Adaptive Incremental Nonlinear Dynamic Inversion
for Attitude Control of Micro Aerial Vehicles
Ewoud J.J. Smeur1and Qiping Chu2and Guido C.H.E. de Croon3
Delft University of Technology, Delft, Zuid-Holland, 2629HS, Netherlands
Incremental Nonlinear Dynamic Inversion (INDI) is a sensor-based control approach
that promises to provide high performance nonlinear control without requiring a de-
tailed model of the controlled vehicle. In the context of attitude control of Micro Air
Vehicles, INDI only uses a control eﬀectiveness model and uses estimates of the an-
gular accelerations to replace the rest of the model. This paper provides solutions for
two major challenges of INDI control: how to deal with measurement and actuator
delays and how to deal with a changing control eﬀectiveness. The main contributions
of this article are: (1) a proposed method to correctly take into account the delays
occurring when deriving angular accelerations from angular rate measurements, (2)
the introduction of adaptive INDI, which can estimate the control eﬀectiveness online,
eliminating the need for manual parameter estimation or tuning and (3) the incorpo-
ration of the momentum of the propellers in the controller. This controller is suitable
for vehicles that experience a diﬀerent control eﬀectiveness across their ﬂight envelope.
Furthermore, this approach requires only very course knowledge of model parameters
in advance. Real-world experiments show the high performance, disturbance rejection
and adaptiveness properties.
1PhD Candidate, Delft University of Technology, Control and Simulation
2Associate Professor, Control and Simulation, member.
3Assistant Professor, Control and Simulation.
b= Width of the vehicle, m
Iv= Moment of inertia matrix of the vehicle, kg m2
Ir= Moment of inertia matrix of the rotor, kg m2
I= Identity matrix
i= Rotor index
k1= Force constant of the rotors, kg m/rad
k2= Moment constant of the rotors, kg m2/rad
l= Length of the vehicle, m
Ma= Aerodynamic moment vector acting on the vehicle, Nm
Mc= Control moment vector acting on the vehicle, Nm
Mr= Moment vector acting on the propeller, Nm
Ts= Sample time of the controller, s
u= Actuator input vector, rad/s
v= Vehicle velocity vector, m/s
µ= Adaptation rate diagonal matrix
Ω= Vehicle angular rate vector, rad/s
Ω= Angular acceleration vector, rad/s2
ω= Angular rate vector of the four rotors around the body zaxis, rad/s
ωi= Angular rate vector of rotor iaround each of the body axes, rad/s
Micro Aerial Vehicles (MAVs) have increased in popularity as low-cost lightweight processors
and inertial measurement units (IMUs) have become available through the smartphone revolution.
The inertial sensors allow stabilization of unstable platforms by feedback algorithms. Typically,
the stabilization algorithm used for MAVs is simple Proportional Integral Derivative (PID) control
[1, 2]. Problems with PID control occur when the vehicle is highly nonlinear or when the vehicle is
subject to large disturbances like wind gusts.
Alternatively, we could opt for a model based attitude controller. A model based controller that
can deal with nonlinear systems is nonlinear dynamic inversion (NDI), which involves modeling all
of the MAV’s forces and dynamics. Theoretically, this method can remove all nonlinearities from the
system and create a linearizing control law. However, NDI is very sensitive to model inaccuracies
. Obtaining an accurate model is often expensive or impossible with the constraints of the sensors
that are carried onboard a small MAV.
The incremental form of NDI, Incremental NDI or INDI, is less model dependent and more
robust. It has been described in the literature since the late nineties [4, 5], sometimes referred to as
simpliﬁed  or enhanced  NDI. Compared to NDI, instead of modeling the angular acceleration
based on the state and inverting the actuator model to get the control input, the angular acceleration
is measured and an increment of the control input is calculated based on a desired increment in
angular acceleration. This way, any unmodeled dynamics, including wind gust disturbances, are
measured and compensated. Since INDI makes use of a sensor measurement to replace a large part
of the model, it is considered a sensor based approach.
INDI faces two major challenges. Firstly, the measurement of angular acceleration is often
noisy and requires ﬁltering. This ﬁltering introduces a delay in the measurement, which should be
compensated for. Secondly, the method relies on inverting and therefore modeling the controls. To
achieve a more ﬂexible controller, the control eﬀectiveness should be determined adaptively.
Delay in the angular acceleration measurement has been a prime topic in INDI research.
A proposed method to deal with these measurement delays is predictive ﬁltering . However, the
prediction of angular acceleration requires additional modeling. Moreover, disturbances cannot be
predicted. Initially, a setup with multiple accelerometers was proposed by Ostroﬀ and Bacon  to
measure the angular acceleration. This setup has some drawbacks, because it is complex and the
accelerometers are sensitive to structural vibrations. Later, they discussed the derivation of angular
acceleration from gyroscope measurements by using a second order ﬁlter . To compensate for the
delay introduced by the ﬁlter, Ostroﬀ and Bacon use a lag ﬁlter on the applied input to the system.
We show in this paper that perfect synchronization of input and measured output can be achieved
by applying the ﬁlter used for the gyroscope diﬀerentiation on the incremented input as well.
Other research focused on compensating delays in the inputs by using a Lyapunov based con-
troller design . In this paper, we show that delayed inputs (actuator dynamics) are naturally
handled by the INDI controller.
The control eﬀectiveness is the sole model still required by INDI. The parameters can be
obtained by careful modeling of the actuators and the moment of inertia, or by analyzing the
input output data from ﬂight logs. However, even if such a tedious process is followed, the control
eﬀectiveness can change during ﬂight. For instance, this can occur due to changes in ﬂight conditions
 or actuator damage . In order to cope with this, we propose a method to adaptively
determine the control eﬀectiveness matrices.
In this paper, we present three main contributions: (1) a mathematically sound way of dealing
with the delays originating from ﬁltering of the gyroscope measurements, (2) the introduction of
an adaptive INDI scheme, which can estimate the control eﬀectiveness online and (3) incorporation
of propeller momentum in the controller design. These contributions are implemented and demon-
strated on a Parrot Bebop quadrotor running the Paparazzi open source autopilot software. This
is a commercially available quadrotor and the code is publicly available on Github.
The presented theory and results generalize to other vehicles in a straightforward manner. We
have applied this control approach successfully to a variety of quadrotors. Some of these MAVs
were able to measure the rotational rate of the rotors (actuator feedback), but some did not have
this ability. The INDI controller is believed to scale well to diﬀerent types of MAVs like helicopter,
multirotor, ﬁxedwing or hybrid.
The outline of this paper is as follows. First, a model of the MAV will be discussed in Section
II. Second, Section III will deal with INDI and the analysis for this controller for a quadrotor.
Section IV is about the adaptive extension of INDI. Finally, in Section V, the experimental setup
is explained, followed by the results of the experiments in Section VI.
II. MAV Model
The Bebop quadrotor is shown in Figure 1 along with axis deﬁnitions. The actuators drive
the four rotors, whose angular velocity in the body frame is given by ωi= [ωix, ωiy, ωiz], where i
denotes the rotor number. The center of gravity is located in the origin of the axis system and the
distance to each of the rotors along the Xaxis is given by land along the Yaxis by b.
Fig. 1 The Bebop Quadcopter used in the experiments with axis deﬁnitions.
If the angular velocity vector of the vehicle is denoted by Ω= [p, q, r]Tand its derivative by ˙
the rotational dynamics are given by Euler’s equation of motion , more speciﬁcally the one that
describes rotation. If we consider the body axis system as our coordinate system we get Eq. (1) for
the angular velocity of the vehicle.
Where Mis the moment vector acting on the vehicle. If we consider the rotating propellers, still
in the body coordinate system, we obtain:
Where ωiis the angular rate vector of the ith propeller in the vehicle body axes and Ωthe angular
rotation of the coordinate system, equal to the vehicle body rates. The rotors are assumed to be ﬂat
in the z axis, such that the inertia matrix Irhas elements that are zero: Irxz =Iry z = 0 . Because
the coordinate system is ﬁxed to the vehicle, Irxx ,Irxy and Iryy are not constant in time. However,
as is shown later on, the terms containing these moments of inertia will disappear. Expanding Eq.
(2) into its three components gives:
Irxx ˙ωix−Iry y Ωzωiy−Irxy Ωzωix+Irzz Ωyωiz=Mrix
Iryy ˙ωiy+Irxx Ωzωix+Irxy Ωzωiy−Irzz Ωxωiz=Mriy
Irzz ˙ωiz−Irxx Ωyωix−Irxy Ωyωiy+Iryy Ωxωiy+Irxy Ωxωix=Mriz
The propellers are light-weight and have a small moment of inertia compared to the vehicle. Relevant
precession terms are therefore those that contain the relatively large ωiz. Since the rotors spin around
the zaxis, it is safe to assume that ωix≪ωizand ωiy≪ωizand that ˙ωixand ˙ωiyare negligible.
Then, the moments exerted on the rotors due to their rotational dynamics are given by Eq. (4).
Note the presence of the term Irzz ˙ωiz, which is the moment necessary to change the angular velocity
of a rotor. In Section VI, it will be shown that this term is important.
This equation holds for each of the four rotors, so the moment acting on a rotor is given a
subscript ito indicate the rotor number. The total moment due to the rotational eﬀects of the
rotors is shown in Eq. (5). Since motors 1 and 3 spin in the opposite direction of rotors 2 and 4, a
factor (−1)iis introduced. As we are left with only the zcomponent for the angular velocity of each
rotor, we will omit this subscript and continue with the vector ω= [ω1z, ..., ω4z]T= [ω1, ..., ω4]T.
0 0 0 0
0 0 0 0
Irzz −Irzz Irzz −Irzz
Irzz Ωy−Irzz ΩyIrzz Ωy−Irzz Ωy
−Irzz ΩxIrzz Ωx−Irzz ΩxIrzz Ωx
Now consider the Euler Equation, Eq. (1), for the entire vehicle. The moments from the rotor
dynamics are subtracted from the other moments yielding:
Ω+Ω×IvΩ=Mc(ω) + Ma(Ω,v)−Mr(ω,˙ω,Ω)(6)
Here, Ivis the moment of inertia matrix of the vehicle, Mr(ω,˙ω,Ω)is the gyroscopic eﬀect of the
rotors, Mc(ω)is the control moment vector generated by the rotors and Ma(Ω,v)is the moment
vector generated by aerodynamic eﬀects, which depends on the angular rates and the MAV velocity
vector v. The control moment Mc(ω)is elaborated in Eq. (7), where k1is the force constant of
the rotors, k2is the moment constant of the rotors and band lare deﬁned in Figure 1.
If we now take Eq. (6), insert Eqs. (4) and (7) and solve for the angular acceleration ˙
Ω, we arrive
at the following
v(Ma(Ω,v)−Ω×IvΩ) + I−1
=F(Ω,v) + 1
where F(Ω,v) = I−1
v(Ma(Ω,v)−Ω×IvΩ)are the forces independent of the actuators and G1,
G2,G3and C(Ω)are given by Eqs. (9), (10), (11) and (12) respectively. Note that the sample
time Tsof the quadrotor is introduced to ease future calculations.
0 0 0 0
0 0 0 0
Irzz −Irzz Irzz −Irzz
Irzz −Irzz Irzz −Irzz
−Irzz Irzz −Irzz Irzz
0 0 0 0
0 0 0
Note that traditionally in the literature, the system solved by INDI has the form of ˙x=f(x) +
g(x, u)where xis the state of the system and uthe input to the system. However, as becomes clear
from Eq. (8), the quadrotor is actually a system of the form ˙x=f(x) + g(x, u, ˙u). In Section III, a
solution to this type of problem will be shown.
III. Incremental Nonlinear Dynamic Inversion
Consider Eq. (8) from the previous section. This equation has some extra terms compared to
previous work , because the gyroscopic and angular momentum eﬀects of the rotors are included.
We can apply a Taylor expansion to Eq. (8) and if we neglect higher order terms this results in Eq.
Ω=F(Ω0,v0) + 1
∂Ω(F(Ω,v0) + C(Ω)G3ω0)|Ω=Ω0(Ω−Ω0)
This equation predicts the angular acceleration after an inﬁnitesimal timestep ahead in time based
on a change in angular rates of the vehicle and a change in rotational rate of the rotors. Now
observe that the ﬁrst terms give the angular acceleration based on the current rates and inputs:
F(Ω0,v0) + 1
Ω0. This angular acceleration can be obtained by
deriving it from the angular rates, which are measured with the gyroscope. In other words, these
terms are replaced by a sensor measurement, which is why INDI is also referred to as sensor based
The second and third term, partial to Ωand v, are assumed to be much smaller than the fourth
and ﬁfth term, partial to ωand ˙ω. This is commonly referred to as the principle of time scale
separation . This assumption only holds when the actuators are suﬃciently fast and have more
eﬀect compared to the change in aerodynamic and precession moments due to changes in angular
rates and body speeds. These assumptions and calculation of the partial derivatives gives Eq. (14):
Ω0+G1diag(ω0)(ω−ω0) + TsG2(˙ω−˙ω0)−C(Ω0)G3(ω−ω0)(14)
Above it is stated that the angular acceleration is measured by deriving it from the angular rates.
In most cases, the gyroscope measurements from a MAV are noisy due to vibrations of the vehicle
due to the propellers and motors. Since diﬀerentiation of a noisy signal ampliﬁes the noise, some
ﬁltering is required. The use of a second order ﬁlter is adopted from the literature , of which a
transfer function in the Laplace domain is given by Eq. (15). Satisfactory results were obtained with
ωn= 50 rad/s and ζ= 0.55. Other low pass ﬁlters are also possible, for instance the Butterworth
H(s) = ω2
The result is that instead of the current angular acceleration, a ﬁltered and therefore delayed angular
Ωfis measured. Since all the terms with the zero subscript in the Taylor expansion
should be at the same point in time, they are all replaced with the subscript f, yielding Eq. (16).
This indicates that these signals are also ﬁltered and are therefore synchronous with the angular
Ωf+G1diag(ωf)(ω−ωf) + TsG2(˙ω−˙ωf)−C(Ωf)G3(ω−ωf)(16)
This equation is not yet ready to be inverted, because it contains the derivative of the angular rate
of the propellers. Since we are dealing with discrete signals, consider the discrete approximation of
the derivative in the zdomain: ˙ω= (ω−ωz−1)T−1
s, where Tsis the sample time. This is shown in
Ωf+G1diag(ωf)(ω−ωf) + G2(ω−ωz−1−ωf+ωfz−1)−C(Ωf)G3(ω−ωf)(17)
Collecting all terms with (ω−ωf)yields Eq. (18):
Ωf+ (G1diag(ωf) + G2−C(Ωf)G3)(ω−ωf)−G2z−1(ω−ωf)(18)
Inversion of this equation for ωyields Eq. (19), where +denotes the Moore-Penrose pseudoinverse:
ωc=ωf+ (G1diag(ωf) + G2−C(Ωf)G3)+(ν−˙
Note that the predicted angular acceleration ˙
Ωis now instead a virtual control, denoted by ν. The
virtual control is the desired angular acceleration, and with Eq. (19), the required inputs ωccan be
calculated. The subscript cis added to ωto indicate that this is the command sent to the motors.
This input is given with respect to a previous input ωf. If we deﬁne the increment in the motor
commands as e
ω=ωc−ωf, it is clearly an incremental control law.
A. Parameter Estimation
Equation (19) shows the general quadrotor INDI control law. The parameters of this equation
are the three matrices G1,G2and G3which need to be identiﬁed for the speciﬁc quadrotor. This
can be done through measurement of each of the components that make up these matrices, including
the moments of inertia of the vehicle and the propellers as well as the thrust and drag coeﬃcients
of the rotors. Identifying the parameters in this way requires a signiﬁcant amount of eﬀort.
A more eﬀective method is to use test ﬂight data to determine the model coeﬃcients. Of course,
to do this the MAV needs to be ﬂying. This can be achieved by initially tuning the parameters.
Alternatively, a diﬀerent controller can be used at ﬁrst to gather the test ﬂight data, such as PID
control. Once a test ﬂight has been logged, Eq. (18) is used for parameter estimation and is written
as Eq. (20). From this equation, a least squares solution is found for the matrices G1,G2and G3.
Here, ∆denotes the ﬁnite diﬀerence between two subsequent samples. From the data, we can also
investigate the importance of some of the terms by comparing the least squares error with and
without the terms. It turns out that on a typical dataset, leaving out the matrix G3only results
in an estimation squared error increase of ∼0.2%. Furthermore, modeling the rotor as linear with
the rotational speed of the rotor instead of quadratic gives an estimation squared error increase of
∼0.9%. Therefore, we can simplify the INDI control law of Eq. (19) to Eq. (21):
With the simpliﬁcations described in subsection III A, the ﬁnal INDI control scheme is shown
in Figure 2. The input to the system is the virtual control νand the output is the angular acceler-
ation of the system ˙
Ω. The angular velocity measurement from the gyroscope is fed back through
the diﬀerentiating second order ﬁlter and subtracted from the virtual control to give the angular
acceleration error ˙
Since the matrices G1and G2are not square, we take the pseudo inverse to solve the problem of
control allocation, denoted by +. The contents of the block ’MAV’ are shown in Figure 3, because
it allows the closed loop analysis in Section I II C. In this diagram, dis a disturbance term that
bundles disturbances and unmodeled dynamics.
Fig. 2 INDI control scheme. A(z)denotes the actuator dynamics and H(z)is the second order
Fig. 3 The contents of the block named ’MAV’ in Figure 2.
Note that Eq. (21) provides a desired angular velocity of the rotors. However, the actuators
do not have an instantaneous response. Instead, it is assumed they have ﬁrst order dynamics A(z).
The reference sent to the motors is denoted by ωcand e
ω=ωc−ωf. In Figure 2, it is assumed
that actuator feedback is available. However, if this is not the case, the actuator state ω0has to be
estimated with a model of the actuator dynamics as is shown in Figure 4. Here A′(z)is a model of
the actuator dynamics.
Fig. 4 Block diagram for estimation of actuator state if actuator feedback is not available.
C. Closed Loop Analysis
Consider the control diagram shown in Figure 2. We can verify that this is a stable controller by
doing a closed loop analysis. First, the transfer function of each of the two small loops is calculated,
shown by Eq. (22) and (23). Here TFx→ydenotes the transfer function from point x to y in the
Ωerr + (G1+G2)+G2z−1e
ω(z) = (G1+G2−G2z−1)+
We deﬁne H(z) = IH(z)and assume that all actuators have the same dynamics, so A(z) = IA(z).
This means that each matrix in TFe
ω→ω(z)is a diagonal matrix and therefore TFe
diagonal matrix function.
ω→ω(z) = (I−A(z)H(z)z−1)−1A(z)
= (I(1 −A(z)H(z)z−1))−1IA(z)
Then, the last part of the open loop is from ωto ˙
Ω, as shown by Figure 3. Using this ﬁgure, the
transfer function is calculated in Eq. (24). Note that for this analysis, disturbances are not taken
Ω(z) = G1+z−1
Using these intermediate results, the open loop transfer function of the entire system is shown in
Ω(z) = TFω→˙
= (G1+G2−G2z−1)I(1 −A(z)H(z)z−1)−1A(z)(G1+G2−G2z−1)+
Using Eq. (25) and Figure 2, we can calculate the closed loop transfer function of the entire system
in Eq. (26):
Ω(z) = (I+TF ˙
= (I+I(1 −A(z)H(z)z−1)−1A(z)IH(z)z−1)−1I(1 −A(z)H(z)z−1)−1A(z)
From this equation, it appears that the closed loop transfer function from the virtual input to the
angular acceleration is in fact the actuator dynamics A(z). In most cases, the actuator dynamics can
be represented by ﬁrst or second order dynamics. Note that this shows the importance of applying
the H(z)ﬁlter on the input as well. By doing this, a lot of terms cancel and all that remains is the
Now, consider the transfer function from disturbances d(see Figure 2) to the angular accelera-
tion. The derivation is given in Eq. (27) in which use is made of Eq. (25).
Ω(z) = (I−TF ˙
= (I+I(1 −A(z)H(z)z−1)−1A(z)IH(z)z−1)−1I
With Eq. (27) we show that disturbances in the angular acceleration are rejected as long as the
actuator dynamics and the designed ﬁlter are stable. The term A(z)H(z)z−1will go to 1 over time,
with a response determined by the actuator dynamics, ﬁlter dynamics and a unit delay. This means
that the faster the angular acceleration is measured, the faster the drone can respond and the faster
the actuators can react, the faster the disturbance is neutralized.
D. Attitude Control
The angular acceleration of the MAV is accurately controlled by the system shown in Figure 2.
To control the attitude of the MAV, a stabilizing angular acceleration reference needs to be passed
to the INDI controller. This outer loop controller can be as simple as a Proportional Derivative
(PD) controller (a gain on the rate error and a gain on the angle error), as shown in Figure 5. Here,
ηrepresents the attitude of the quadcopter. The beneﬁt of the INDI inner loop controller is that
the outer PD controller commands a reference, independent of the eﬀectiveness of the actuators
(including the inertia of the quadrotor).
This means that the design of this controller depends only on the speed of the actuator dynamics
A(z). In case the actuator dynamics are known (through analysis of logged test ﬂights for instance),
a value of Kηand KΩcan be determined that give a stable response.
This outer loop controller does not involve inversion of the attitude kinematics as has been done
in other work . However, the attitude angles for a quadrotor are generally small, in which case
the inversion of the attitude kinematics can be replaced with simple angle feedback.
ηref Ωref ˙
Fig. 5 The design of the attitude controller based on the closed loop response of the INDI
E. Altitude Control
The INDI controller derived in the beginning of this section controls the angular acceleration
around the axes x,yand z, which corresponds to roll, pitch and yaw. However, there is a fourth
degree of freedom that is controlled with the rotors, which is the acceleration along the z-axis.
Control of this fourth axis is handled by a separate controller. This controller scales the average
input to the motors to a value commanded by the pilot, after the input has been incremented by
the INDI controller.
IV. Adaptive INDI
The INDI approach only relies on modeling of the actuators. The control eﬀectiveness depends
on the moment of inertia of the vehicle, the type of motors and propellers. A change in any of these
will require re-estimation of the control eﬀectiveness. Moreover, the control eﬀectiveness can even
change during ﬂight, due to a change in ﬂight velocity, battery voltage or actuator failure.
To counteract these problems and obtain a controller that requires no manual parameter es-
timation, the controller was extended with onboard adaptive parameter estimation using a Least
Mean Squares (LMS)  adaptive ﬁlter. This ﬁlter is often used in adaptive signal ﬁltering and
adaptive neural networks.
The LMS implementation is shown in Eq. (28), where µ1is a diagonal matrix whose elements
are the adaptation constant for each input and µ2is a diagonal matrix to adjust the adaptation
constants per axis. This is necessary as not all axes have the same signal to noise ratio.
The LMS formula calculates the diﬀerence between the expected acceleration based on the
inputs and the measured acceleration. Then it increments the control eﬀectiveness based on the
error. The control eﬀectiveness includes both G1as well as G2, as is shown in Eq. (29). Clearly,
when there is no change in input, the control eﬀectiveness is not changed. The reverse is also
true: more excitation of the system will result in a faster adaptation. This is a beneﬁt of the LMS
algorithm over, for instance, recursive least squares with a ﬁnite horizon because recursive least
squares will ’forget’ everything outside the horizon.
G(k) = G(k−1) −µ2
Note that the ﬁltering can be diﬀerent for the online parameter estimation than for the actual
control. Equation (28) makes use of ∆˙
Ωf, which is the ﬁnite diﬀerence of ˙
Ωfin the control Eq.
(21). Since diﬀerentiating ampliﬁes high frequencies, a ﬁlter that provides more attenuation of these
high frequencies is necessary. We still use the second order ﬁlter described by Eq. (15), but with
ωn= 25 rad/s and ζ= 0.55.
When an approximate control eﬀectiveness is given before takeoﬀ, the adaptive system will
estimate the actual values online, and thereby tune itself. The only knowledge provided to the
controller is an initial guess of the control eﬀectiveness. It is generally not possible to take oﬀ
without any estimate of the control eﬀectiveness, because the UAV might crash before the adaptive
system has converged.
The choice of the adaptation constants µ1and µ2determines the stability and the rate of
adaptation. By making these constants larger, a faster convergence is achieved. By making them
too large, the adaptation will no longer be stable. The theoretical limit has been discussed in the
literature  and it depends on the autocorrelation matrix of the input to the ﬁlter. In practice,
the ﬁlter stability deteriorates before the theoretical limit, so in order to ﬁnd a good adaptation
constant some tuning is required.
V. Experimental Setup
To validate the performance of the INDI controller developed in Section III and the adaptive
parameter estimation from Section IV, several experiments were conducted. These experiments were
performed using the Bebop quadcopter from Parrot shown in Figure 1. The Bebop weighs 396.2
grams and can be equipped with bumpers, which are 12 grams per bumper. For these experiments,
the bumpers where not equipped unless explicitly stated. The quadcopter was running the Paparazzi
open source autopilot software, which contains all the code for wireless communication, reading
sensor measurements etc. The accelerometer, gyroscope and control loops were running at 512 Hz.
Four experiments test the key properties of the controller:
• Disturbance rejection
During these experiments, the reference attitude and average thrust level were controlled by
a pilot and sent to the drone over WiFi. All other computations were done on the drone itself,
including the online adaptation.
In order to put the responsiveness of the system to the test and make sure that the angular
acceleration reference is tracked by the INDI controller, a doublet input was applied on the attitude
roll angle. The amplitude of the doublet is 30 degrees and the period is half a second (0.25 seconds
positive and 0.25 seconds negative). This test is only done for the roll and not for the pitch,
because there is no fundamental diﬀerence between these axes. The yaw axis is covered separately
in experiment V D. Note that this experiment is performed without the adaptation.
The performance is compared to a manually tuned PID controller. The INDI controller is not
expected to be faster or slower than a traditional PID controller, because the result of Eq. (26)
shows that the response of the INDI inner loop is simply the actuator dynamics. Considering that
the outer loop is a PD controller, the rise time and overshoot should be similar.
Finally, this test will also be performed with an INDI controller that does not contain the ﬁlter
delay compensation, so by using ω0in the controller increment instead of ωf. It is expected that
this will not ﬂy well, because in Section III C we showed that with this compensation all terms
cancel and the closed loop transfer function reduces to IA(z).
By inspection of Figure 2, we can get a feel for what will happen if we omit this ﬁlter compen-
sation. When there is an angular acceleration error, a control increment e
ωwill be the result, which
is added to ω0to produce ωc.ωcgoes through the actuator dynamics to produce the new ω. The
next time step, the result of this new ωdoes not yet appear in ˙
Ωf, because it is ﬁltered and therefore
delayed. Therefore, e
ωwill be the same. However, ω0did update, so ωcwill be incremented even
more, while we are still waiting to see the result of the ﬁrst increment in ˙
B. Disturbance Rejection
The disturbance rejection property is validated by adding a disturbance to the system. One
possibility would be to apply aerodynamic disturbances by ﬂying in the wake of a big fan. The
disturbances occuring would be realistic, but not very repeatable. Moreover, the magnitude of the
disturbance would be unknown.
Instead, it is possible to apply a disturbance in the form of a step function to the system. This
is done by adding a weight of 42.5 grams to a container located in an oﬀ-centered position on the
quadrotor while it is ﬂying, as shown in Figure 6. The container is located on the front of the drone
and has a distance of about 11 cm to the center of gravity, so any weight added will shift the center
of gravity forward. This will cause a misalignment of the thrust vector with respect to the center of
gravity and therefore a pitch moment. This moment will be persistent and therefore have the form
of a step disturbance. This is indicated with din Figure 2. Although this moment is created with a
center of gravity shift, the situation is the same as in the case of a persistent gust or an unmodeled
Fig. 6 The container attached to the nose of the quadrotor with one weight inside.
A normal PID controller would respond to such a disturbance very slowly, because it takes time
for the integrator to accumulate. But the introduction of the INDI inner loop leads to a cascaded
control structure, which is much more resistant to disturbances than a single loop design .
Fig. 7 The Bebop quadrotor with bumpers.
Because of this, the reference pitch angle is expected to be tracked shortly after the disturbance.
The Bebop quadcopter has the possibility to ﬂy with bumpers, as is shown in Figure 7. Though
these bumpers only weigh 12 grams a piece, they are located far from the center of gravity and
therefore increase the moment of inertia. Furthermore, they can inﬂuence the airﬂow around the
propellers. These system changes aﬀect the G1and G2matrices. Therefore, the adaptive algorithm
from Section IV should deal with adding or removing the bumpers.
First, two ﬂights are performed to show the eﬀect of adding or removing the bumpers when
the adaptive algorithm is not active. For the ﬁrst ﬂight, the bumpers are added, while the G1and
G2matrices correspond to the quadrotor without bumpers. For the second ﬂight, the bumpers are
removed and the Gmatrices from the quadrotor with bumpers are used. In both ﬂights, doublets
are performed like in Section V A. The performance is expected to degrade compared to the previous
results for both cases, as the Gmatrices do not correspond to what they should be.
Second, the ability of the quadrotor to adapt its G1and G2matrices is tested. In this ex-
periment, the drone starts with bumpers equipped, but with system matrices that represent the
conﬁguration without bumpers. The pilot ﬂies the drone in a conﬁned area while performing some
pitch, roll and yaw maneuvers to excite the system. While ﬂying, the correct matrices should be
estimated. Then, the Bebop is landed and the bumpers are removed. After take oﬀ, the matrices
should converge to their original state.
Finally, doublets are performed with and without the bumpers equipped, while the adaptation
algorithm is active. We expect the same performance as in Section V A.
D. Yaw Control
The purpose of this experiment is to show the improvement in yaw performance due to the
incorporation of the rotor spin-up torque in the controller design. This is done by applying a
doublet input on the yaw setpoint. The amplitude of the doublet is 5 degrees and the period is one
second (0.5 seconds positive and 0.5 seconds negative). As a comparison, the same experiment is
performed with a traditional PID controller. This PID controller is manually tuned to give a fast
rise time with minimal overshoot.
Additionally, the same test is performed with a zero G2matrix. Here we expect an oscillation,
because the persistent eﬀect of a change in rotor angular velocity on the yaw axis is small. We take
the pseudoinverse in Eq. 21, so the resulting gain will be very large. Because there is the angular
momentum eﬀect of the propellers, the initial angular acceleration will be larger than expected, and
the controller will start to oscillate.
This section deals with the results of the experiments described in Section V. The angular
acceleration shown in the plots in this section is not the onboard estimate of the angular acceleration,
because it is delayed through ﬁltering. Instead, it is computed after the experiment from the ﬁnite
diﬀerence of the gyroscope data. The signal is ﬁltered with a fourth order Butterworth ﬁlter with
a cutoﬀ frequency of 15 Hz. It is ﬁltered twice, forward and reverse, resulting in a zero phase
(non-causal) ﬁlter. For the actual control, the onboard ﬁltered (and delayed) angular acceleration
Figure 8 shows the angular acceleration around the xaxis denoted by ˙pand the reference angular
acceleration denoted by ˙pref . Additionally, the reference is ﬁltered with the actuator dynamics,
resulting in ˙pref A. This signal is the angular acceleration that is expected based on the calculations
in Section III C, speciﬁcally Eq. (26). It might seem that the controller does not track the reference
well because it lags behind the reference, but this was expected based on the model of the actuator
dynamics. The angular acceleration is actually very close to the expected angular acceleration ˙pref A.
Finally, we also show the angular acceleration as calculated on board the quadrotor using the second
order ﬁlter. The ﬁltered angular acceleration on board the quadrotor is signiﬁcantly delayed with
respect to the actual angular acceleration, which is why we will run into problems if we don’t take
this delay into account in the INDI controller.
8.2 8.4 8.6 8.8 9 9.2
Fig. 8 Angular acceleration in the roll axis during doublet input.
The outer loop controller, which generates the angular acceleration reference to track, was
designed such that the resultant accelerations give a desired response of the roll angle, shown in
Figure 9. From this ﬁgure, it can be seen that the quadcopter reaches its reference roll angle within
0.2 seconds with a very small overshoot.
Roll angle [deg]
8.2 8.4 8.6 8.8 9 9.2
Fig. 9 The roll angle during the doublet for the INDI controller.
The roll angle response of the PID controller is shown in Figure 10. As expected, the PID
controller performs very similar to the INDI controller in terms of rise time and overshoot. The
integral gain included in the PID controller, which needs to eliminate steady state oﬀsets, degrades
the dynamic performance of the closed loop system. This shows that the INDI controller marginally
improves the performance of a traditional PID controller in terms of responsiveness for the roll.
Roll angle [deg]
10.4 10.6 10.8 11 11.2 11.4
Fig. 10 The roll angle during the doublet for the PID controller.
As discussed above, the onboard ﬁltered measurement of the angular acceleration is signiﬁcantly
delayed. If we remove the ﬁlter delay compensation from the INDI controller, the quadrotor was
severely oscillating, as can be seen in Figure 11. The doublet was not performed as this did not
seem safe. The oscillation might be reduced by lowering Kηand KΩ, but this will make the response
slower as well. From this ﬁgure, we can conclude that the ﬁlter delay compensation is an important
part of the INDI controller and is crucial in obtaining good performance with an INDI controller.
Roll angle [deg]
13 13.5 14 14.5 15 15.5 16
Fig. 11 The roll angle for the INDI controller without ﬁlter compensation.
B. Disturbance Rejection
The weight, shown in Figure 6, was placed in the container attached to the nose of the quadrotor
by hand. The weight was placed in the container gently, but it probably arrived in the container
with some small velocity. The disturbance in the angular acceleration is therefore a combination of
a step and a delta pulse.
Figure 12 shows the angular acceleration that is the result of the disturbance. From the ﬁgure,
it is clear that the disturbance happened just after 13 seconds. As the angular acceleration increases
in the negative direction, the reference angular acceleration starts to go the opposite way, because
now an angular rate and a pitch angle error start to arise. About 0.1 seconds after losing track
of the reference, the angular acceleration again coincides with the expected angular acceleration,
having overcome the disturbance in the angular acceleration.
12.6 12.8 13 13.2 13.4 13.6 13.8
Fig. 12 The angular acceleration during the disturbance.
This results in a pitch angle with no steady state error as can be seen from Figure 13. After
0.3 seconds, the pitch angle is back at zero. To show that the weight in the container really is
a step disturbance, which can be compared to a constant aerodynamic moment, consider Figure
14. It shows the diﬀerence of the rotational rate of the front and rear motors divided by four:
(ω1+ω2−ω3−ω4)/4. This indicates the average magnitude in Rounds Per Minute (RPM) that
each motor contributes to the pitch control (see Eq. (7)). Clearly, there is a diﬀerence before and
after the disturbance which can be quantiﬁed as an average change of 578 RPM over the interval
[12.6 13.0] versus [13.4 13.8]. This demonstrates that the disturbance was really a step and that the
Pitch angle [deg]
12.6 12.8 13 13.2 13.4 13.6 13.8
Fig. 13 The pitch angle during the disturbance.
INDI controller can rapidly cope with such a disturbance.
12.6 12.8 13 13.2 13.4 13.6 13.8
Fig. 14 The diﬀerence between the rotational rate of the front motors and the rear motors.
Figure 15 shows the same experiment performed with a PID controller. Of course, the weight
was not dropped in exactly the same manner and with the same velocity, so the initial disturbance
was probably diﬀerent. However, the persisting disturbance is the same, because the weight has
exactly the same mass. It takes about 1.5 seconds before the pitch angle is back at zero again,
which is approximately 5 times longer than for the INDI controller. One might say that the integral
gain of the PID controller should be larger, but this will deteriorate the performance in the previous
Pitch angle [deg]
15 15.5 16 16.5 17 17.5
Fig. 15 The pitch angle during the disturbance for the PID controller.
Figures 16 and 17 show the response to a roll doublet without adaptation if there is a mismatch
in the control eﬀectiveness. Even though the bumpers are lightweight, their eﬀect is signiﬁcant
because they are located far from the center of gravity. In Figure 16, we see what happens if the
actuators are less eﬀective than in the model, because the inertia is higher. Additional increments
of the input are needed to reach a desired angular acceleration. The oscillation occurs because this
takes more time. The oscillation can be reduced by reducing the Kηand Kωgains, at the cost of
having a slower response.
Roll angle [deg]
27.4 27.6 27.8 28 28.2 28.4 28.6
Fig. 16 Flight without adaptation, with bumpers equipped, while the control eﬀectiveness has
been determined without bumpers
In Figure 17, we see the opposite: the control eﬀectiveness is higher than what was modeled.
This results in a fast oscillation, which cannot be removed by reducing the attitude gains. This is
because the cause of the oscillation is diﬀerent: now too much input is applied to reach a certain
angular acceleration. This will happen regardless of what angular acceleration is requested by the
Roll angle [deg]
18.4 18.6 18.8 19 19.2 19.4 19.6
Fig. 17 Flight without adaptation, without bumpers equipped, while the control eﬀectiveness
has been determined with bumpers
We can conclude that the performance degrades when the modeled control eﬀectiveness does not
closely correspond to the actual control eﬀectiveness. When the adaptation algorithm is enabled,
Figures 18 through 20 show how each row of the G1matrix evolves over time as a result of the
second experiment described in section V C. The same is shown in Figure 21 for the third row of
the G2matrix. Each line represents one of the elements of that row, indicating the eﬀectiveness of
that motor on the speciﬁed axis.
Note that the drone is ﬂying in the interval of [8 54] seconds and again in [66 125] seconds; in
between these times, the drone is landed and the bumpers are removed. This is indicated by vertical
lines in the ﬁgures. A large change in eﬀectiveness due to the addition and removal of the bumpers
can be seen in the third row of the G1matrix, shown in Figure 20, which corresponds to the yaw.
Also in Figure 18 a change in eﬀectiveness can be seen between the ﬂights with and without
bumpers. Once converged, the eﬀectiveness values are stable with little noise. Upon take-oﬀ and
landing the eﬀectiveness seems to diverge for a short period of time. This is not a failure of the
adaptation algorithm, but merely the result of the interaction with the ﬂoor.
The controller is engaged once the pilot gives a thrust command that exceeds idle thrust. At
0 50 100 150
Fig. 18 The ﬁrst row of the G1matrix corresponding to the roll.
0 50 100 150
Fig. 19 The second row of the G1matrix corresponding to the pitch.
that point, the quadrotor does not produce enough lift to take oﬀ, so it is still standing on the
ﬂoor. When the INDI controller tries to attain certain angular accelerations, the quadrotor does
not rotate and the adaptation algorithm will adapt to this. When landing, these interactions with
the ﬂoor can also occur.
Notice the large diﬀerence in eﬀectiveness between the actuators in the second part of the ﬂight
in Figure 20. This illustrates the added value of adaptive INDI, as often the actuators are assumed to
perform equal to each other, while in this case they do not. These diﬀerences between the actuators
are also observed with the estimation metho d describ ed in subsection III A for multiple ﬂights. The
diﬀerences may be caused by small imperfections that are not clearly visible on some of the rotors.
Finally, we can observe how the online parameter estimation aﬀects the response to a roll doublet
in Figures 22 and 23. Regardless of whether the bumpers are equipped or not or with what control
0 50 100 150
Fig. 20 The third row of the G1matrix corresponding to the yaw.
0 50 100 150
Fig. 21 The third row of the G2matrix corresponding to the yaw.
eﬀectiveness model the quadrotor starts ﬂying, the same performance is achieved as in Section V A.
This shows the robustness of the adaptive algorithm against control eﬀectiveness changes.
D. Yaw control
Finally, consider Figure 24. It shows for each timestep the change in angular acceleration in
the yaw axis, ∆ ˙r, during the large control inputs discussed above. A careful reader up until this
point may wonder: ’Is the rotor spin-up torque really signiﬁcant? Can we not omit the G2matrix?’.
The ﬁgure shows the predicted change in angular acceleration based on the change in motor speeds
according to Eq. (21), which is a close match. In green, the ﬁgure shows the predicted change in
angular acceleration if we neglect G2, denoted by ∆ ˙rsimple . Clearly, the motor spin-up torque is
Roll angle [deg]
24.2 24.4 24.6 24.8 25 25.2 25.4
Fig. 22 Flight with adaptation, with bumpers equipped, while the control eﬀectiveness has
been determined without bumpers
Roll angle [deg]
24.2 24.4 24.6 24.8 25 25.2 25.4
Fig. 23 Flight with adaptation, without bumpers equipped, while the control eﬀectiveness has
been determined with bumpers
24 24.5 25 25.5 26
Fig. 24 The change in angular acceleration in the yaw axis along with the predicted change.
Moreover, if we try to ﬂy with a zero G2matrix, the resulting oscillation is so strong that a
takeoﬀ is not possible. In order to ﬂy without this matrix, we cannot use the estimated values for
the control eﬀectiveness in the yaw axis. Instead, we can take a higher eﬀectiveness for the model
parameters than in reality in order to avoid overshooting the reference angular acceleration due to
the rotor spin-up torque that is now not taken into account. Figure 25 shows that it is possible to
ﬂy with a zero G2matrix, at the cost of a severe performance penalty.
Yaw angle [deg]
23.5 24 24.5 25 25.5
Fig. 25 The yaw angle during the doublet for the INDI controller without G2matrix.
If we do take the rotor angular momentum into account, Figure 26 shows the resultant doublet
response of the yaw angle. Compare this with Figure 27, which shows the doublet response for the
PID controller. The INDI controller clearly has a faster rise time and less overshoot.
Yaw angle [deg]
24 24.5 25 25.5 26
Fig. 26 The yaw angle during the doublet for the INDI controller.
Yaw angle [deg]
12.5 13 13.5 14 14.5 15
Fig. 27 The yaw angle during the doublet for the PID controller.
Adaptive Incremental Nonlinear Dynamic Inversion is a very promising technique for control
of Micro Aerial Vehicles (MAV). Due to incorporation of the spin-up torque, fast yaw control is
possible, which is typically very slow on a quadrotor. The disturbance rejection capabilities are
vital when ﬂying in windy conditions or with MAVs that have complex aerodynamics. Because
unmodeled aerodynamic moments are measured with the angular acceleration, no complex aerody-
namic modeling is needed. Even the control eﬀectiveness matrices are shown to be adapted online,
resulting in a controller that can handle changes in the MAV conﬁguration and needs little eﬀort to
set up on a new platform. Only when a high performance outer loop is required is some knowledge
of the actuator dynamics needed. These properties result in a very ﬂexible and powerful controller.
This work was ﬁnanced by the Delphi Consortium. The authors would like to thank Bart Remes
and the MAVLab for their support.
 Mahony, R., Kumar, V., and Corke, P., “Multirotor Aerial Vehicles: Modeling, Estimation and Control
of Quadrotor,” IEEE Robotics &Automation Magazine, Vol. 19, 2012, pp. 20–32,
 Fresk, E. and Nikolakopoulos, G., “Full Quaternion Based Attitude Control for a Quadrotor,” in “Eu-
ropean Control Conference,” IEEE, 2013, pp. 3864–3869.
 da Costa, R., Chu, Q., and Mulder, J., “ Reentry Flight Controller Design Using Nonlinear Dynamic
Inversion,” Journal of Spacecraft and Rockets, Vol. 40, No. 1, 2003, pp. 64–71,
 Smith, P. R., “A Simpliﬁed Approach To Nonlinear Dynamic Inversion Based Flight Control,” in “At-
mospheric Flight Mechanics Conference and Exhibit,” AIAA Paper 1998-4461, pp. 762–770,
 Bacon, B. J. and Ostroﬀ, A. J., “Reconﬁgurable Flight Control Using Nonlinear Dynamic Special
Accelerometer Implementation,” AIAA Paper 2000-4565,
 Cox, T. H. and Cotting, M. C., “ A Generic Inner-Loop Control Law Structure for Six-Degree-of-Freedom
Conceptual Aircraft Design,” AIAA Paper 2005-31,
 Ostroﬀ, A. J. and Bacon, B. J., “Enhanced NDI Strategies For Reconﬁgurable Flight Control,” in
“Proceedings of the American Control Conference,” , 2002, pp. 3631–3636,
 Sieberling, S., Chu, Q. P., and Mulder, J. A., “ Robust Flight Control Using Incremental Nonlinear
Dynamic Inversion and Angular Acceleration Prediction,” Journal of Guidance Control and Dynamics,
Vol. 33, No. 6, 2010, pp. 1732–1742,
 Bacon, B. J., Ostroﬀ, A. J., and Joshi, S. M., “ Reconﬁgurable NDI Controller Using Inertial Sensor
Failure Detection & Isolation,” IEEE Transactions On Aerospace And Electronic Systems, Vol. 37,
No. 4, 2001, pp. 1373–1383,
 Koschorke, J., Falkena, W., van Kampen, E.-J., and Chu, Q. P., “ Time Delayed Incremental Nonlinear
Control,” AIAA Paper 2013-4929,
 B. Theys, G. D., Andrianne, T., Hendrick, P., and Schutter, J. D., “Wind Tunnel Testing of a VTOL
MAV Propeller in Tilted Operating Mode,” in “International Conference on Unmanned Aircraft Sys-
tems,” IEEE, 2014, pp. 1064–1072,
 Chowdhary, G., Johnson, E. N., Chandramohan, R., Kimbrell, M. S., and Calise, A., “Guidance and
Control of Airplanes Under Actuator Failures and Severe Structural Damage,” Journal of Guidance
Control and Dynamics, Vol. 36, No. 4, 2013, pp. 1093–1104,
 Bedford, A. and Fowler, W., Engineering Mechanics Dynamics, Prentice Hall, Pearson Education South
Asia Pte Ltd, ISBN 981-06-7940-8, pp. 507–515, 2008.
 Simplicio, P., Pavel, M., van Kampen, E., and Chu, Q., “An acceleration measurements-based approach
for helicopter nonlinear ﬂight control using Incremental Nonlinear Dynamic Inversion,” Control Engi-
neering Practice, Vol. 21, No. 8, 2013, pp. 1065–1077,
 Haykin, S. and Widrow, B., Least-Mean-Square Adaptive Filters, Wiley, ISBN: 978-0-471-21570-7, pp.
 Love, J., Process Automation Handbook, Springer-Verlag London, pp. 173–178, 2007,
 Reader can download the code at https://github.com/EwoudSmeur/paparazzi in the branch