Received February 20, 2019, accepted February 24, 2019, date of publication February 28, 2019, date of current

version April 5, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2902219

Kinematic Analysis and Control Algorithm

for the Ballbot
CHENGTAO CAI1 , JIAXIN LU 1, AND ZUOYONG LI 2,3
1 College of Automation, Harbin Engineering University, Harbin 150001, China
2 Fujian Provincial Key Laboratory of Information Processing and Intelligent Control, Minjiang University, Fuzhou 350121, China
3 Institute of Internet Innovation, Fujian College’s Research Base of Humanities and Social Science, Minjiang University, Fuzhou 350121, China

Corresponding author: Jiaxin Lu ([email protected])

This work was supported in part by the National Natural Science Foundation of China under Grant 61673129 and Grant 51674109.

ABSTRACT The ballbot is a dynamically stable mobile robot designed to balance on a single ball, whose
dynamic stability enables improved navigability in narrow, crowded, and dynamic environments. Through its
single contact point with the ground, ballbot is omnidirectional and exceptionally agile, maneuverable, and
organic in motion compared to other ground vehicles. For dealing with the challenging and imperative issues
about ballbot, such as balancing, yaw, position control algorithms as well as the mathematical kinematic
model, a novel model employing the Lagrange Equation is derived and control algorithm based on the similar
principle as the second-order inverted pendulum is proposed, which is used for tackling the balancing, yaw,
and position control. A cascade fuzzy proportional derivative (PD) controller and another controller with
proportional integral (PI) control and PD feedback are designed for position and speed, respectively. Yaw
control is presented, including head-hold mode and head-free mode. Some experiments are carried out to
validate the effectiveness of the mathematical model and control algorithm.

INDEX TERMS Ballbot, Lagrange equation, cascade fuzzy control.

I. INTRODUCTION the vertical axis. It used linear quadratic regulator (LQR)

With the rapid development of robotics, all kinds of robots controller for full state feedback [9]. ’BallIP’ was developed
are gradually entering people’s lives, especially dynamic bal- in 2008, the first ballbot used the structure of three Omnidi-
ancing robot [1]. One of the most popular dynamic balancing rectional wheels. It achieves rotation around the vertical axis,
robots is the two-wheeled Segway Balancing robot [2]. As a which is a breakthrough in the ballbot. However, the paper
Vehicle, one-wheeled Balancing robot also won a huge mar- does not describe the modeling method in detail. In terms of
ket. However, these balancing robots have turn limits, only control algorithms, ‘BallIP’ used classical PD controller [10].
move forward and backward, human can’t move in accor- Ballbot with a manipulator has been proposed by Asgari et al
dance with owns wishes, in other words, these robots can only in 2013. It has been proved that the dynamics equations of the
do one dimensional motion [3], [4]. The ballbot improves the assumed mobile robot. At the same time, a control algorithm
mobile efficiency result of moving in any direction without is proposed to realize the stable motion control of the system.
a radius of rotation. It is supported by a ball that move in all Finally, the execution of the simulation program is to verify
directions, which means that it is not necessary to turn around the advantages of the algorithm [11]. Aiming at the ballbot
like a wheeled or multi-legged robot to change the way is an underactuated, nonholonomically constrained system
forward. Moving omnidirectionality and the characteristics of (there are more degrees of freedom than independent control
single contact point with the ground makes it more suitable inputs), Aykut et al presented the linearized equations of
for navigating in the limited space [5]–[8]. motion and the controllability analysis is carried out. They
In 2005, the first ballbot was invented at Carnegie Mellon consider that ballbot was a linear system except for rotating
University, which used the inverse mouse-ball drive mech- about the vertical axis of the ballbot. They provided a PD
anism to drive the ball to achieve a dynamic balance, but controller and presented simulation results. Unfortunately,
this structure is more cumbersome, and cannot rotate along they did not implement the algorithm on a real robot [12].
There is some recent work on the application of the ball robot
The associate editor coordinating the review of this manuscript and in a specific environment, for example, paper focuses on
approving it for publication was Zhen Li. the control methods of the ball robot climbing and proposes
2169-3536
2019 IEEE. Translations and content mining are permitted for academic research only.
38314 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 7, 2019
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
C. Cai et al.: Kinematic Analysis and Control Algorithm for the Ballbot

ballbot with an annular support leg, which can keep statically

stable when powered off [13]. Paper presents the equations
of motion for the ballbot system on a sloped surface with a
center-of-mass offset [14].
In this paper, we propose the control method and analyze
the kinetics of the ballbot. Firstly, we derive the Lagrange
equation of the plane model of ballbot and the mathemat-
ical relationship between the speed of the omnidirectional
wheel and the speed of the plane model. The only con-
tact point between the ballbot and the ground is the ball,
thus it only relies on the inertial force of the body to
maintain the dynamic stability. Cascade fuzzy PD control FIGURE 2. Structure of system.

used for dynamic balancing in high angle non-linear areas.

Head-free model and head-hold model are presented for yaw
controller with little attention in previous work. Outer-loop we set the measurable range of the gyroscope to be 2000◦ /s,
position controller and inner-loop speed controller outputs and the measurable range of acceleration to be 8g. Triaxial
desired attitude angle, which couples balancing and posi- magnetic field intensity is gauged by three axis magnetometer
tion control. The key is to achieve position and yaw con- HMC5883L measurement. Event-triggered sampling method
trol while maintaining its own balance. Finally, the ballbot is used to save the power energy [15]. The measurement
is implemented shown in Fig. 1 and the experiments are ranges from milli Gauss to 8Gauss, then the Euler angle
carried out, including self-balancing, station-keeping, anti- of the robot is calculated by quaternion algorithm. To get
interference and position-moving. The paper is organized as a more accurate Euler angle, we used a sliding average fil-
follows: section II describes the ballbot system construction ter for angle velocity. Roll and pitch are input of the bal-
and dynamic model, section III introduces the ballbot control ancing controller, because the ballbot body balances on the
method, including balancing control and position control, ball dynamically. the yaw angle used to achieve the ballbot
in addition, kinetics of ballbot was analyzed. Section IV body rotate around a vertical axis. At the same time, ZigBee
shows the result and the Experimental results are analyzed. is used for wireless data transmission between ballbot and
Section V concludes the paper. computer, including real-time sensor data, system configu-
ration parameters. We select the direct current gear motor
to provide enough torque and the Motor reduction ratio is
1: 16. Robomasters RM35 motor is a dedicated 5 ∼ 20kg
robot customized power motor. The encoder feedback the
motor speed integrated in the back end of motor. The three
motors are 120◦ on the xoy plane, and the angle between
the ball and plane of omnidirectional wheel α is 45◦ , if α
is too large, the ballbot body is easy to fall; on the contrary,
the mobile robot will be subject to certain restrictions. The
omnidirectional wheel is used on this robot. Two alternating
driven wheels ensure that the omnidirectional wheel has no
sliding friction when moving parallel to the center axis, and
there is no twitching phenomenon of the ballbot body. The
FIGURE 1. Ballbot balancing. diameter of the omnidirectional wheel is 100mm and the
weight are 290g [16].

II. SYSTEM DESCRIPTION AND MODELING B. DYNAMIC MODELING

A. SYSTEM DESCRIPTION The dynamic analysis of the system used the Lagrange equa-
Fig. 2 shows the structure of the system. The ballbot is divided tion which is an important method for dynamic modeling.
into two parts: ball and body. The ballbot body included It simplifies the modeling process obviously when the degree
processor, sensor, driver, motor, omnidirectional wheel. The of freedom of the system is less, because it only needs
processor adopts Freescale Kinetis K60 series MCU, whose to analyze the system kinetic energy, potential energy and
hardware interfaces is wealthy and operating speed can be generalized force, rather than the complicated force analy-
overclocked up to 150MHz. Robot measures the triaxial angle sis [17], [18]. To facilitate the analysis, the ballbot can be
velocity and acceleration of the ballbot body through the regarded as a uniform rigid body and a uniform rigid ball,
Inertial measurement unit (IMU) MPU-6050. The measure- and we decompose the three-dimensional dynamic system
ment range can be controlled by a program write register. into three planes. It is worth noting that modeling is based
In order to adapt to the fast-dynamic movement of the ballbot, on the following two assumptions:

VOLUME 7, 2019 38315

 C. Cai et al.: Kinematic Analysis and Control Algorithm for the Ballbot

(i)The mechanical structure is completely symmetric, and of the motor, so the kinetic energy of the system can ignore the
the motion of the xoz plane and the yoz plane is uncoupled. kinetic energy generated by the rotation of the drive wheel.
(ii)There is rolling, no sliding and spin between the sup- Using the Euler Lagrange equations, the dynamic equations
porting ball and the ground. of the planar motion of the ballbot can be written in matrix
form, as follows.
 
0
M (q)q̈ + C(q, q̇) + G(q) + D(q̇) = (7)
τ
Define the system configuration vector q = [θ, ψ]T , M (q)
is the mass matrix, (q, q̇) is the Coriolis force vector. G(q) rep-
resents gravity vector, D(q̇) is the friction torque, τ is the
torque between the body and the support ball in the direction
normal to the plane. The expressions of the above terms are
as follows.
 
M11 M12
M (q) = (8)
M21 M22
M11 = Ib + IB + mb rb2 + mB (rb2 + l 2 ) (9)
M12 = M11 + mB rb l cos(θ + ψ) (10)
FIGURE 3. The plane model of the ballbot.
M21 = M12 (11)
The ballbot has 5 degrees of freedom, including 2 degrees M22 = M11 + 2mB rb l cos(θ + ψ) + mB l + IB (12) 2

−mB rb l sin(θ + ψ)(θ̇ + ψ̇)2

 
of freedom in the plane position and 3 degrees of freedom C(q, q̇) = (13)
of the Euler angle. The plane model is shown in the Fig. 3. −mB rb l sin(θ + ψ)(θ̇ + ψ̇)2
θx,y,z , ψx,y,z represent the rotate angle of the ballbot body and
" #
0
the supporting ball around each axis. The xoz plane is the D(q̇) = gmB rb l sin(θ +ψ) (14)
− I +m l 2
same as the yoz plane model. The kinetic energy Kb and mass B B

energy Vb of the supporting ball are given below.

C. SPEED CONVERSION BETWEEN WHEEL AND BODY
2 The ballbot’s movement is achieved through the rotation of
1 1
Kb,XOZ = Ib ψ̇y + mb (rb ψ̇y )2 (1) the omnidirectional wheel that drives the ball to roll. The
2
| {z } | 2 {z } relationship between the speed of the robot and the rotational
Rotation Translation speed of the omnidirectional wheel are obtained by the fol-
Vb,XOZ = 0 (2) *
lowing methods. We define ω as the angle velocity of the
*
Kinetic energy Kb includes rotation parts and translation rotation of the ball, Pi as the coordinate vector of contact point
parts. As the origin of the coordinate system is in the center of the support ball.
of the ball, the mass-energy Vb = 0. Ib , mb , rb is the moment P1 (rb cos α, 0, rb sin α)
of inertia, the mass and the radius of the supporting ball. √
1 3
The kinetic energy of the support ball xoy plane only P2 (− rb cos α, rb cos α, rb sin α)
includes the rotation part, and the mass energy is 0. 2 2√
1 3
1 P3 (− rb cos α, − rb cos α, rb sin α) (15)
Kb,XOY = mb (rb ψ̇z )2 (3) 2 2
2
Evi is the speed of a point Pi on the ball, we obtain the following
The kinetic energy and mass energy of the ballbot body are
Equation.
as follows:
ωy Piz − ωz Piy
 
1 1
KB,XOZ = mB (rb2 ψ̇y2 + l 2 θ̇y2 ) + IB θ̇y2 (4) Evi = ω
E ×P E i =  ωz Pix − ωx Piz  (16)
2 2
VB,XOZ = mB gl cos θy (5) ωx Piy − ωy Pix
1 Esi represents the driving direction of the three omnidirectional
KB,XOY = IB θz2 (6)
2 wheels shown in Fig. 4, their value are as follows:
IB , mB , rB represent the moment of inertia, mass, radius of
s1 (0, −1, 0)
the ballbot body respectively. √
3 1
The angle velocity of the omnidirectional wheel is like that s2 ( , , 0)
of the support ball and the robot body. However, the moment 2√ 2
of inertia of the ballbot body and the support ball is far greater 3 1
s3 (− , , 0) (17)
than the moment of inertia of the drive wheel around the shaft 2 2

38316 VOLUME 7, 2019

C. Cai et al.: Kinematic Analysis and Control Algorithm for the Ballbot

roll forward by controlling three motors and the roll speed

of the ball is faster than the dumping speed of the body. It is
possible to use the similar principle of the inverted pendulum
to simulate the ballbot. Balancing control can be equivalent
to two inverted pendulum of x, y direction, which is a typical
nonlinear, multivariate, strong coupling and unstable system.
We use the cascade PD controller, and use fuzzy control in
the outer-loop angle control. The inner loop of the cascade
control system is a follow-up control system whose set value
varies with the output of the outer loop controller. The outer
loop controller can continuously adjust the set value of the
FIGURE 4. Driving direction of the three omnidirectional wheels. inner loop controller according to the operating conditions
and load changes, to ensure that the control system still has
better control effects under the condition that the operating
The relationship between the speed of three omnidi- conditions and load change. The fuzzy PD control automati-
rectional wheels and the body shown as Equation (18). cally adjusts the PD value according to the preset parameter
rb cos αωz is the speed of rotation around the z axis and table according to the input error. Different PD values act
appears at the speed of the three omnidirectional wheels. It is on nonlinear systems to accelerate the response speed of
not in the balancing control of x or y direction. nonlinear systems.

vs1 = −vy sin α − rb cos αωz

√
3 1
vs2 = vx sin α + vy sin α − rb cos αωz
2√ 2
3 1
vs3 = − vx sin α + vy sin α − rb cos αωz (18)
2 2
III. BALLBOT CONTROL
This section describes the various controllers used on the
ballbot. The complete control block diagram of the ballbot FIGURE 6. The block diagram of balancing control.
is shown as Fig. 5, including position control of x, y axes and
yaw control. The ballbot feedback roll, pitch, roll rate, pitch The block diagram of balancing control is shown as Fig. 6,
rate, speed and position to position control and yaw, yaw rate Equation (19) shows the error between the desired angle
to yaw control. The speed of the three wheels is determined and the Euler angle, which is considered as the input of the
by the superposition of the outputs of the yaw control and outer-loop angle controller. While there is only the balancing
position control. controller, the desired angle is 0deg. Desired angle velocity is
output of the outer-loop angle controller. The angle velocity
interference is more likely to cause instability of the ballbot
than the angle, cascade controller plays the role of anticipa-
tory control and effectively suppresses the inner loop inter-
ference. The inner-loop angle velocity is controlled by PD.
As shown in Equation (20), the output of the inner-loop is
the acceleration in the direction of x and y. The perform
frequency of the inner-loop is usually 2 ∼ 5 times of the
outer-loop frequency in cascade control. We choose 2 times
according to the experience of the experiment. The con-
troller in xoz plane is the same as yoz plane, the parameters
setting is the same. The fuzzy control rule table is shown
in Table 1, While the angle and angle velocity are in the
FIGURE 5. The complete control block diagram of the ballbot. same direction, the system becomes more and more unstable,
and the error should be eliminated quickly, Kp should be
the largest one. On the contrary, if the system has a trend
A. BALANCING CONTROL of gradual stabilization, and Kp will be set to be small to
It was proposed to use the control methods of inverted pendu- avoid overshoot. In addition to this, the adjustment of Kp is
lum to control the robot. For the balance of ballbot, it holds the also related to the magnitude of the angle and angle velocity.
balance around the unstable point by controlling the rotation The speed is obtained by integrating acceleration in sampling
of three omnidirectional wheels. The ballbot drive balls to time, as show in Equation (21). The speed in the direction

VOLUME 7, 2019 38317

 C. Cai et al.: Kinematic Analysis and Control Algorithm for the Ballbot

TABLE 1. Attitude control fuzzy rule. ëyaw is the difference between the current sampling value
ėyaw and the last sampling value. It is worth noting that in
head-hold mode, the desired angle of balancing control is
affected by the yaw angle. The specific steps are to rotate
the desired angle and transform the geographic coordinate
system into its own coordinate system by using the rotation
matrix R ∈ SO(2).

θRoll θRoll
 ∗    ∗ 
cos eYaw − sin eYaw
= (24)
θPitch
∗ sin eYaw cos eYaw θPitch
∗

of x and y can be divided into three directions of motors

according to the Equation (18) so that the ballbot body can
be dynamically balanced on the ball.
θ̇ ∗ = kpAngle (θ ∗ − θ ) (19)
F = kpAngle_rate (θ̇ ∗ − θ̇ ) + kdAngle_rate θ̈ (20)
F
v = v+ t (21)
m

FIGURE 8. The block diagram of position control.

C. POSITION CONTROL
The ballbot will move erratically at equilibrium despite the
balancing controller, which is still incomplete to a robot used
for people. Position control and balancing control are mutu-
ally coupled because the ballbot tilts the body to move in the
floor just like balancing robot. As shown in Fig. 8, position
FIGURE 7. The block diagram of yaw control.
cascade controller includes outer-loop position controller and
inner-loop speed controller. the outer-loop position controller
B. YAW CONTROL is PI controller that outputs the desired speed which depend-
The biggest innovation of the ballbot is that there is no turning ing on the error between desired position and real position.
radius, which is closely related to the control of the yaw angle. The PI controller effectively reduces the difference of the
Yaw control is an independent control, decoupling of balanc- position point during the movement. The inner-loop speed
ing control. Yaw control includes head-hold mode and head- controller is a manually tuned PD controller that can reach
free mode and the control block diagram is shown in Fig. 7. the desired speed quickly. The output of the speed controller
In the head-free mode, the ballbot’s reference coordinate is desired angle determined by the error between the output of
system is the geographic coordinate system. When initialized, the position controller and the actual speed. The actual speed
the ballbot will calculate an initial yaw angle Yawinit through of the robot is measured by the encoder, and the equations for
a magnetic sensor, then the yaw angle is controlled by PD calculation are as follows.
controller.
vx = speedm1 − 0.5speedm2 − 0.5speedm3
Fyaw = kpfree · eyaw + kdfree · ėyaw (22) vy = 0.86(speedm2 − speedm3 ) (25)
The error eyaw is the difference between the yaw angle and
To decrease the error caused by the mechanical gap of
the initial yaw angle during the movement. ėyaw represents
encoder, the real speed is determined by converging the
the yaw angle rate measured by IMU. The output is limited to
speeds in Equations (21) and (25). Limiting the output ampli-
avoid excessive output. In the head-hold mode, the ballbot’s
tude of position control prevents the ballbot from falling
reference coordinate system is its body coordinate system.
due to excessive tilt. The size of the limit determines the
The desired yaw angle of the robot is artificially locked in the
maximum speed of ballbot movement. When the desired
positive direction of the x, the control equation is as follows.
position is (0,0), the ballbot keeps the station by position
Fyaw = kpfree · ėyaw + kdfree · ëyaw (23) controller.

38318 VOLUME 7, 2019

C. Cai et al.: Kinematic Analysis and Control Algorithm for the Ballbot

FIGURE 11. A series of picture of the ballbot smashed by volleyball.

FIGURE 9. First version of the ballbot.

IV. RESULT
Fig. 9 is the first version of ballbot, we used hollow cup
reducer as drive motor. After the system modeling analy-
sis and the design of the control system, the ballbot was
debugged and the dynamic balance was reached. Meanwhile,
when we tested the anti-interference, the ballbot body was
easy to fall from the support ball. Then, the ballbot can
only balance in situ, be unable to move. At the same time,
the driven wheel of rubber material is easy to be damaged,
and it cannot ensure that there is no sliding friction when the
omnidirectional wheels rotate parallel to the center axis of the
motor. We make the second version. The supporting ball is
changed into a basketball, so there is enough friction between
the wheel and the ball and between the ball and the floor.
We carry out several experiments to verify the stability and
robustness of revision of the ballbot. Firstly, we only retain
the balancing control and set the desired attitude angle to FIGURE 12. The roll, roll rate. pitch and pitch rate of the ballbot smashed
by volleyball.
(0◦ , 0◦ ). Attitude angle of the ballbot is updated with the
frequency of 400Hz. The balancing control period is 10ms.
As shown in the Fig. 10, the roll and pitch of the ballbot can
volleyball. Obviously, the whole process is divided into three
be kept in the range of (−0.4◦ ∼ 0.4◦ ). Due to the lack of
stages. The first stage is that before the impact, the roll of
feedback on the position, the ballbot moves wildly.
the ballbot is balanced near 0 deg, the roll rate is between
(-4deg/s,4deg/s); The second is that the roll rate becomes
15deg and the roll rate reaches 93.5deg/s drastically at the
impact moment; Finally the roll and roll rate is restored to
the same level as the first stage after about 4 seconds.

FIGURE 10. The attitude angle of the ballbot under balancing control.

To verify anti-interference of the ballbot, we took a violent

impact experiment by using volleyball. Because the robot is
‘‘riding’’ on the ball, there is no other mechanical constraint, FIGURE 13. Position and attitude angle of the ballbot during station
the interference can only be applied to the ballbot body. keeping.
As shown in Fig. 11, we smash the volleyball into the robot
body. The attitude angle of the ballbot increases obviously, As shown in Fig. 13, the ballbot moves between (-2cm, -
and the dynamic balance is resumed after moving for a period 2.6cm) and (2.5cm, 3cm) after combining with the position
along the direction be smashed. The Fig. 12 shows the roll closed loop and balancing controller with setting a desired
and pitch, roll rate and pitch rate of the ballbot smashed by position (0cm, 0cm). The diameter of support ball is 24.5cm

VOLUME 7, 2019 38319

 C. Cai et al.: Kinematic Analysis and Control Algorithm for the Ballbot

and the ball contacted with the ground through is a very small the position tracking has little hysteresis. After reaching the
plane. Therefore, the range of motion error is acceptable. desired position, the ballbot balances within a narrow range
When keeping the original station, the range of attitude angle of oscillations. The biggest error between desired and actual
is (−0.7◦ ∼ 0.6◦ ), which is slightly larger than −0.4◦ ∼ 0.4◦ trajectory is 10.8cm. Yaw angle shown in Fig. 15, we rotate
under the balancing control. The reason is that the desired the desired yaw angle by -90deg after each vertex of the
angle is the output of the position control, and is no longer square is reached. Overshoot is the key factor of large error
the 0 under the balancing control. at the vertex of the final trajectory.
Table 2 shows the comparison of the position and attitude
angle error with other control methods when the ballbot is
keeping station, which are currently the state-of-the-art LQR
control method respectively [19]. The red numbers in the
table represent that our method is superior to LQR. Our posi-
tion control method makes the ballbot dynamically balance
in a smaller range.

TABLE 2. The comparison of the position and attitude angle error with
other control methods. FIGURE 15. Yaw angle of the ballbot where the preset trajectory is a
square of 150cm.

FIGURE 16. position of the ballbot where the preset trajectory is a circle
The trajectory of the ballbot is as shown in the case where with a radius of 100cm.
the preset trajectory is a square of 150cm. Because of the
omnidirectionality of the ballbot, there are two methods to The position of the ballbot is shown in Fig. 16 where the
move along the preset trajectory: one is to set the desired preset trajectory is a circle with a radius of 100cm. The head-
values of positionX and positionY in headless mode; the hold mode is used in yaw control as above. The ballbot starts
other is to set the expected values of yaw and positionX in from (0,100) at a speed of 0.15m/s and returns to the end
headless mode. Because experiments show that instability (0,100). Throughout the process, the expected yaw angle is
sometimes occurs when the first method is used, the yaw always 0, so the maximum error is less than the case where
control is in the head-hold mode during the entire process, the preset trajectory is square.
which means that the ballbot always moves in the direction of
its head. The ballbot is instructed to move to the desired posi-
tion (150cm, 0), (150cm,150cm), (0,150cm), (0,0) at a speed
of 0.15m/s under position control and yaw control. As shown
in Fig. 14, Because PI controller is used for position control,

FIGURE 17. Reverse movement of the ballbot when starts and stops.

As the ballbot balancing control is an inverted pendulum

problem, it must incline toward the direction of the movement
to accelerate. At the same time, the inclination of the main
FIGURE 14. Position of the ballbot where the preset trajectory is a square
body causes the ball to retreat in the opposite direction. Just
of 150cm. as shown in Fig. 17, there is a large reverse movement when

38320 VOLUME 7, 2019

C. Cai et al.: Kinematic Analysis and Control Algorithm for the Ballbot

the ballbot starts and stops. The recoil movement will cause [14] B. Vaidya et al., ‘‘Operation of the ballbot on slopes and with center-of-
instability of the ballbot. We reduce the parameter Kp and the mass offsets,’’ in Proc. IEEE Int. Conf. Robot. Automat. (ICRA), May 2015,
pp. 2383–2388.
integral time appropriately to reduce overshoot, the overshoot [15] S. Li et al., ‘‘Event-trigger heterogeneous nonlinear filter for wide-area
of system decreases by strengthening the integral function. measurement systems in power grid,’’ IEEE Trans. Smart Grid, to be
published.
[16] A. Bonci et al., ‘‘Embedded system for a Ballbot robot,’’ in Proc. Int.
V. CONCLUSIONS Workshop Intell. Solutions Embedded Syst., Oct. 2015, pp. 157–161.
Cascade fuzzy control method is designed to solve the bal- [17] H. Zabihi, H. A. Talebi, and A. A. Suratgar, ‘‘Open-loop trajectory plan-
ning and nonlinear control for underactuated spherical wheel mobile robot
ance of the ballbot which moves in any direction flexibly (Ballbot),’’ in Proc. 24th Iranian Conf. Elect. Eng. (ICEE), May 2016,
and turn more efficiently than traditional wheel-supported pp. 1–6.
robot. Position control uses cascade PI and PD controller [18] G. Seyfarth et al., ‘‘Initial results for a ballbot driven with a spherical
induction motor,’’ in Proc. IEEE Int. Conf. Robot. Automat., May 2016,
and yaw control divided into head-hold and head-free modes pp. 3771–3776.
are proposed. The ballbot is completed according to the [19] C. C. Tsai, C. K. Chan, and L. C. Kuo, ‘‘LQR motion control of a
mathematical model and the designed controller. Using the ball-riding robot,’’ in Proc. IEEE/ASME Int. Conf. Adv. Intell. Mechatron.,
Jul. 2012, pp. 861–866.
cascade fuzzy control, the ballbot achieves a dynamic balance
of −0.4◦ ∼ 0.4◦ and completes anti-interference experi-
ments with amazing performance. The ballbot walking along
the preset trajectory with an error of 7.2% under the joint
operation of position control and yaw control. Due to the
excellent omnidirectional movement of the ballbot, it moves
freely in a narrow environment. An interesting topic for future CHENGTAO CAI received the Ph.D. degree from
Harbin Engineering University, Harbin, China,
research would be to improve control accuracy by applying in 2008, where he was a Postdoctoral Research
deep learning to control methods. Scientist in shipbuilding and oceanography engi-
neering, from 2009 to 2011. From 2006 to 2007,
he was a Visiting Scholar with the National
REFERENCES Research Council Canada. He is currently a Pro-
[1] U. Nagarajan, G. Kantor, and R. L. Hollis, ‘‘Trajectory planning and fessor with the College of Automation, Harbin
control of an underactuated dynamically stable single spherical wheeled Engineering University. His research interests
mobile robot,’’ in Proc. IEEE Int. Conf. Robot. Automat., May 2009, include robot intelligent control and panoramic
pp. 3743–3748. vision applications.
[2] H. G. Nguyen et al., ‘‘Segway robotic mobility platform,’’ Proc. SPIE,
vol. 5609, pp. 207–220, Dec. 2004.
[3] A. Lotfiani, M. Keshmiri, and M. Danesh, ‘‘Dynamic analysis and control
synthesis of a spherical wheeled robot (Ballbot),’’ in Proc. 1st RSI/ISM Int.
Conf. Robot. Mechatron. (ICRoM), Feb. 2013, pp. 481–486.
[4] U. Nagarajan et al., ‘‘State transition, balancing, station keeping, and yaw
control for a dynamically stable single spherical wheel mobile robot,’’ in
Proc. IEEE Int. Conf. Robot. Automat., May 2009, pp. 998–1003. JIAXIN LU was born in Inner Mongolia, China,
[5] M. Shomin and R. Hollis, ‘‘Fast, dynamic trajectory planning for a dynami- in 1996. He received the B.S. degree in engineer-
cally stable mobile robot,’’ in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., ing from Harbin Engineering University, in 2017,
Sep. 2014, pp. 3636–3641. where he is currently pursuing the M.S. degree
[6] A. Mampetta, ‘‘Automatic transition of ballbot from statically stable in control science and engineering. His research
state to dynamically stable state,’’ Robot. Inst., Carnegie Mellon Univ., interests include robot and computer vision.
Pittsburgh, PA, USA, Tech. Rep. CMU-RI-TR-06-41, Sep. 2006.
[7] L. Hertig, D. Schindler, M. Bloesch, C. D. Remy, and R. Siegwart,
‘‘Unified state estimation for a ballbot,’’ in Proc. IEEE Int. Conf. Robot.
Automat., May 2013, pp. 2471–2476.
[8] G. Seyfarth, A. Bhatia, O. Sassnick, M. Shomin, M. Kumagai, and
R. Hollis, ‘‘Initial results for a ballbot driven with a spherical induc-
tion motor,’’ in Proc. IEEE Int. Conf. Robot. Automat., May 2016,
pp. 3771–3776.
[9] T. B. Lauwers, G. A. Kantor, and R. L. Hollis, ‘‘A dynamically stable
single-wheeled mobile robot with inverse mouse-ball drive,’’ in Proc. IEEE
Int. Conf. Robot. Automat., May 2006, pp. 2884–2889. ZUOYONG LI received the B.S. and M.S. degrees
[10] M. Kumaga and T. Ochiai, ‘‘Development of a robot balanced on a ball— in computer science and technology from Fuzhou
Application of passive motion to transport–,’’ in Proc. IEEE Int. Conf. University, Fuzhou, China, in 2002 and 2006,
Robot. Automat., May 2009, pp. 916–921. respectively, and the Ph.D. degree from the School
[11] P. Asgari et al., ‘‘Dynamics modelling and stable motion control of a of Computer Science and Technology, Nanjing
Ballbot equipped with a manipulator,’’ in Proc. 1st RSI/ISM Int. Conf. University of Science and Technology (NUST),
Robot. Mechatron., Feb. 2013, pp. 21–36. Nanjing, China, in 2010. He is currently a Pro-
[12] A. C. Satici, F. Ruggiero, and V. Lippiello, ‘‘Intrinsic Euler-Lagrange fessor with the College of Computer and Control
dynamics and control analysis of the ballbot,’’ in Proc. IEEE Amer. Control Engineering, Minjiang University, Fuzhou. He has
Conf., Jul. 2016, pp. 5685–5690. published over 60 papers in international/national
[13] J. Jian et al., ‘‘Standing-up control and ramp-climbing control of a spher- journals. His current research interests include image processing, pattern
ical wheeled robot,’’ in Proc. Int. Conf. Control Automat. Robot. Vis., recognition, and machine learning.
Dec. 2014, pp. 1386–1391.

VOLUME 7, 2019 38321