Calibration Procedure for an Inertial Measurement Unit Using a

6-Degree-of-Freedom Hexapod
Øyvind Magnussen, Morten Ottestad and Geir Hovland

Abstract— In this paper a calibration procedure for an Iner-

tial Measurement Unit (IMU) mounted on an Unmanned Aerial
Vehicle (UAV) is presented. Calibration of the sensor when it
is mounted on the UAV is attractive because it combines cali-
bration of the internal sensor parameters with the translational rotational transformation relative to the UAV’s body frame
and rotational offsets of the IMU relative to the body frame of are identified.
the UAV. A stepwise calibration procedure is presented, based Many authors have proposed different methods for IMU
on motion profiles and measurements generated by a 6-degree- calibration, see [8] and the references therein. Common for
of-freedom hexapod. The experimental results demonstrate that these methods is the utilization of the fact that ideally the
the linear acceleration and angular velocity measurements of
the IMU sensor ADIS16400 can be calibrated to typically 0.17% norm of the measured output of the accelerometer and gyro
and 0.25% of the ADIS16400 measurement range for linear cluster is equal to the magnitude of the applied force and
acceleration and rotational velocity, respectively. rotational velocity. As noted in [8] the main drawback of
this approach is that not all the sensor parameters of the IMU
I. INTRODUCTION are observable. This in turn implies that these uncalibrated
parameters must be taken into account in the integration of
Unmanned aerial vehicles (UAVs) are often designed the IMU in the UAV, for example with advanced filtering
with an inherent instability and such systems can only be algorithms. There are also papers describing the process of
stabilized by feedback controllers using accurate sensors. In calculating the new CG of the UAV if its shifted to a new
order to achieve high performance and robustness of such location during operation [7]. Algorithms aligning two IMU
systems, it is essential that the sensor measurements are in two reference frames has also been established [11]. But
as accurate as possible. A typical sensor used in UAVs is this does not count for misalignment of the internal axis-to-
an Inertial Measurement Unit (IMU) which measures linear axis or scaling of the IMU. The research topic is not new, and
accelerations and angular velocities in three dimensions. The old articles of how to automatically align an IMU relative to
IMU calibration is considered as one of the main challenges the earth [6] is also developed.
in inertial navigation. Even if highly accurate IMUs are The main benefit of the work presented in this paper,
available from sensor manufacturers it is still desirable to is the flexibility of the hexapod to generate motions in
perform an independent calibration and verification of the selected directions. Hence, all the parameters of the IMU
performance against a more accurate measurement device. can be calibrated and the need for complex, online filtering
It is also desirable to calibrate the IMU after it has been techniques in the UAV real-time controller is reduced.
mounted on the UAV, since the actual location and orientation
of the sensor relative to the body frame of the UAV are II. SYSTEM DESCRIPTION
additional error sources in the models and the feedback
control algorithms. To ease the attitude calculations the IMU A. Inertial Measurement Unit
is often put in the center of gravity (CG) of the UAV with An IMU, ADIS16400 from Analog Devices, measures
alignment of the sensor axes with the body frame. But UAVs, acceleration aI and angular velocity ωI for the origin of
especially quadrotors, have limited space available and it can the IMU frame for the three axes XI , YI , ZI as illustrated
be difficult to place the IMU at the desired position. in Fig. 1. The subscripts I refer to the IMU’s coordinate
In this paper an approach for calibration and verification axes. In this paper, if the subscript does not begin with I,
of an IMU mounted on a UAV is presented. The approach the variable is always referred to the body frame. In general,
uses the motion and sensing capabilities of a 6-degree-of- depending on how the IMU is mounted on the UAV, there
freedom (6-DOF) hexapod. By using a hexapod, it is possible will be rotations of the IMU’s coordinate axes relative to the
to generate pure translational or rotational motion in selected body axes (φI , θI , ψI ), illustrated in Fig. 1.
directions, and this feature is exploited in a stepwise calibra- In Fig. 2 the origin of the body frame is located at CG of
tion procedure presented in this paper. Sensor parameters the UAV. This is also the center of rotation for any rotational
such as offsets, scaling factors, the alignment of the sensor’s motion. The IMU frame is located off-center.
internal coordinate axes as well as the translational and A off-center rotation of the IMU relative to the body
frame, as in Fig 3, will result in additional accelerations
The authors are with Faculty of Engineering and Science, Department of
Engineering, University of Agder, Norway. E-mail of corresponding author: perpendicular to the rotational axis, i.e. a radial acceleration
[email protected] aR and a tangential acceleration aT .
 Zb
acceleration of the body frame, ag is the gravity vector in
ψI
Yb the body frame. Ideally, the matrix Ka is a pure rotational
matrix [2] given by the three angles φI , θI , ψI . However,
ZI YI θI if the IMU’s coordinate axes XI , YI , ZI have an axis-to-
axis misalignment, or the scaling of at least one of the axes
XI φ I are different than the others Ka will no longer be a proper
Xb rotational matrix [5].
The IMU’s gyroscopes measure the angular rates ωI ,
which are not affected by the location of the IMU relative
to the CG, only its orientation given by φI , θI , ψI .

ωb = Kω · ωI (5)
Fig. 1. IMU frame relative to the body frame ωI fI − ωI,o − ωI,n
= ω (6)

where Kω is the gain and cross correlation matrix ωb is

the angular rate of the body frame, ωI,n is the noise of the
Yb gyroscope, ωI,o is the gyroscope bias and ω
fI is the gyroscope
YI measurement including noise and bias.
XI
ZI B. Stewart Platform

Z b Xb A Stewart Platform shown in Fig. 4 is used to calibrate

the IMU. The Stewart Platform has six degrees of freedom
IMU
and can manipulate rotations and translations. The Stewart
Platform is also able to set the center of rotation to a specific
location. This makes it possible to mount any UAV on top
of it, and to align the Stewart Platforms center of rotation
Fig. 2. Rotation and translation of IMU frame relative to body frame. with the UAV’s CG. Different types of motion profiles
with different frequencies and amplitudes were generated.
Six degrees of freedom data were recorded for positions,
These accelerations specified in the body frame are de-
velocities and accelerations in the fixed world frame. These
scribed by [1] and [10]:
measurements were transformed to the UAV body frame.
aR = ωb × (ωb × R) (1)
aT = αb × R (2)
where × is the vector cross product, R = [Rx Ry Rz ]T is
the distance from the origin of the body frame to the IMU
frame, ωb = [ωx ωy ωz ]T is the angular velocity of the body
frame and αb = [αx αy αz ]T is the angular acceleration
of the body frame. The relationship between the various
accelerations are
ab + aR + aT + ag = Ka · aI (3)
aI = aeI − aI,n − aI,o (4)
where aI is the raw acceleration measurement from the IMU,
including noise aI,n and bias aI,o in the IMU frame, Ka ∈
R3 is a gain and cross correlation matrix, ab is the translation

aT
R
aR
ω

Fig. 4. Stewart Platform used for IMU calibration experiments.

Fig. 3. Radial and tangential accelerations when IMU is mounted off-
center compared to body frame.
 fixed position. The bias is subtracted from the gyroscope
as given by (6). Acceleration bias can be calculated by
C. Laser Tracker taking two measurements, where one measurement is taken
To verify the motion of the Stewart platform a FARO laser while the sensor is turned 180 degrees compared to the first
tracker as shown in Fig. 5 was used to verify its position. In measurement, this is however not done in this paper.
the verification it was found that the accuracy of the internal
B. Scaling and Correlation
measurements of the Stewart Platform were comparable with
the laser tracker, typically in the range 10-30 µm. Hence, In order to calculate the two matrices Ka and Kω , two
in the experimental results presented in this paper, only the separate tests had to be done, where the first test calculates
internal measurements from the Stewart Platform were used. Ka and the second calculates Kω . The series of measurement
data were synchronized in time for both the Stewart Platform
and the IMU. The motion to calculate Ka consists of
rapid translative motions, combined with a slow rotation
around every body axis. The translative motions manipulate
the accelerometer directly, while the slow rotations makes
the gravity affect each axis of the IMU dynamically. The
Stewart Platform has a force limit for each actuator which
in turn limits the maximum acceleration. Adding the rotating
gravity vector relative to the IMU increases the acceleration
measured by the accelerometer and hence the the signal-
noise ratio. The rotation frequency of 0.01Hz is very low,
as a result ag in eq.(3) dominates aR and aT which can be
neglected from the equations.
The motion to calculate Kω was a pure rotation about the
Fig. 5. FARO laser tracker used to verify the internal measurements of the body axes Xb , Yb , Zb . The amplitude and frequencies were
Stewart Platform. higher than for Ka resulting in larger rotational velocities.
Typical frequencies generated by the Stewart Platform were
in the range 0.6-0.7 Hz with an amplitude of 10 degrees.
III. CALIBRATION CONCEPT
For both the accelerometer and gyroscope the problem can
The UAV is mounted on top of the Stewart Platform with be written on the form:
aligned axes and CG, shown in FIG 6.
Ωb = K · ΩI (7)

where Ωb is a 3 × n matrix with the true datasets from

the Stewart Platform, K ∈ R3 is the correlation and scaling
matrix, ΩI is a 3 × n matrix with the measured data, either
Yb acceleration or rotational velocity, from the IMU and n is
the number of measurement samples. The problem is to find
YI XI the best fit for K over the data sets. As noted in [9], several
ZI
Z b Xb
optimization criteria have been used for estimation purposes
over the years, but the most important ones, at least in the
sense of having had the most applications, are criteria that
are based on quadratic cost functions. The most common
among these is the linear least-squares criterion.
Each column in (7) represents a data set, X, Y and Z
respectively. The objective of the least-squares problem is
Fig. 6. UAV mounted for calibration to determine the vectors in the column space of K that is
closest to the vectors in Ωb in the least-squares sense.
The calibration consists of three different stages. min ||Ωb − KΩI ||2 (8)
A) Remove offsets ΩI

B) Calculate scaling and correlation matrices The final solution to the problem is:
C) Find the location of the inertial measurement units Rx ,
Ry and Rz positions Ωb = K · Ω I
Ωb · ΩTI = K · ΩI · ΩTI
A. Offset Calibration
K = Ωb · ΩTI · (ΩI · ΩTI )−1
The gyroscope bias are removed by calculating the average
output over a period of time, with the gyroscope in a K = Ωb Ω+
I (9)
 where Ω+ I is the Moore-Penrose pseudoinverse and will
exist and be unique for any ΩI [4]. This approach (9) is
used in order to calculate the two matrices Ka and Kω . axes, the method will not be able to separate the rotation
around the x and the y-axis. The resultant K-matrix will be
C. Location of the inertial measurement unit      
ωb,x 0.5 0.5 0 ωI,x
The last step is to calibrate the position R = [Rx Ry Rz ]T ωb,y  = 0.5 0.5 0 · ωI,y 
of the IMU relative to CG. The distance vector R can be ωb,z 0 0 1 ωI,z
calculated from (1) and (2) by rotating around one axis at a
time. Rotating around the body frame x-axis yields: and if the phase difference of ωb,x and ωb,y is π rad, the
rotation of the Xb - and Yb -axis will be opposite and a
ab,x = 0 negative sign will occur in the correlation matrix:
2
ab,y = −αb,x · Rz − ωb,x · Ry
     
ωb,x 0.5 −0.5 0 ωI,x
2
ab,z = αb,x · Ry − ωb,x · Rz ωb,y  = −0.5 0.5 0 · ωI,y 
ωb,z 0 0 1 ωI,z
and in matrix form:
    In order to overcome this, a sine-function with different
ab,x 0 0   frequencies were used.
α
ab,y  = −Rz −Ry  · b,x
2 (10)
ωb,x ωb,x = A sin (2πfx t)
ab,z Ry −Rz
ωb,y = A sin (2πfy t)
Solving the equations using the approach from eq.(9)
ωb,z = A sin (2πfz t)
resolves two sets of Ry and Rz , because of the structure
of the matrix in eq.(10). Limitations of the Stewart Platform where the amplitude A = 4 degrees and the frequency fx =
results in a very low value of ω 2 compared to α. This makes 0.6 Hz, fy = 0.65 Hz, fz = 0.7 Hz. Using the approach
the estimates of Ry , Rz which are multiplied with the angular from (9) with a number of 10.000 sampled datasets from the
acceleration αx more trustworthy than the same estimates Stewart Platform and the IMU gives:
multiplied with the angular velocity ωx2 . Because of this  
result the estimates multiplied with the accelerations are kept 0.95 0.29 0.01
as the final calibrations. The resulting equations are: Kω = −0.29 0.95 0.01 (13)
    −0.01 −0.01 1.00
ab,x 0
ab,y  = −Rz  · αb,x The same approach was followed for the accelerometer. The
(11)
resultant Ka matrix is:
ab,z Ry  
0.97 0.27 0.02
Doing this for each of the three axes results in two calcu- Ka = −0.30 0.97 −0.01 (14)
lations for each distance. If the two datasets are identical it 0.01 0.01 1.00
is a verification of the test method of the IMU.
A result of the measured angular velocity wI multiplied
IV. EXPERIMENTS with the correction matrix Kω is shown in Fig. 7, and for
the accelerometer in Fig. 8. The figures are for a different
A. Remove bias dataset than for the calculation of Kω and Ka .
The first experiment was to remove the gyro offsets. The 15
Stewart Platform
offset was calculated as the mean value of 10.000 samples: IMU Corrected
10 IMU Raw
Angular velocity (deg/s)

 
−0.0043 5
rad
ωI,o =  0.0010  (12) 0
sec
0.0048 −5

B. Scaling and correlation −10

−15
0 2 4 6 8 10 12 14 16
The scaling and cross correlation matrix of the gyroscopes Time(s)

were then calculated. For the method to be able to separate

the data the test has to be as complex as possible. Running Fig. 7. Gyroscope scaling and correction for ωb,x , ωI,x corrected and
ωI,x raw.
a sine function at each body axis with a the same frequency,
only phase shifted, will not result in a correct calculation
of the K-matrix. For instance doing a test with the same The IMU’s orientation was measured to be:
frequency for each body axis only with a π2 rad phase on the φI = 0 deg
y-axis will not result in correct calculations. Assuming that
θI = 0 deg
the IMU reference frame is perfectly aligned with the body
frame, and the IMU is perfectly scaled and has perpendicular ψI = 17 deg
 1
Stewart Platform
TABLE I
IMU Corrected
0.5
IMU Raw R ESULT OF CALCULATION OF THE DISTANCES Rx , Ry AND Rz ,
Acceleration (m/s2)

RESULTS ARE IN METERS ( M )

0

Rotation axis Rx Ry Rz
−0.5 ωx : − −0.237 0.227
ωy : 0.362 − 0.233
−1
ωZ : 0.362 −0.239 −
−1.5
0 2 4 6 8 10 12 14 16
Time(s)

Fig. 8. Accelerometer scaling and correction for ab,x , aI,x corrected and
aI,x raw.
A comparison of the measured values with the calculated
ones, shows that the error range is less than 5 mm.
D. Full system test
Assuming the IMU is ideal (scaling is perfect and the axis- A full system test was done in order to verify the cal-
to-axis misalignment is zero) then the resultant K-matrices culated correction matrices and distances. Ka , Kω and the
would be a rotation matrix. A XYZ rotation matrix is defined calculated distances Rx , Ry , Rz was used according to (3)
as [12] and [3]: with rotation about the three body axes Xb , Yb , Zb :
 
cθI cψI sφI sθI cψI − cφI sψI cφI sθI cψI + sφI sψI ωIC = Kω · ωI (17)
C = cθI sψI sφI sθI sψI + cφI cψI cφI sθI sψI − sφI cψI 
−sθI sφcθI cφI cθI ab + ag = Ka · aI − ωIC × (ωIC × R) − ωIC ˙ × R (18)
(15) where ωIC is the corrected rotational velocity measured by
where C is the rotation matrix from the IMU frame to the the gyroscope and ωIC ˙ is the time derivative of the rotational
body frame, c is the cosine function and s is the sine function. velocity.
Inserting the measured values in the rotation matrix C yields: The results are shown in Fig. 9
 
0.96 −0.29 0 0.8
Stewart Platform
0.6 IMU Corrected
C = 0.29 0.96 0 (16) 0.4
IMU Raw
Acceleration [m/s2]

0 0 1 0.2

−0.2

The measured correction matrices are very close to the ideal −0.4

−0.6

one, Kω is the one which is closest. There are many reasons −0.8

why Ka is a bit off. One reason could be the scaling, and −1

−1.2

correlation are different for the accelerometer than for the 0 2 4 6 8

Time(s)
10 12 14 16

gyroscope. The ADIS16400 is however a high-end IMU 1

with less than 0.2 degrees misalignment of the internal Stewart Platform
IMU Corrected
IMU Raw
axes. Another issue could be that the accelerometer has a 0.5
Acceleration [m/s2]

range of ±18 g, which makes it impossible for the Stewart 0

Platform to use the whole range. The range of the excited −0.5

acceleration to the IMU compared to the accelerometer

noise is also significantly less than for the gyroscopes. −1

−1.5
0 2 4 6 8 10 12 14 16
Time(s)

C. Location of the inertial measurement unit

10.3
Stewart Platform
The result of Ka from (14) was used in order to correct 10.2 IMU Corrected
IMU Raw
10.1
the acceleration measurement when calculating the position
Acceleration [m/s2]

10

of the IMU, R, as in (3). The Stewart Platform was run 9.9

9.8
in three steps, where each step was a rotation around the 9.7

Xb , Yb , Zb axis respectively. The amplitude was 5 degrees 9.6

9.5
and with a frequency of 0.7 Hz. The same approach as for 9.4

the calculation of Ka and Kω (9) was used to calculate 9.3

0 2 4 6 8 10 12 14 16
Time(s)
R. According to (11) two sets of each distance will be
calculated. The results are displayed in table I:
Fig. 9. Comparison of the acceleration from the Stewart Platform, the
The IMU’s position was measured in meters to be: corrected acceleration from the IMU and the non-corrected acceleration
from IMU, for ax , ay and az respectively
 
0.365(m)
R =  −0.235(m) 
0.230(m)
 15
were used to calibrate different types of IMU parameters,
Stewart Platform
10
IMU Corrected
IMU Raw
such as offsets, scaling, axis-to-axis misalignment as well
Angular velocity [deg/s]

as translational and rotational offsets relative to the body

5
frame of the UAV. The experimental results show an achieved
0
accuracy of typically 0.1-0.3 m/s2 for the linear accelerations
−5
and typically 0.01-0.015 rad/s for the rotational velocities
−10
during a test motion profile with motion frequencies of 0-
−15
0 2 4 6 8 10 12 14 16
1Hz, 5o angular range and 0.1m positional range. The test
Time(s)
motion profile was independent from the motion profiles used
15
Stewart Platform
for calibration. The achieved accuracies correspond to 0.17%
IMU Corrected
10 IMU Raw and 0.25% of the ADIS16400 measurement range for linear
Angular velocity [deg/s]

5
acceleration and rotational velocity, respectively. The work
0
presented in this paper demonstrates that a 6-DOF hexapod
−5
is well suited for calibration of an IMU when mounted on a
UAV.
−10

−15
0 2 4 6 8 10 12 14 16
R EFERENCES
Time(s)
[1] M.D Ardema. Newton-Euler Dynamics. Springer, 2005.
15 [2] R.W Beard and T.W McLain. Small Unmanned Aircraft Theory and
Stewart Platform
IMU Corrected Practice. Princeton University Press, 2012.
10 IMU Raw
Angular velocity [deg/s]

[3] G. Genta. Introduction to the Mechanics of Space Robots. Springer,

5 2012.
[4] G.H Golub and C.F Van Loan. Matrix Computations, 3rd ed. The
0
Johns Hopkins University Press, 1996.
−5 [5] M.S. Greval and A.P Andrews. Kalman Filtering Theory and Practice
using MATLAB. Wiley-IEEE Press, 2008.
−10
[6] J.C. Hung and H.V. White. Self-alignment techniques for inertial mea-
−15
0 2 4 6 8 10 12 14 16
surement units. Aerospace and Electronic Systems, IEEE Transactions
Time(s) on, AES-11(6):1232 –1247, nov. 1975.
[7] Daniel Mellinger, Quentin Lindsey, Michael Shomin, and Vijay Ku-
mar. Design, modeling, estimation and control for aerial grasping
Fig. 10. Comparison of the rotational velocity from the Stewart Platform, and manipulation. In Intelligent Robots and Systems (IROS), 2011
the corrected rotational velocity from the IMU and the non-corrected IEEE/RSJ International Conference on, pages 2668 –2673, sept. 2011.
rotation velocity from IMU, for ωx , ωy and ωz respectively [8] L. Sahawneh and M.A. Jarrah. Development and calibration of
low cost mems imu for uav applications. In Mechatronics and Its
The root-mean-square (RMS) of the signals is a verifica- Applications, 2008. ISMA 2008. 5th International Symposium on,
pages 1 –9, may 2008.
tion of how good the estimates are as a mean. The RMS [9] A.H Sayed. Digital Signal Processing Fundamentals. CRC Press,
defined below is an overall verification of how close the 2012.
estimated values are to the true signal. The signal from the [10] A.A Shabana. Dynamics of Multibody Systems , 2nd ed. Cambridge
Unversity Press, 1998.
Stewart Platform is subtracted from the signals from the [11] K.J. Shortelle, W.R. Graham, and C. Rabourn. F-16 flight tests
IMU. The difference is squared so that the negative signals of a rapid transfer alignment procedure. In Position Location and
will be positive. The RMS is defined as: Navigation Symposium, IEEE 1998, pages 379 –386, apr 1998.
v [12] B. Siciliano, L. Sciavicco, L. Villani, and G. Oriolo. Robotics
u n Modelling, Planning and Control. Springer, 2011.
u1 X
RM S = t (xI − xJ )2 (19)
n i=1

where n is the number of samples, xI is the IMU sampling

data and xJ is the Stewart Platform sampling data.
TABLE II
R ESULTS OF THE RMS CALCULATION FOR THE CORRECTED DATA SET
ax 0.264
ay 0.058 (m/s2 )
az 0.177
ωx 0.0123
ωy 0.0123 (o /s)
ωz 0.0107

V. CONCLUSIONS
This paper has demonstrated a stepwise calibration method
for an ADIS16400 IMU mounted on a UAV by using a 6-
DOF Hexapod motion generator. Separate motion profiles