2012 IEEE/RSJ International Conference on

Intelligent Robots and Systems

October 7-12, 2012. Vilamoura, Algarve, Portugal

The Cubli: A Cube that can Jump Up and Balance

Mohanarajah Gajamohan, Michael Merz, Igor Thommen and Raffaello DAndrea

Abstract This paper introduces the Cubli, a 151515 cm Once the Cubli is balancing on its edge, the next two wheels
cube that can jump up and balance on a corner. Momentum are instantaneously stopped to make it stand on its corner.
wheels mounted on three faces of the cube (Fig. 1) rotate at high Finally, an inertial measurement unit (IMU) based state
angular velocities and then brake suddenly, causing the Cubli to
jump up. Once the Cubli has almost reached the corner stand- estimation is performed and the motor torques are precisely
up position, controlled motor torques are applied to make it controlled to make the Cubli continue standing.
balance on its corner. This paper tracks the development of the Whereas a relatively large portion of research tries to avoid
Cublis one dimensional prototype at ETH Zurich and presents impact or overcome the effects of it [10][12], only a small
preliminary results. portion [13], including the Cubli, takes advantage of it. While
I. I NTRODUCTION a non-impulsive force is limited by the actuator torque limits,
an impact gives forces that are well beyond the actuator
Inverted pendulum systems have a very rich history [1] and limitations.
have been widely used to test, demonstrate and benchmark
new control concepts and theories [2]. Furthermore, devel-
opment of new control algorithms for the pendulum system
itself is still an active area of research [3][5].
Compared to other 3D inverted pendulum test-beds [8],
[9] the Cubli has two unique features. One is its relatively
small footprint, (hence the name Cubli, which is derived from
the Swiss German diminutive for cube). The other feature 45
is its ability to jump up from a resting position without 35.3
any external support, not only an interesting concept for the
control engineer but also also an appealing demonstration for (a) (b)
the general public.
Fig. 2. The Cubli jump-up strategy: (a) Flat to Edge: Initially lying flat
on its face, the Cubli jumps up to stand on its edge. (b) Edge to Corner:
The Cubli goes from balancing about an edge to balancing on a corner.

Once complete, the Cubli will provide an inexpensive,

open source test-bed with a relatively small footprint for
research and education in estimation and control.
The three dimensional Cubli design started off with the
following question:
How to build a 15 cm sided cube that can jump
up and balance on its corner, using off-the-shelf
motors, batteries, and electronic components ?
Due to the structures rigidity and off-the-shelf component
constraint, only the momentum wheels allowed enough de-
sign flexibility in terms of mass distribution properties of the
Cubli. Necessary angular velocities of the momentum wheels
for jump-up were calculated assuming a perfectly inelastic
collision between the wheels and the pendulum body.
Although high angular velocities of the wheels before
Fig. 1. The CAD drawing of the Cubli with covers removed braking can be reduced by increasing the wheel inertia, i.e.,
increasing wheel mass since the wheel size is constrained,
Fig. 2 shows the jumping up strategy of the Cubli. Initially this was not taken to the extreme since this will result in
the Cubli, lying flat on its face, will jump about its edge reduced recovery angles while balancing.
by instantaneously stopping one of its momentum wheels. A gear chain between the wheel and the motor was
avoided since it would not allow the high angular velocities
The authors are with the Institute for Dynamic Systems and Control,
Swiss Federal Institute of Technology Zurich, Switzerland. The contact for jump-up and would add extra weight and volume. Al-
author is M. Gajamohan, e-mail: gajan@ethz.ch though the balancing maneuver, which requires high torques,

978-1-4673-1735-1/12/S31.00 2012 IEEE 3722

was affected by the choice of no gear, the brushless DC pivot point, Iw is the moment of inertia of the wheel and the
motors were still able provide enough torque for recovery motor rotor around the rotational axis of the motor, l is the
angles up to 7 . distance between the motor axis and the pivot point, lb is the
The objective of this paper is to present the concept of distance between the center of mass of the pendulum body
the three dimensional Cubli along with the development of and the pivot point, g = 9.81 m s2 is the gravitational
the one dimensional design, modelling, identification, and acceleration, Tm is the torque produced by the motor, and
control. This paper also presents a control procedure to Cw , Cb are the dynamic friction coefficients of the pendulum
eliminate sensor offsets during the balance maneuver. body and wheel. Since we used a motor that allowed current
set points controlled by an inner loop running at 10 kHz, the
II. T HE O NE D IMENSIONAL P ROTOTYPE
current-torque relationship can modelled as
Fig. 3 illustrates the one dimensional prototype that was
built to examine the feasibility and develop control algo- Tm = Km u, (2)
rithms for the Cubli. Similar momentum exchange wheel where, Km = 25.1 103 Nm A1 [14] is the torque
based inverted pendulum designs, except for the braking constant of the brushless DC motor used and u is the current
mechanism, can be found in [5][7]. The prototype consists input.
of a square plastic plate that holds the momentum exchange
wheel through the motor at its center and the braking B. Parameter identification
mechanism at one of its corners. The dimension of the plastic This subsection describes procedures for identifying the
plate matches the dimension of the proposed Cubli face, and parameters in (1) that cannot be directly measured.
it will be referred to hereafter as the pendulum body. The lb was identified by freely hanging the pendulum body
plate is attached to a bearing at the bottom that gives it a from different corners. To identify Iw and Cw , the momen-
single degree of freedom to pivot around its corner on a tum wheel was driven with different current steps, while
horizontal plane. pendulum body was rigidly fixed, and the time trace of w
A. System dynamics was recorded. A least squares fit with the measurements to
Iw w (t) = Km u(t) Cw w (t)
gave Iw and Cw .
b To identify Ib and Cb , after rigidly fixing the momentum
wheel with the pendulum body the whole setup was hung
w upside down and was made to swing. A least squares fit
with the measurements to
l
(Ib +Iw +mw l2 )b (t) = Cb b (t)+(mb lb +mw l)g sin b (t)
lb
gave Ib and Cb . The following table shows all the measured
and identified parameters of the prototype:
l 0.085 m
lb 0.075 m
mb 0.419 kg
mw 0.204 kg
Fig. 3. Illustration of the one dimensional prototype consisting of a square Ib 3.34 103 kg m2
plastic plate that holds the momentum exchange wheel through the motor Iw 0.57 103 kg m2
at its center. The plate is attached to a bearing at the bottom. Cb 1.02 103 kg m2 s1
Cw 0.05 103 kg m2 s1
Let b denote the tilt angle of the pendulum body and
w represent the rotational displacement of the momentum C. Braking mechanism
wheel with respect to the pendulum body. The non-linear Fig. 4 shows the current iteration of the braking mecha-
dynamics of the the setup shown in Fig. 3 is given by nism where an RC servo is used to quickly collide a metal
barrier with the bolt head attached to the momentum wheel.
(mb lb + mw l) g sin b Tm Cb b + Cw w
b = , The metal barrier and the RC servo are connected using a
Ib+ mw l2  thin metal sheet to ensure that most of the impact is taken


Ib + Iw + mw l2 Tm Cw w by the metal barrier. Furthermore, the design guarantees that
w = the metal barrier is easily replaceable. The RC servo-based
Iw (Ib + mw l2 )
braking mechanism had several advantages over the initial
(mb lb + mw l) g sin b Cb b solenoid-based braking mechanism in terms of weight (a
, (1)
(Ib + mw l2 ) reduction of 39g), power and durability.
where mb , mw are the pendulum body and wheel masses, To test the durability of the stopping mechanism, and to
Ib is moment of inertia of the pendulum body around the select the appropriate material for the metal barrier, several

3723
 and a free development environment for embedded systems
named ODeV [16], based on the Eclipse IDE, was used for
the software development.

Remote PC

Development/Debugging
UART JTAG
STM3210E

SPI1
Fig. 4. The CAD drawing of the RC servo-based braking mechanism: An IMU1 PPM RC Servo
RC servo is used to quickly collide a metal barrier (blue) with the bolt head
(red) attached to the momentum wheel. SPI2
IMU2

Encoder TIMER
jump up trials were done with additional steel plates attached
to the pendulum body such that the mass and the inertia CAN CAN
properties are equal to the full Cublis face to edge jump as
shown in Fig. 8. Maxon motor controller

D. Electronics and software

Fig. 5. The schematic diagram of the electronics setup
The power, computation and control components were not
mounted on the one dimensional prototype to avoid making
III. E STIMATION AND C ONTROL
the first prototype too complex. Figure 5 shows the overall
electronics setup except for the power, which was provided A. IMU-based tilt angle estimation and IMU calibration
by a constant voltage supply. To estimate the tilt angle b of the pendulum body, a purely
We selected the STM3210E evaluation board, (which accelerometer-based tilt estimation was implemented using
houses a Cortex-M3 clocked at 72 MHz) from STMicroelec- two accelerometers. Note that this method can be extended
tronics as the main controller for its rapid prototyping and [9] to the 3D tilt estimation of the Cubli.
readily available community support. The IMU comprised of Two accelerometers were placed along the diagonal of the
a 3-axis accelerometer, ADXL345 from Analog Devices, and pendulum body at distances ri , i = 1, 2 from the pivot point
a 3-axis rate-gyro using the IDG-500/ISZ-500 series from as shown in Fig. 6. The two most sensitive measurement axis
InvenSense. The two IMUs, mounted on the pendulum body b~exi , b~eyi were placed on the pendulum body plane with
as shown in Fig. 6, was connected to the evaluation board b exi pointing towards tangential and b~
~ eyi towards the radial
using two separate Serial Peripheral Interface (SPI) buses. direction. With this setup, the measured acceleration of the
A 50 W brushless DC motor, EC-45-flat, from Maxon accelerometer i, i = 1, 2, with respect to the accelerometer
Motor AG was selected to drive the momentum wheels for frame is given by
their high energy density as compared to the brushed DC := (ri b + g sin b , ri b2 g cos b , 0),
ai m i = 1, 2.
motors. Embedded hall sensors of the motor were used for
wheel speed, w , sensing. The motor was controlled using an The dynamic terms can be eliminated by
off-the-shelf digital four quadrant motor controller, EPOS2 a1 m a2 m = ((1 )g sin b , (1 )g cos b , 0),
50/5, from Maxon. The CANopen protocol was used for
=: (mx , my , 0),
the communication between the motor controller and the
evaluation board. Note that for the full Cubli, a miniature where = r1 /r2 and the estimated tilt angle of the
version of the above controller, DEC module 36/2, will be pendulum body is given by b := tan1 (mx /my ).
used. The mechanical placement errors of the accelerometers
The RC Servo, HSG-5084MG, of the braking mechanism were identified by the following procedure:
was driven by the PPM signals from the evaluation board. For Record the accelerometer measurements when the pen-
debugging purposes a high speed rotary magnetic encoder, dulum body is fixed (b = 0) at various tilt angles
RE36 [15], was used to measure b and was connected to b [/4, +/4].
the timing/counter unit of the evaluation board. Do principle component analysis to find the rotational
The STM32 port of the FreeRTOS scheduler was used in plane with respect to the accelerometer frame.
the software framework for its prioritized multitasking func- Calculate the orientation offset by projecting the mea-
tionalities provided by a scheduler, and for the small footprint surements into the rotational plane and defining a ref-
of the binary kernel image (4 kB). A completely open source erence frame.

3724
 where x := (b , b , w ),

0 1 0
b~
ey1 (mb lb +mw l)g Cb Cw
r1 A := Ib +mw l2 Ib +mw l2 Ib +mw l2 ,

b~
ex1 2

b~
ez1 (mIbbl+m
b +mw l)g
wl
2
Cb
Ib +mw l2 CwI((I b +Iw +mw l )
2
w (Ib +mw l )

b 0
Km
and B :=
Ib +mwl
2 .

Km (Ib +Iw +mw l2 )
r2 Iw (Ib +mw l2 )

b~
ey2 A 20 ms sampling time was selected for the digital control,
b~
ez2 considering the open loop unstable pole of (4) (637.4 ms)
b~
ex2 and a safety factor of 30. Using this sampling time the
continuous time system given by (4) was discretized using
zero-order hold and the resulting discrete time model is given
by
Fig. 6. Illustration of the tilt angle estimation setup using two accelerom- x[k + 1] = Ad x[k] + Bd u[k], k N0 , (5)
eters: Two accelerometers were placed along the diagonal of the pendulum
body at distances ri , i = 1, 2 and the ratio between the distances is used to where Ad and Bd are the discrete-time counterparts of the
eliminate the dynamic terms, ri i , ri i2 , i = 1, 2 from the measurements.
continuous time state matrix A and input matrix B. For the
sake of simplicity we use the same notation to represent the
continuous and discrete time versions of the state x and input
B. Jumping Up
u.
Assuming a perfectly inelastic collision i.e., a zero coeffi- Using the above discrete time model, a Linear Quadratic
cient of restitution, the angular velocity w of the momentum Regulator (LQR) feedback controller of the form
wheel required for jump-up is calculated using: (1) The

conservation of angular momentum during the impact u[k] = Kd (b [k], b [k], w [k]) (6)

Iw w = (Iw + Ib + mw l2 )b , was designed, where Kd = (Kd1 , Kd2 , Kd3 ) is the LQR

feedback gain that minimizes the cost function
where b is the angular velocity of the pendulum body after
X
the impact. (2) The conservation of energy after the impact J(u) := xT [k]Qx[k] + uT [k]Ru[k], Q > 0, R > 0,
(b = /4) and until the pendulum body reaches the top k=1
(b = 0)

b is the angular velocity estimate of the body from the
1 1
(Iw + Ib + mw l2 )b2 = (mb lb + mw l)g(1 ). rate gyro, and w is the angular velocity estimate of the
2 2
momentum wheel from the hall sensors. Fig. 9 illustrates
Eliminating b from both equations gives the balancing performance under (6). As it can be seen
from the plot, the estimated state of the system converges
(Iw + Ib + mw l2 )
2
w = (2 2) (mb lb + mw l)g. (3) to (b , b , w ) = (0.058, 0.0, 37.0). This behavior can be
Iw2 explained by the following property of the system:
The uncertainties in the parameters or a non zero coefficient Km
of restitution can make w deviate from the above calculated (I3 Ad + Bd Kd )1 Bd = (0, 0, ), (7)
Cw + Kd3 Km
value. In this case a simple bisection-based trial and error Cw + Kd3 Km 6= 0,
learning procedure can be applied to learn the required w ,
exploiting the monotic relationship between b achieved after where I3 is the identity matrix of size 3. The physical
the impact and w . The w given by (3) can be used as the interpretation of this property is that if Ad , Bd , and Kd gives
initial condition in this case. After the collision, once the a stable system, any constant offset in the measurement will
pendulum body has arrived near the equilibrium position, result in a non-zero wheel velocity at steady state. Using the
an LQR based controller, explained in the following section, readings from the absolute encoder attached to the pendulum
will start balancing the system. body at the pivot point, we localized the offset to the offset
in the tilt angle and redefined the tilt angle estimate as
C. Balancing control
b := tan1 (mx /my ) + d = b + d, d R.
Linearizing (1) around (b , b , w ) = (0, 0, 0) gives
To automatically eliminate the above constant tilt angle
x(t) = Ax(t) + Bu(t). (4) offset, the discrete time system given in (5) was extended

3725
by introducing a low-pass filter with state xf [k] and the
following dynamics:
xf [k + 1] = (1 )xf [k] + b [k], xf [0] = 0, (8)
where (0, 1) and the controller was modified to

u[k] = Kd (b [k] xf [k], b [k], w [k]). (9)
The closed loop system given by (5), (8) and (9) is stable
with = 0.02. Now, let ((b , b , w ) =: x, xf ) denote the
steady states of the system, which is the solution to
x = Ad x Bd Kd x Bd Kd1 (d xf )
xf = (1 )xf + (b + d).
This, with (8), gives
x = (I Ad + Bd Kd )1 Bd Kd1 (d xf )
Fig. 8. The picture shows another one dimensional prototype that was
0 built with a metal pendulum body with extra masses (on the left side of
= 0 (d xf ) (10) the pendulum body) to test the durability of the braking mechanism. With
Km Kd1 all the weights attached, it resembles the full Cublis flat to edge jump, the
Cw +Kd3 Km most critical jump for the Braking mechanism, in terms of mass and inertia.
xf = (1 )xf + (b + d). (11)

Finally, (10) and (11) gives (b , b , w , xf ) = (0, 0, 0, d)

showing that the system reaches the desired equilibrium state 0.1
and the filter state xf converges to the tilt angle offset d with 0
b

the proposed controller. Although the zero mean sensor noise 0.1
on the measurements were avoided in the formulation for the 0 5 10 15 20 25 30 35
sake of simplicity, the results are still valid under this type 10
of noise. 0
b

Fig. 10 shows a balancing experiment with the offset 10

20
correction. The estimated states (b , b , w ) converge to
(0.058, 0, 0) and the filter state xf converges to the offset 0 5 10 15 20 25 30 35 40
50
value 0.058 rad. Fig. 11 shows a balancing experiment
0
w

where the system was externally disturbed at time t = 26[s].

Furthermore, fig 12 shows the full jump-up and balancing

50
maneuver starting from rest at b = /4. 0 5 10 15 20 25 30 35 40
Time [s]

Fig. 9. Time traces of the one dimensional prototypes estimated

states during a balancing run with no offset correction on the body tilt

angle. The estimated state of the system converges to (b , b , w ) =
(0.058, 0.0, 37.0).

IV. C ONCLUSION AND F UTURE W ORK

This paper presented the Cubli, a unique 3D inverted

pendulum that jumps up using a self-generated impact and
balances using the exchange of angular momentum. The one
dimensional prototype has demonstrated various important
aspects of the system, including: the hardware, jump-up
using impact, accelerometer-based tilt estimation, control,
and software. The Cubli is currently undergoing its second
iteration with on-board electronics and power. Once success-
ful, three copies of the one dimensional prototype will be put
Fig. 7. A picture of the first one dimensional prototype while balancing together to produce the 3D version (See Fig. 1). These results
will be presented in future work.

3726
 V. ACKNOWLEDGEMENTS
0.1 The Authors would like to thank Sebastian Trimpe and
0
b

Raymond Oung for all the support and discussions sharing

0.1 their experience on building the Balancing Cube [9] and
0 5 10 15 20 25 30 35 the Distributed Flight Array [17]. Furthermore, the authors
10 would like to extend their gratitude to Philipp Reist for the
0 idea of the impact based jump-up strategy, Carolina Flores
b

10 for the graphics, and Tobias Widmer for helping with the
20 experiments.
0 5 10 15 20 25 30 35 40
50 R EFERENCES
0 [1] A. Stephenson, On a new type of dynamical stability, Proceedings:
w

Manchester Literary and Philosophical Society, vol. 52, pp. pp. 110,
1908.
50 [2] P. Reist and R. Tedrake, Simulation-based lqr-trees with input and
0 5 10 15 20 25 30 35 40
Time [s] state constraints, in Robotics and Automation (ICRA), 2010 IEEE
International Conference on, may 2010, pp. 5504 5510.
[3] D. Alonso, E. Paolini, and J. Moiola, Controlling an inverted pendu-
Fig. 10. Time traces of the one dimensional prototypes estimated states
lum with bounded controls, in Dynamics, Bifurcations, and Control,
during a balancing run with offset correction on the body tilt angle. The es-
ser. Lecture Notes in Control and Information Sciences. Springer
timated state of the system converges to (b , b , w ) = (0.058, 0.0, 0.0). Berlin / Heidelberg, 2002, vol. 273, pp. 316.
[4] M. V. Bartuccelli, G. Gentile, and K. V. Georgiou, On the stability of
the upside-down pendulum with damping, Proceedings: Mathemati-
cal, Physical and Engineering Sciences, vol. 458, no. 2018, pp. pp.
0.1 255269, 2002.
0 [5] J. Meyer, N. Delson, and R. de Callafon, Design, modeling and
b

stabilization of a moment exchange based inverted pendulum, in 15h

0.1 IFAC Symposium on System Identification, Saint-Malo, France, 2009,
0 5 10 15 20 25 30 35 pp. 462 467.
10 [6] B. Andrievsky, Global stabilization of the unstable reaction-
wheel pendulum, Automation and Remote Control, vol. 72, pp.
0 19811993, 2011. [Online]. Available: http://dx.doi.org/10.1134/
b

10 S0005117911090189
[7] A. Beznos, A. Grishin, A. Lenskij, D. Okhotsimskij, and A. For-
20 malskij, A flywheel use-based control for a pendulum with a fixed
0 5 10 15 20 25 30 35 40
50 suspension point, . , no. 1, pp. 2738, 2004.
[8] D. Bernstein, N. McClamroch, and A. Bloch, Development of air
0 spindle and triaxial air bearing testbeds for spacecraft dynamics
w

and control experiments, in American Control Conference, 2001.

Proceedings of the 2001, vol. 5, 2001, pp. 3967 3972 vol.5.
50 [9] S. Trimpe and R. DAndrea, Accelerometer-based tilt estimation of
0 5 10 15 20 25 30 35 a rigid body with only rotational degrees of freedom, in Robotics
Time [s]
and Automation (ICRA), 2010 IEEE International Conference on, May
2010, pp. 2630 2636.
Fig. 11. Time traces of the one dimensional prototypes estimated states [10] D. N. Nenchev and K. Yoshida, Impact analysis and post-impact
during a balancing run with offset correction on the body tilt angle and an motion control issues of a free-floating space robot subject to a
external disturbance at time t = 26 [s]. force impulse, IEEE Transactions on Robotics and Automation,
vol. 15, no. 3, pp. 548557, 1999. [Online]. Available: http:
//ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=768186
0.2 [11] T.-J. Tarn, Y. Wu, N. Xi, and A. Isidori, Force regulation and contact
0 transition control, Control Systems, IEEE, vol. 16, no. 1, pp. 32 40,
0.2 feb 1996.
b

0.4 [12] P. Pagilla and B. Yu, An experimental study of planar impact of

0.6
0.8 a robot manipulator, Mechatronics, IEEE/ASME Transactions on,
0 10 20 30 40 50 60 70 vol. 9, no. 1, pp. 123 128, march 2004.
10 [13] A. Konno, T. Myojin, T. Matsumoto, T. Tsujita, and M. Uchiyama,
An impact dynamics model and sequential optimization to
5 generate impact motions for a humanoid robot, Int. J. Rob. Res.,
b

vol. 30, no. 13, pp. 15961608, Nov. 2011. [Online]. Available:
0 http://dx.doi.org/10.1177/0278364911405870
0 10 20 30 40 50 60 70 [14] (2012, Feb) Maxon motor ag, maxon motor catalog, ec-45-flat.
40 [Online]. Available: http://www.maxonmotor.com/
[15] (2012, Feb) Rotary and linear motion sensors (rls), re36. [Online].
20 Available: http://www.rls.si/document/RE36D01.pdf
w

[16] S. Oliveri. (2012, Feb) Open development environment for

20 embedded application. [Online]. Available: http://www.stf12.org/
0 10 20 30 40 50 60 70 developers/ODeV.html
Time [s] [17] R. Oung and R. DAndrea, The distributed flight array, Mechatronics,
vol. 21, no. 6, pp. 908 917, 2011. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S095741581000139X
Fig. 12. Time traces of the one dimensional prototypes estimated states
during a jump-up and balance maneuver.

3727