Skuba 2009 Team Description 

Kanjanapan Sukvichai

1

, Piyamate Wasuntapichaikul

2

, Jirat Srisabye

2

, 

and Yodyium 

Tipsuwan

2

 

 

 

1

 Dept. of Electrical Engineering, Faculty of Engineering, Kasetsart University. 

2

 Dept. of Computer Engineering, Faculty of Engineering, Kasetsart University. 

50 Phaholyothin Rd, Ladyao Jatujak, Bangkok, 10900, Thailand

 

baugp@hotmail.com

 

http://iml.cpe.ku.ac.th/skuba

 

Abstract.

 This paper is used to describe the Skuba Small-Size League robot 

team. Skuba robot is designed under the World RoboCup 2009 rules in order to 
participate in the ssl competition in Graz, Austria. The overview describes both 
the robot hardware and the overall software architecture of our team.  

Keywords: 

Small-size, Robocup, Vision, Robot Control, Artificial Intelligence.  

1   Introduction 

Skuba is a small-size league robot team from Kasetsart University [1], which entered 
the World RoboCup competition since 2006. Skuba got the third place in the world 
ranking last year from the World RoboCup 2008 in Suzhou, China. During the last 
year competition, problems about the robot low level controller and multi-agent game 
plans were revealed.  

This year, robot low level controller is redesigned along with the new open loop 

skills. Both are implemented in Skuba robot 2009. Omni-directional wheels robot is 
one of the most popular mobile robot which is used in most of the teams because of 
its maneuverability. The major problem for many teams is how to tune the low level 
controller gains. The surface parameters are changed according to the time because 
the carpet is damaged from the robot wheels. Therefore, all of the low level controller 
gains for every wheel have to be adapted every match. Torque control scheme is 
implemented in this year in order to solve this problem. Torque controller consists of 
the PI control and the torque converter. The new idea of the modified robot 
kinematics is implemented in order to make the open loop game plans possible. 

The vision system process two video signals from the cameras mounted on top of 

the field. It computes the positions and the orientations of the ball and robots on the 
field then transmit the information back to the AI system 

The AI system receives the information and makes strategic decisions. The 

decisions are converted to commands that are sent back to the robots via a wireless 
link. The robots execute these commands and set actions as ordered by the AI system. 

2   Robot  

The major issue about Skuba 2008 robot is the robustness. Most mechanical parts are 
general purpose grade aluminum which can be damaged easily. The Skuba 2009 robot 
is built based on Skuba 2008 design with some modified parts. The major parts were 
built using aircraft grade aluminum alloy to improve strength such as the kicker and 
the chip-kicker. 

Each robot consists of four omni-directional wheels which are driven by 30 watt 

Maxon flat brushless motors. The kicker has ability to kick the ball at speeds up to 14 
m/s using a solenoid. The chip-kicker is a large powerful flat solenoid attached with a 
45 degree hinged wedge located on the bottom of the robot which can chip the ball up 
to 7.5 m before it hits the ground. Both of the solenoids are driven from two 2700

μ

F 

capacitors charged to 250V. Kicking devices are controlled by a separate board 
located below the middle plate. The kicking speed is fully variable and limited to 10 
m/s according to the rule. 

The controller of the robot hardware is done by using a single-chip Spartan 3 

FPGA from Xilinx. The FPGA contains a soft 32-bit microprocessor core and 
peripherals. This embedded processor executes the low level motor control loop, 
communication and debugging. The motor controller, quadrature decoder, kicker 
board controller, PWM generation and onboard serial interfaces are implemented 
using FPGA. The robot receives control commands from the computer and sends back 
the status for monitoring using a bidirectional 2.4GHz wireless module. A Kicker 
board is a boost converter circuit using a small inductor, the board is seperated from 
the main electronics for safety. 

The size of robot has a diameter of 178 mm and a height of 144mm.The dribbler 

covers to 20% of the ball. The 3D model of the robot and the real robot are shown in 
Fig. 1 and Fig. 2 respectively. 

 

 

Fig. 1.

 3D mechanical model of Skuba 2009 robot 

 

1

α

2

α

3

α

4

α

1

f

2

f

3

f

4

f

d

d

d

d

r

x

r

y

e

x

e

y

θ

1

α

2

α

3

α

4

α

1

f

2

f

3

f

4

f

d

d

d

d

r

x

r

y

1

α

2

α

3

α

4

α

1

f

2

f

3

f

4

f

d

d

d

d

r

x

r

y

e

x

e

y

θ

 

Fig. 2.

 Skuba 2009 robot 

2.1   Robot Dynamics 

Dynamics of a robot is derived in order to provide information about its behavior. 
Kinematics alone is not enough to see the effect of inputs to the outputs because the 
robot kinematics lacks information about robot masses and inertias. The dynamic of a 
robot can be derived by many different methods such as Newton’s law [2] [3] and 
Lagrange equation [4]. In this paper, Newton’s law is used to solve the robot dynamic 
equation.  

Newton’s second law is applied to robot chassis in Fig 2 and the dynamic equation 

can be obtained as (1) though (3). 

 

(

sin

sin

sin

sin

)

1

1

2

2

3

3

4

4

1

v

&&

f x

x

f

f

f

f

f

M

α

α

α

α

=

−

−

+

+

−

                     (1) 

( cos

cos

cos

cos

)

1

1

2

2

3

3

4

4

1

v

&&

f y

y

f

f

f

f

f

M

α

α

α

α

=

−

−

+

−

                  (2) 

(

)

1

2

3

4

&&

trac

J

d f

f

f

f

T

θ

=

+

+

+

−

                                        (3) 

where, 

 

&&

x

   is the robot linear acceleration along the x-axis of the global reference frame  

&&

y

   is the robot linear acceleration along the y-axis of the global reference frame 

M

  is the total robot mass 

i

f

  is the wheel i motorized force 

v

f

f

  is the friction force vector 

i

α

  is the angle between wheel i and the robot x-axis  

&&

θ

 

is the robot angular acceleration about the z-axis of the global frame 

J

 

is the robot inertia 

d

 

is the distance between wheels and the robot center 

trac

T

  is the robot traction torque 

 The robot inertia, friction force and traction torque are not directly found from the 

robot mechanic configuration. These parameters can be found by experiments. The 
robot inertia is constant for all different floor surfaces while the friction force and 
traction torque are changed according to floor surfaces.  

The friction force and traction torque are not necessary found at this point because 

these two constraints are different for different floor surfaces and their effect can be 
reduced by using the control scheme which is discussed in the next topic. The wheel 
force can be written in a motor torque form as:  

  

(

sin

sin

sin

sin

)

1

2

3

4

1

2

3

4

1

v

&&

m

m

m

m

f x

x

f

M

r

r

r

r

τ

τ

τ

τ

α

α

α

α

=

−

−

+

+

−

                          (5) 

 

(

cos

cos

cos

cos

)

1

2

3

4

1

2

3

4

1

v

&&

m

m

m

m

f

y

y

f

M

r

r

r

r

τ

τ

τ

τ

α

α

α

α

=

−

−

+

−

                          (6)  

(

)

1

2

3

4

&&

m

m

m

m

trac

d

T

J

r

r

r

r

τ

τ

τ

τ

θ =

+

+

+

−

                                     (7)  

where,  

r

     is a wheel radius 

 

 

Equation (5) though (7) show that the dynamic of the robot can be directly 

controlled by using the motor torques.  

2.2   Modified Robot Kinematics 

The pervious topic, the dynamics of the robot is derived. Although the dynamics can 
be correctly used to predict the robot behavior but it is hard to directly implement and 
it needs a long computing time. In this topic, the regular mobile robot kinematics is 
modified. First the friction force and traction torque vector are defined as a system 
disturbance. The normal kinematics can be written as: 

  

r

Designed

ζ

ψ ζ

= ⋅

                                                      (13) 

where, 

T

r

ζ

φ φ φ φ

⎡

⎤

= ⎣

⎦

1

2

3

4

&

&

&

&

 

T

Designed

x y

ζ

θ

⎡

⎤

= ⎣

⎦

&

& &

 

cos

sin

cos

sin

sin

sin

cos

cos

cos

sin

cos

sin

sin

sin

cos

cos

cos

sin

cos

sin

sin

sin

cos

cos

cos

sin

cos

sin

sin

sin

cos

cos

d
d
d
d

θ

α

α

θ

θ

α

α

θ

θ

α

α

θ

θ

α

α

θ

ψ

θ

α

α

θ

θ

α

α

θ

θ

α

α

θ

θ

α

α

θ

⋅

+

⋅

⋅

−

⋅

−

⎡

⎤

⎢

⎥

⋅

+

⋅

⋅

−

⋅

−

⎢

⎥

=

⎢

⎥

⋅

+

⋅

⋅

−

⋅

−

⎢

⎥

⋅

+

⋅

⋅

−

⋅

−

⎣

⎦

1

1

1

1

2

2

2

2

3

3

3

3

4

4

4

4

 

 

Designed robot velocity (

Designed

ζ

) is used to generate robot’s wheel angular 

velocity vector (

r

ζ

). This wheels angular vector is the control signal which is sent 

from PC to interested mobile robot. The output linear velocity (

Captured

ζ

) is captured 

by a bird eye view camera. The output velocity contains information about 
disturbances, therefore by comparing the designed velocity and the output velocity. 
The output velocity can be defined as (14) when assuming that disturbance is constant 

for the specific surface. The disturbance is modeled and separated to the disturbance 
from the robot coupling velocity and the disturbance from the surface friction. 

 

†

(

)

Captured

r

ζ

ψ

ε ζ

=

+ ⋅

+ Δ

                                            (14) 

where, 

†

ψ

  is the pseudo inverse of the kinematic equation 

ε

  is the disturbance gain matrix due to the robot coupling velocity  

Δ

 

is the disturbance vector due to the surface friction 

The disturbance matrices can be found from experiments. The first designed robot 

velocity (

Designed

ζ

1

 ) is applied to the robot in order to get the first output velocity 

(

Captured

ζ

1

) in the first experiment. The first experiment is repeated in the second 

experiment with the second designed robot velocity (

Designed

ζ

2

) and the second output 

velocity (

Captured

ζ

2

) is captured. The disturbance matrices now can be found by adding 

(13) to (14) for both experiments. 

†

(

)

Captured

Designed

ζ

ψ

ε ψ ζ

=

+ ⋅ ⋅

+ Δ

1

1

                                    (15) 

†

(

)

Captured

Designed

ζ

ψ

ε ψ ζ

=

+ ⋅ ⋅

+ Δ

2

2

                                    (16) 

Subtract (15) by (16): 

 

†

†

(

)

(

)

Captured

Captured

Designed

Designed

ζ

ζ

ψ

ε ψ ζ

ψ

ε ψ ζ

−

=

+ ⋅ ⋅

−

+ ⋅ ⋅

1

2

1

2

 

†

†

((

) (

)

)

Captured

Captured

Designed

Designed

I

ε

ζ

ζ

ζ

ζ

ψ

=

−

⋅

−

− ⋅

1

2

1

2

                 (17) 

 
Substitute (17) to (15) and 

Δ

 is found. 

2.3   Motor Model and Torque Control 

A Maxon brushless motor is selected for the robot. The dynamic model of the motor 
can be derived by using the energy conservation law as shown in [5]. The dynamic 
equation for the brushless motor is  

 

,

2

30 000

&

m

m

m

m

m

u

R

k

k

τ

τ

π

φ τ

⎛

⎞

⋅

=

⋅ ⋅

+ ⋅⎜

⎟

⎝

⎠

                                 (8) 

where,  

u

 

is the input voltage 

m

τ

  is the motor output torque 

m

k

  is the motor torque constant 

&

φ

 

is the motor angular velocity 

R

  is the motor coil resistance 

 

 

Equation (8) is not easy to implement to the control law. Therefore, this equation 

has to be modified by using the Maxon parameters relationship, which is shown in its 
datasheet, and the final dynamic equation of the motor is  

 

&

m

m

m

n

k

k

u

R

R k

τ

φ

⎛

⎞

⎛

⎞

=

⋅ −

⋅

⎜

⎟

⎜

⎟

⋅

⎝

⎠

⎝

⎠

                                     (9) 

where, 

n

k

  is the motor speed constant  

The control scheme is set using the discrete Proportional-Integral control law and 

torque dynamic equation (9). The control system runs at 600Hz cycle [1].  The error 
between desired angular velocity and real filtered angular velocity of each wheel is 
the input of the PI controller with the PI gains 

p

K

and 

I

K

  respectively. The 

controller is shown in Fig. 2 and the control law can be described as (10) though (12). 

 

real

φ

&

desired

φ

&

[

]

real

filtered

φ

&

PI

err

FIR

e

τ

Torque

Convertor

*

u

Motor

∑

real

φ

&

desired

φ

&

[

]

real

filtered

φ

&

PI

err

FIR

e

τ

Torque

Convertor

*

u

Motor

∑

 

Fig. 3.

 Torque controller scheme 

 

[ ]

[ ]

[

[ ]]

&

&

desired

real

err j

j

filtered

j

φ

φ

=

−

                                      (10) 

[ ]

[ ]

(

[ ])

1

N

d

p

I

j

j

k err j

k

err j

τ

=

=

⋅

+ ⋅

∑

                                       (11) 

[ ]

*[ ]

[

[ ]]

&

d

m

m

cc

real

n

j

u

j

k

k

V

filtered

j

R

R k

τ

φ

=

⎛

⎞

⎛

⎞ ⋅ −

⋅

⎜

⎟

⎜

⎟

⋅

⎝

⎠

⎝

⎠

                           (12) 

 

where, 

N

  is the number of samples 

cc

V

  is the driver supply voltage 

 

 

 
The output from (12) is converted to the Pulse Width Modulation (PWM) signal 

and directly used as input signal for every poles of the motor. The difference between 
the regular discrete PI controller for the wheel angular velocity and the torque 
controller is the torque converter block which is shown in Fig 3 and defined as (12). 

 
 
 
 
 

3   Vision  

Our vision structure diagram is shown in Fig 4. 

 

•

 

Capture Device

  

Our team applies the global vision and 

uses the output signal of two cameras. We 
employ 

AVT Stingray F-046C

 1394b firewire 

camera which is capable of grabbing 780 x 
580 images at 62 fps 

 

•

 

Preprocessing

 

The preprocessing is used to improve the 

quality of the image. 

 

•

 

Transform Color Space  

We transform color model to the HSV 

space, which consists of a hue, a saturate and 
a value. The HSV space is more stable than 
RGB space in different light properties. 

 

•

 

Color Segmentation 

The color segmentation assigns each 

image pixel into color classes. Currently, we 
classify and segment color by CMVision2.1 
library [6]. 

 

•

 

Object Localization 

After color segmentation, we receive all 

the color regions. The filtering process 
discards incorrect regions. Then, object 
localization computes the position and 
orientation of objects in the field from the 
final regions. 

 

•

 

Tracking Update 

Objects which is received from 

localization has a lot of noise, so we need to 
track it. Our approach is working by the 
Kalman Filter. 

 

•

 

Transmit to AI  

  Fig. 4.

Vision system structure             

 This component consists of network link  

         

 

 

      communication between the vision system and  

                                                               system 

3.1   Camera Calibration 

Camera calibration is a part in Object localization. We compute the internal and 
external parameters of the cameras using the Tsai [7] algorithm. These parameters are 
used to correct the distortion produced by the camera lenses. 

4   Multilayer, Learning-based Artificial Intelligence  

Multi-layered learning based agent architecture is applied to the RoboCup domain. 
Upper layer are used to control the activation and priority of behaviors in layers 
below, and only the lowest layer interacts directly with the robot. This year, the 
program is rebuilt from scratch by using strategy structure based on 
“StrategyModule” from Cornell Big Red 2002. 

 

 

Fig. 5. 

Strategy structure 

4.1   Play  

Play illustrates a specific global state of the AI and the general goal the positions are 
attempting to achieve at a given time. The system will transit from one play to another 
by learning-based method. We score the successful play more than the failed play.  

4.2   Skill  

Skill is a basic action of robot, such as “MoveToBallskill” or “Kickskill”. We can use 
a Neural Network for train each skill independently for the best efficiency. The 
modified robot kinematics is used in our new skill such as open loop pass and kick 
skill. 
 

 

 

Fig. 6. 

Skuba’s user interface 

5   Conclusion 

The new robot hardware design and the new approach of low level controller have 
been implemented and they improved the speed, precision, and flexibility of the 
robots. With some filters, we could acquire precisely coordinates of all players. The 
modified robot kinetics is used in the simulator and it can improve the robot overall 
efficiency. We believe that the RoboCup Small-Size League is and will continue to be 
an excellent domain to drive research on high-performance real-time autonomous 
robotics. We hope that our robot performs better in this competition than the last year 
competition. We are looking forward to share experiences with other great teams 
around the world. 

References 

1. Srisabye, J., Hoonsuwan, P., Bowarnkitiwong, S., Onman, C., Wasuntapichaikul, P., 

Signhakarn, A., et al., Skuba 2008 Team Description of the World RobCup 2008, Kasetsart 
University, Thailand.  

2.  Oliveira, H., Sousa, A., Moreira, A., Casto, P., Precise Modeling of a Four Wheeled Omni-

directional Robot, Proc. Robotica’2008 (pp. 57-62), 2008. 

3. Rojas, R., Forster, A., Holonomic Control of a robot with an omnidirectional drive, 

Kunstliche Intelligenz, BottcherIT Verlag, 2006. 

 4. Klancer, G., Zupancic, B., Karba, R., Modelling and Simulation of a group of mobile robots, 

Simulation and Modelling Practice and Theory vol. 15 (pp. 647-658), ScienceDirect, 
Elsevier, 2007. 

5. Maxon motor, Key information on – maxon DC motor and maxon EC, Maxon Motor 

Catalogue 07(pp. 157–173), 2007. 

6.  Bruce, J.: CMVision realtime color vision system. (The CORAL Group’s Color Machine 

Vision Project) http://www.cs.cmu.edu/˜jbruce/cmvision/. 

7.  Tsai, R.Y.: A versatile camera calibration technique for high accuracy 3D machine vision 

using off-the-shell TV cameras and lenses, IEEE Journal of robotics and Automation, 1987.