Tq
dt
k
r
2
dt (8)
s.t. r
T q
(t
w
) = r
wq
, w = 0, ..., n
w
; ∀q
d
j
x
T q
dt
j
|
t=t
w
= 0 or free, w = 0, n
w
; j = 1, ..., k
r
; ∀q
d
j
y
T q
dt
j
|
t=t
w
= 0 or free, w = 0, n
w
; j = 1, ..., k
r
; ∀q
d
j
z
T q
dt
j
|
t=t
w
= 0 or free, w = 0, n
w
; j = 1, ..., k
r
; ∀q
Here r
T q
= [x
T q
, y
T q
, z
T q
] represents the trajectory for
quadrotor q and r
wq
represents the desired waypoints for
quadrotor q. We enforce continuity of the first k
r
derivatives
of r
T q
at t
1
,...,t
n
w
−1
. As shown in [17] writing the trajectories
as piecewise polynomial functions allows [8] to be written as
a quadratic program (or QP) in which the decision variables
are the coefficients of the polynomials.
For quadrotors, since the inputs u
2
and u
3
appear as
functions of the fourth derivatives of the positions, we generate
trajectories that minimize the integral of the square of the
norm of the snap (the second derivative of acceleration,
k
r
= 4). Large order polynomials are used to satisfy such
additional trajectory constraints as obstacle avoidance that are
not explicitly specified by intermediate waypoints.
2) Integer Constraints for Collision Avoidance: For col-
lision avoidance we model the quadrotors as a rectangular
prisms oriented with the world frame with side lengths l
x
, l
y
,
and l
z
. These lengths are large enough so that the quadrotor
can roll, pitch, and yaw to any angle and stay within the prism.
We consider navigating this prism through an environment
with n
o
convex obstacles. Each convex obstacle o can be
represented by a convex region in configuration space with
n
f
(o) faces. For each face f the condition that the quadrotor’s
desired position at time t
k
, r
T q
(t
k
), be outside of obstacle o
can be written as
n
of
· r
T q
(t
k
) ≤ s
of
, (9)
where n
of
is the normal vector to face f of obstacle o in
configuration space and s
of
is a scalar that determines the
location of the plane. If (9) is satisfied for at least one of the
faces then the rectangular prism, and hence the quadrotor, is
not in collision with the obstacle. The condition that quadrotor
q does not collide with an obstacle o at time t
k
can be enforced
with binary variables, b
q of k
, as
n
of
· r
T q
(t
k
) ≤ s
of
+ M b
q of k
∀f = 1, ..., n
f
(o) (10)
b
q of k
= 0 or 1 ∀f = 1, ..., n
f
(o)
n
f
(o)
X
f=1
b
q of k
≤ n
f
(o) − 1
where M is a large positive number [23]. Note that if b
q of k
is 1 then the inequality for face f is always satisfied. The last
inequality in (10) requires that the non-collision constraint be
satisfied for at least one face of the obstacle which implies
that the prism does not collide with the obstacle. We can
then introduce (10) into (8) for all n
q
quadrotors for all n
o
obstacles at n
k
intermediate time steps between waypoints.
The addition of the integer variables into the quadratic program
causes this optimization problem to become a mixed-integer
quadratic program (MIQP).
3) Inter-Quadrotor Collision Avoidance: When transition-
ing between waypoints quadrotors must stay a safe distance
away from each other. We enforce this constraint at n
k
intermediate time steps between waypoints which can be
represented mathematically for quadrotors 1 and 2 by the
following set of constraints:
∀t
k
: x
T 1
(t
k
) − x
T 2
(t
k
) ≤ d
x12
(11)
or x
T 2
(t
k
) − x
T 1
(t
k
) ≤ d
x21
or y
T 1
(t
k
) − y
T 2
(t
k
) ≤ d
y 12
or y
T 2
(t
k
) − y
T 1
(t
k
) ≤ d
y 21
Here the d terms represent safety distances. For axially sym-
metric vehicles d
x12
= d
x21
= d
y 12
= d
y 21
. Experimentally
we have found that quadrotors must avoid flying in each
other’s downwash because of a decrease in tracking perfor-
mance and even instability in the worst cases. Therefore we
do not allow vehicles to fly underneath each other here. Finally,
we incorporate constraints (11) between all n
q
quadrotors in
the same manner as in (10) into (8).
4) Integer Constraints for Optimal Goal Assignment: In
many cases one might not care that a certain quadrotor goes
to a certain goal but rather that any vehicle does. Here we
describe a method for using integer constraints to find the
optimal goal assignments for the vehicles. This results in a
lower total cost compared to fixed-goal assignment and often
a faster planning time because there are more degrees of
freedom in the optimization problem. For each quadrotor q
and goal g we introduce the integer constraints:
x
T q
(t
n
w
) ≤ x
g
+ M β
q g
(12)
x
T q
(t
n
w
) ≥ x
g
− M β
q g
y
T q
(t
n
w
) ≤ y
g
+ M β
q g
y
T q
(t
n
w
) ≥ y
g
− M β
q g
z
T q
(t
n
w
) ≤ z
g
+ M β
q g
z
T q
(t
n
w
) ≥ z
g
− M β
q g
Here β
q g
is a binary variable used to enforce the optimal goal
assignment. If β
q g
is 0 then quadrotor q must be at goal g at
t
n
w
. If β
q g
is 1 then these constraints are satisfied for any final
position of quadrotor q. In order to guaruntee that at least n
g
quadrotors reach the desired goals we introduce the following
constraint.
n
q
X
q =1
n
g
X
g =1
β
q g
≤ n
g
n
q
− n
g
(13)
Note that this approach can be easily adapted if there are more
quadrotors than goals or vice versa.
5) Relaxations for Large Teams: The solving time of the
MIQP grows exponentially with the number of binary vari-
ables that are introduced into the MIQP. Therefore, the direct
use of this method does not scale well for large teams. Here
we present two relaxations that enable this approach to be used
for large teams of vehicles.
a) Planning for Groups within a Team: A large team
of vehicles can be divided into smaller groups. We can then
use the MIQP method to generate trajectories to transition
groups of vehicles to group goal locations. This reduces
the complexity of the MIQP because instead of planning
trajectories for all n
q
vehicles we simply plan trajectories
for the groups. Of course we are making a sacrifice here by
not allowing the quadrotors to have the flexibility to move
independently.
b) Planning for Sub-Regions: In many cases the en-
vironment can be partitioned into n
r
convex sub-regions
where each sub-region contains the same number of quadrotor
start and goal positions. After partitioning the environment
the MIQP trajectory generation method can be used for the
vehicles inside each region. Here we require quadrotors to
stay inside their own regions using linear constraints on the
positions of the vehicles. This approach guarantees collision-
free trajectories and allows quadrotors the flexibility to move
independently. We are gaining tractability at the expense
of optimality since the true optimal solution might actually
require quadrotors to cross region boundaries while this re-
laxed version does not. Also, it is possible that no feasible
trajectories exist inside a sub-region but feasible trajectories
do exist which cross region boundaries. Nonetheless, this
approach works well in many scenarios and we show its
application to formation transitions for teams of 16 vehicles
in Sec. VI.