8th IFAC Symposium on Nonlinear Control Systems

University of Bologna, Italy, September 1-3, 2010

Traversing from point-to-point along a

straight line with a ballbot
Pedro La Hera ∗ Leonid Freidovich ∗ Anton Shiriaev ∗,∗∗
Uwe Mettin ∗∗
∗
Department of Applied Physics and Electronics
Umeå University, SE-901 87, Umeå, SWEDEN,
{Xavier.LaHera|Leonid.Freidovich|Anton.Shiriaev|Uwe.Mettin}@tfe.umu.se.
∗∗
Department of Engineering Cybernetics
Norwegian University of Science and Technology, NO-7491 Trondheim,
NORWAY.

Abstract:
The planar motion to traverse from point to point with a ballbot is discussed. Our aim is
to propose an analytical method to plan motions that can be robustly executed by feedback
control. The approach is based on defining geometric virtual relations among the generalized
coordinates. These relations form the basis of a reparametrization of the system dynamics for
motion planning, and an introduction of coordinates transverse to the desired trajectories useful
for controller design. Results of numerical simulation are presented to illustrate the performance
and robustness of the closed-loop system.

Keywords: Underactuated mechanical systems, virtual holonomic constraints, motion planning.

1. INTRODUCTION “To maintain balance when still, the ballbot must keep its
center of gravity directly over its center of support (step
1). During movement, the ballbot manipulates its center
Ballbots have been introduced as a novel type of loco- of gravity to best effect. To go from one point to another
motive systems, with the purpose to introduce dynamic on level ground, for example, the drive ball first rotates
stability in mobile robots (Lauwers et al. [2005]). Their slightly in the direction opposite to the intended direction
mechanics intend to simplify the complexity of wheeled of travel, which tilts the body forward a bit to initiate the
robots by letting a tall rigid cylindrical body to balance move (step 2). Next, the ball spins in the direction of
on a single omnidirectional spherical ball. This would in motion to accelerate ahead (step 3). While the ballbot is
principle allow to execute more agile motions to ease the at constant velocity, the body must remain nearly vertical
navigability of the robots around complex flat ground (step 4). The opposite actions must occur to decelerate the
environments. machine and then prepare it to halt (step 5), which together
From the theoretical point of view, this system belongs bring it to a stop (step 6). When traversing inclines, the
to the class of underactuated mechanical systems, where body must lean into slopes to keep its equilibrium.”
the degree of underactuation depends on the task to be
performed. The basic scenarios to cope with this systems
are: dynamic balancing around the unstable equilibrium
and traversing from point-to-point in the vertical plane
(straight line) and when the body behaves as a spherical
pendulum (3D motions).
Balancing around upright equilibrium is a problem that
has been theoretically and practically considered in Na-
garajan et al. [2009b], Lauwers et al. [2005]. However,
locomotion is a problem with no straightforward solution,
and it remains challenging due to the underactuted nature
of the system dynamics. A publication which is relevant
to this work is Nagarajan et al. [2009a], where a solution
based on dynamic programming is presented to design mo-
tions from point to point in the vertical plane. The model
considered is planar and uses simplifying assumptions to Fig. 1. Traveling from point-to-point.
a model of underactuation degree one, in order to solve
the particular problem postulated by the ballbots creator Our aim is to propose an analytical motion planning
Hollis [2006], which main concept is cited below (refer to technique for a similar motion. To this end we consider the
Fig. 1): application of the virtual holonomic constraints approach

978-3-902661-80-7/10/$20.00 © 2010 IFAC 125 10.3182/20100901-3-IT-2016.00185

NOLCOS 2010
Bologna, Italy, September 1-3, 2010

(Westervelt et al. [2007], Freidovich et al. [2008]). We the desired motion is embedded into half of a periodic
provide the design steps for both motion planning and cycle as illustrated in Fig. 3. The motion is executed by
controller design for vertical plane motions, which can switching appropriately between the periodic trajectories
be readily extended to 3D motions, where underactuation and the balancing controller that stops the motion.
higher than one is to be faced (Shiriaev et al. [2010]).
The rest of the paper is organized as follows. The problem
is formulated in Section 2. The main contribution in
motion planning is presented in Section 3, giving an
example of application. The step by step procedure to
derive the controller is described in Section 4. Different
simulation results are presented in Section 5. Concluding
remarks are given in Section 6.

2. PROBLEM FORMULATION

Fig. 3. Trajectories required to achieve a point-to-point

motion.
To design these motions we first observe that a particular
enforced solution of the system dynamics (1), can be
represented by
qa = qa∗ (t), qu = qu∗ (t), t ∈ [0, T ]. (2)

However, there exist always a geometrical reparametriza-

tion of the same path without an explicit time dependence.
In order to demonstrate this, let us suppose that there
Fig. 2. Planar simplified model. exist two C 2 -smooth functions φ1 (·) and φ2 (·), such that

The dynamics of the ballbot in the vertical plane can be qa = φ1 (θ),

modeled as an underactuated mechanical system with two
degrees of freedom, see Fig. 2. Under the assumption of qu = φ2 (θ), (3)
non-sliding ground contact, the Euler-Lagrange equations where θ = θ∗ is some independent scalar variable used
yield the following dynamics, to reparameterize the motion (2). The identities in (3)
 
u are known as virtual holonomic constraints, and express
M (q)q̈ + C(q, q̇)q̇ + G(q) = , (1) geometric relations among the generalized coordinates
0
along a motion. These constraints are forced into dynamics
where u is the external torque that allows to control the of the robot by feedback action.
ball’s rotation, q = [qa , qu ] with qa being the ball’s angular The dynamics of the system in terms of the new variable
position and qu the body angle with respect to the vertical θ can be found by substituting (3), with their first and
axis. M (q) is a positive definite matrix of inertias, G(q) is second time derivatives
the gravity matrix, C(q, q̇) is a matrix representing Coriolis
and centrifugal forces. The components of these matrices, d ′ d
dt qdof (t)=φi (θ(t)) dt θ(t),
as well as the numerical values of the physical parameters d 2
d2 ′′ 2
are presented in Appendix A. dt2 qdof (t)=φi (θ(t)) + φ′i (θ(t)) dt
dt θ(t)
d
2 θ(t),

The motion depicted in Fig. 1 can in principle be thought where dof = [qa , qu ], i = [1, 2], into the the non actuated
as composed by three separated parts: differential equation in the second row of (1), yielding the
second-order differential equation
(1) initialization, which is represented by steps 1 and 2,
α(θ, φi , φ′i )θ̈ + β(θ, φi , φ′i , φ′′i )θ̇2 + γ(θ, φi ) = 0. (4)
(2) the forwarding (or backwarding) motion, visualized
from steps 2 to 5, and The set of achievable motions of the robot is defined
(3) the halting, which consists of stabilizing the robot by the solutions [θ∗ (t), θ̇∗ (t)] of the differential equation
around upright equilibrium once the forwarding mo-
above, initiated at [θ0 , θ̇0 ]. From this point of view, the
tion has been executed.
motion planning is reformulated as the task of defining the
Our main idea is to design periodic trajectories for the constraint relations φi , which will result in an appropriate
first two parts, and a stabilizing controller to halt. In desired motion. If such motion is found, the systematic
this form we introduced bounded motions, for which the controller design steps introduced in Shiriaev and Frei-
design of feedback controllers can be done by following a dovich [2009] can be followed to achieve orbital exponential
systematic procedure (Shiriaev et al. [2005]). As a result, stability.

126
NOLCOS 2010
Bologna, Italy, September 1-3, 2010

3. MOTION PLANNING (C) Suppose ω > 0, then the dynamical system (6) has
periodic solutions if for some a1 , a2 the integrals
 θ 
Following the theoretical results presented in Freidovich  Z ∗ β(τ ) 
et al. [2008], this section introduces the sufficient condi- 2 dτ
tions that guarantee existence of periodic solutions for the Zθ∗  α(τ ) 
ai 2γ(s)
system (4). Furthermore, one type of constraint function e · ds, (9)
α(s)
is proposed to exemplify the motion planning steps. ai

are equal to zeros. Furthermore, these curves can be

3.1 Conditions for existence of periodic solutions found as solutions of (6) with initial conditions
θ(0) = a1 , θ̇(0) = 0 and θ(0) = a2 , θ̇(0) = 0. (10)
If we consider the following geometric relations Proof 1. Statement (A) is obvious. Statement (B) is
proved in Shiriaev et al. [2005]. To validate (C) we observe
qa = φ1 (θ) = θ, that the system (6) is integrable (Shiriaev et al. [2005]),
qu = φ2 (θ) = φ(θ), (5) and that the integral function I(θ(t), θ̇(t), θ0 , θ̇0 ), defined
by
such that the pendulum motion is parametrized according  
to the ball’s position, the second order reduced dynamics Zθ
2γ(s)
(4) takes the form: I = θ̇2 (t) − Ψ(θ) θ̇02 − Ψ(θ) ds, (11)
α(s)
θ0
α(θ, φ, φ′ )θ̈ + β(θ, φ, φ′ , φ′′ )θ̇2 + γ(θ, φ) = 0, (6)
with
with 
 Zθ β(τ ) 


Ψ(θ) = exp −2 dτ , (12)

α(·) = (2b cos φ + a + c) φ′ + b cos φ + a,  α(τ ) 
θ0
β(·) = (2b cos φ + a + c) φ′′ − b sin φφ′2 , (7)
b · g sin φ remains zero on the solution θ(t) of (6) initiated at [θ0 , θ̇0 ].
γ(·) = − , If we assume that a curve θ∗ (t) exist, then by evaluating
r
(11) at θ̇ = 0, we obtain the expression (9).
for which the following conditions hold (Shiriaev et al.
[2005]). 3.2 Planning the forwarding motion

To exemplify this approach, let us consider the constraint

function
φ(θ) = K · arctan(θ + θm ), (13)
such that the angle of inclination of the body are limited
by the constant K, irrespective of the ball’s motion.With
this choice the coefficient α(θ) > 0 for all θ ∈ [−2π, 2π].
Substituting (7) in the expression (8) results in
− cos φ · φ′
> 0, (14)
(2b cos φ + a + c) φ′ + b cos φ + a
which analyzed at the equilibrium θ = −θm , allows to
compute the values of K for which periodic solutions exist
according to the following inequality
Fig. 4. Desired periodic motion for the ball. −b − a
< K < 0, (15)
2b + a + c
Lemma 1. Let φ(θ) be a C 2 -smooth function such that
α(θ, φ, φ′ ) is separated from zero, i.e. either positive or numerically equivalent to −0.0629 < K < 0, which implies
negative. Then (refer to Fig. 4): a maximum inclination of the upper body of approxi-
mately 5.66 degrees. A visualization of the effect of this
(A) The equilibria of the dynamical system (6), i.e. the coefficient is given in Fig. 5, showing the motion depen-
solutions of the equation γ(θ, φ) = 0, are located at some dence on this value. An important issue in the design is
middle range of the motion θm , with φ(θm ) = 0 (step 4 related to the motor’s maximum torque |umax | = 29Nm 1 .
in Fig. 1). Therefore, an optimization problem can be postulated to
(B) Considering the equilibrium of the dynamical system find the value K, that allows to execute the motion in some
(6), if the constant prescribed time without exceeding the torque limitations.
γ ′ (θ, φ) However, this is not considered here, since by the choice
ω= , (8) of −0.035 < K < 0 the requirement on available torque is
α(θ, φ, φ′ ) θ=θm ,φ=0 within the allowed interval.
is positive, then this equilibrium is a center. If, in 1 This limitation is considered for the ballbot platform of Carnegie

opposite, ω is negative, then this equilibrium is a saddle. Mellon University.

127
NOLCOS 2010
Bologna, Italy, September 1-3, 2010

Phase portrait of Zero Dynamics Periodic Trajectories Phase portrait of Zero Dynamics Periodic Trajectories
15 10 4 1.4

K = −0.04 K = −0.04 3 1.2

8
2
1
10 K = −0.008
6 1
0.8

θ [rad]

θ [rad]
dθ / dt

dθ / dt
0
0.6
4 −1
5
0.4
−2
2
−3 0.2
K = −0.008
0 0 −4 0
0 2 4 6 8 10 0 1 2 3 4 5 6 7 8 0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 1 2 3 4 5
theta[rad] time[sec] theta[rad] time[sec]

Constraint function Nominal Torque Constraint function Nominal Torque

4 50 2 5
K = −0.04 K = −0.04
3 1.5

2 1
qu [degrees]

qu [degrees]
0.5
τ [Nm]

τ [Nm]
0 0 0 0
K = −0.008 K = −0.008
−1 −0.5

−2 −1

−3 −1.5

−4 −50 −2 −5
0 2 4 6 8 10 0 2 4 6 8 10 0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 0.2 0.4 0.6 0.8 1 1.2 1.4
theta [rad] theta [rad] theta [rad] theta [rad]

Fig. 5. The motion of the ball for 1[m] distance (the rate Fig. 6. The motion of the ball for 0.12[m] distance (os-
between ball angular rotation and horizontal traveled cillations with an amplitude of 0.06[m], i.e. θmini =
distance is x = qa · r). Here we show half of a periodic −0.6[rad]) for initialization. Top right: time evolution
trajectory with initial conditions [θ(0), θ̇(0)] = [0, 0]. of θ. Bottom left: body evolution as function of the
Top left: Phase portrait solution of the reduced dy- ball’s rotation θ. Bottom right: nominal input torque
namics (6). Top right: time evolution of θ∗ (t). Bottom for the different trajectories
left: body motion qu as function of the ball’s rotation
θ. Bottom right: nominal input torque for the different
8
trajectories. Switching point
6

3.3 Planning the initialization of motion

4

The same analysis is applied for initialization; however, the Forwarding motion

value of K is chosen according to the following equality, 2

Starting position
Dtheta [rad/s]

K · arctan(θm ) 0

Kini = , (16)
arctan(θmini ) cycle for initialization stoping point
−2

such that both motions connect at [θ, θ̇] = [0, 0]. −4

The value of θmini is chosen so that we achieve small −6

oscillations around upright equilibrium considering torque

limitations. As example, in Fig. 6 we show different solu- −8
0 1 2 3 4 5 6 7 8 9 10

tions of the reduced dynamics for different values of K. theta [rad]

The final periodic solutions chosen for stabilization can be

visualized in Fig. 7, where the trajectories are chosen for Fig. 7. Final trajectories for the ball’s motion.
a value of K = −0.0222 for the forwarding motion, and
Kini = −0.0565 for initialization. The period of oscillation Below we briefly present the general steps in constructing
for the main motion is T = 7.86[sec]. Both trajectories a transverse linearization, and then show how it is used
share a common point at θ = 0, θ̇ = 0, where the switch for controller design.
is intended to be applied. The robot is assumed to start
at a static configuration located at θ = 0.6. The body 4.1 Changing the Generalized Coordinates
movement during both stages is also shown in Fig. 8.
Given the scalar function φ(·), we can introduce new
generalized coordinates for (1) in a vicinity of the motion
4. CONTROLLER DESIGN
as:
For the synthesis of a controller, the technique recently θ, y = qu − φ(θ), (17)
proposed in Shiriaev and Freidovich [2009] is applied. This
design relies on concepts of a moving Poincaré section and with the derivatives
a linearization of the dynamics transverse to a desired
orbit. This linearization in a vicinity of a cycle can be θ̇ , ẏ = q̇u − φ′ (θ)θ̇, (18)
efficiently constructed for forced motions of mechanical  
systems and further used for controller design. θ̈ , ÿ = q̈u − φ′′ (θ)θ̇2 + φ′ (θ)θ̈ . (19)

128
NOLCOS 2010
Bologna, Italy, September 1-3, 2010

2
elements are bounded continuous for all τ ≥ 0. The goal
is to find a C 1 -smooth vector of gains K(τ ), such that the
1.5
control input
1

û⊥ (τ ) = K(τ )x̂⊥ (τ ), (27)

0.5
qu [degrees]

0
Body motion during forwarding
stabilizes the equilibrium of the linear auxiliary system
−0.5 (26).
−1 Theorem 1. A controller that locally achieves a similar
−1.5
Body motion during initialization property for the trajectory (20) on the nonlinear system
−2
(24) is
0 1 2 3 4 5 6 7 8 9 10
theta [rad]
u⊥ (t) = K(s)x⊥ (t), s = s(θ(t)), (28)

Fig. 8. Constraint functions for both initization and mov- where x⊥ is the vector of transverse coordinates defined
ing forward motions. by (25), and s(·) is an index parameterizing the particular
Observe that as long as a control action makes the desired leaf of the moving Poincaré section, to which the vector
trajectory invariant and the initial conditions are on the x⊥ belongs at time moments t (see Shiriaev and Freidovich
target motion, we have that: [2009]).

θ = θ∗ (t), θ̇ = θ̇∗ (t), y = 0, ẏ = 0. (20)

5. SIMULATION RESULTS
Dynamics in the new coordinates can be found by intro-
ducing the expression from (17), (18) and (19) into dynam- Following the control design steps two different auxiliary
ics of the robot (1). The dynamics of the y-variable can be systems (26) are found for the initialization and the
then computed and denoted in short form as follows: forwarding motion. Two sets of controllers in the form (27)
are required to stabilize separately each cycle. A switching
ÿ = R(y, θ, ẏ, θ̇) + N (y, θ)u. (21) between controllers is applied in the vicinity of [θ, θ̇] =
[0, 0]. In Fig. 9 the simulation results of this scenario shows
Proposition 2. Feedback Transformation - There exist the the convergence of the robot to the trajectories planned
feedback transformation (Shiriaev et al. [2005]) in the form by appropriately combining the controllers. Notice that
u = N (q)−1 [u⊥ − R(q, q̇)] , (22) we consider the robot to start at an offset of 0.6[rad],
which is the equilibrium of the initialization cycle, but
well defined in the vicinity of the motion, such that the it is also a wrong initial condition of the entire motion.
dynamics (1) can be rewritten in the new coordinates (17) However, the controller is able to handle this disturbance
as follows and exponentially converge to the desired motion. The
transverse coordinates given by (25) are also shown in Fig.
α(θ)θ̈ + β(θ)θ2 + γ(θ) = g(θ, θ̇, y, ẏ, ÿ), (23) (10) to visualize this effect. The time evolution of these
coordinates allow us to verify the exponential stability of
ÿ = u⊥ , (24) the motion.
where g(·) is a smooth function that is equal to zero in the
desired orbit, and the left hand side of (23) matches (6). 8
phase portrait
2.5
Body trajectory

6 2

1.5
4.2 Tranverse Linearization 4
1
2
q [rad]
Dtheta

0.5
0
u

0
The dynamical system (24), (23) possesses a natural choice −2
−0.5

of (2n − 1) - transverse coordinates −4

−6
−1

−1.5

x⊥ = [I(θ, θ̇, θ∗ (0), θ̇∗ (0)), y, ẏ]T , (25) −8

0 2 4
theta [rad]
6 8 10
−2
0 2 4
θ [rad]
6 8 10

where I(·) is defined in (11), and represents a measure of 10

ball trajectory
10
Input Torque

the Euclidean distance to the desired trajectory. Lineariza- 8 5

tion of (24) and (23) along (20) can be represented by the

linear system 6 0
tau(Nm)
theta

4 −5
d
x̂⊥ (τ ) = A(τ )x̂⊥ (τ ) + B(τ )û⊥ (τ ), τ ∈ [0, T ] (26) 2 −10
dτ
0 −15
0 5 10 15 20 0 5 10 15 20

where the matrices A(τ ) and B(τ ) are computed following time [sec] time (sec)

the generic procedure proposed in Shiriaev and Freidovich

[2009].
Fig. 9. Simulation results for initialization and forwarding
4.3 Design of a feedback controller motion. Top left: Phase portrait. Top right: body
motions as a function of the ball’s rotation. Bottom
left: ball’s trajectory as a function of time. Bottom
Consider the time-varying control system (26) with initial
right: torque.
conditions x̂⊥ (0) = x̂0⊥ . A(τ ) and B(τ ) are matrices whose

129
NOLCOS 2010
Bologna, Italy, September 1-3, 2010

Ralph Hollis. Ballbots. Scientific American Magazine,

Tranverse dynamics
6
2006.
4

2
Tom Lauwers, George Kantor, and Ralph Hollis. One is
enough! In 12th Int’l Symp. on Robotics Research, San
I(t)

−2 Francisco, Oct. 12 - 15 2005.

−4
0 2 4 6 8 10 12 14 16 18 20 Umashankar Nagarajan, George Kantor, and Ralph L.
0.01
Hollis. Trajectory planning and control of an underactu-
0 ated dynamically stable single spherical wheeled mobile
robot. In IEEE International Conference of Robotics
y(t)

−0.01

−0.02 and Automation, pages 3743–3748, Kobe, Japan, May

−0.03
0 2 4 6 8 10 12 14 16 18 20
12 - 17 2009a.
Umashankar Nagarajan, Anish Mampetta, George A.
0.2

0.15
Kantor, and Ralph L. Hollis. State transition, balanc-
0.1 ing, station keeping, and yaw control for a dynamically
Dy(t)

0.05
stable single spherical wheel mobile robot. In IEEE
−0.05
0
International Conference of Robotics and Automation,
0 2 4 6 8 10
time(sec)
12 14 16 18 20
pages 998 – 1003, Kobe, Japan, May 12 - 17 2009b.
A.S. Shiriaev and L.B. Freidovich. Transverse linearization
for impulsive mechanical systems with one passive link.
Fig. 10. Exponential convergence of the transverse coordi- IEEE Trans. on Automatic Control, 54(12):2882–2888,
nates. 2009.
6. CONCLUSIONS A.S. Shiriaev, J.W. Perram, and C. Canudas-de-Wit. Con-
structive tool for orbital stabilization of underactuated
From the motion planning perspective, we have suggested nonlinear systems: Virtual constraints approach. IEEE
a new procedure for designing point-to-point motions in Transactions on Automatic Control, 50(8):1164–1176,
the case of a ballbot. This is done by a reparametrization 2005.
of the system according to a geometric relation between A.S. Shiriaev, L.B. Freidovich, and S.V. Gusev. Transverse
the body and the ball. Solutions of the reduce dynamics linearization for controlled mechanical systems with
resulting from this reparametrization represent the family several passive degrees of freedom. IEEE Transactions
of motions achievable for the particular constraint func- on Automatic Control, 55(4), 2010.
tion. E.R. Westervelt, J.W. Grizzle, C. Chevallereau, J.H. Choi,
and B. Morris. Feedback Control of Dynamic Bipedal
To control any of the desired motions, we have suggested Robot Locomotion. CRC Press, Taylor and Francis
to introduce coordinates transverse to the desired cycle. Group, 2007.
Stabilization of the dynamics transverse to the desired
trajectory, together with a partial feedback linearization, Appendix A. MATRICES FOR DYNAMICS (1)
allows to ensure the stability of the specified motion. Due
to the property of orbital exponential stability of the The matrices of (1) are given below:
controller, it is important to realize that the design can be  
reduced to a forwarding motion and its controller. How- a a + b cos(qu )
M= ,
ever, we avoid this design, since in practical applications it a + b cos(qu ) a + c + 2b cos(qu )
would be important to have control over the initialization  
of the motion. −b sin(qu )q̇u2
C= ,
An example of a constraint relation has been presented −b sin(qu )q̇u2
to illustrate the method, and the simulation results were " #
0
shown to verify its performance. However, equivalent re-
G= bg sin(qu ) .
sults with a bigger range of body inclination angle can be −
found by means of different other constraint equations, e.g. r
linear constraint. Here, we restricted the analysis to one The parameters a = Iball + (mball + mbody )r2 , b = mbody rl,
example. c = Ibody + mbody l2 , and the physical values Iball , mball ,
mbody , Ibody , r, l are taken from Nagarajan et al. [2009a].
ACKNOWLEDGEMENTS
Symbol Parameter Value(unit)
The authors would like to acknowledge Umashankar Na- mbody Body mass 51.663(kg)
garajan, from the Robotics Institute of Carnegie Mellon mball Ball mass 2.437(kg)
University, for his collaboration providing the required r Ball radius 0.1058(m)
information about the Ballbot Research Platform. l Body center of mass 0.69(m)
Ibody Body Inertia 37.1873(kgm2 )
REFERENCES Iball Ball Inertia 0.0174(kgm2 )
Table A.1. Model Parameters
L.B. Freidovich, A. Robertsson, A.S. Shiriaev, and R. Jo-
hansson. Periodic motions of the Pendubot via virtual
holonomic constraints: Theory and experiments. Auto-
matica, 44(3):785–791, 2008.

130