CYCLOP SYSTEM: MOTION CONTROL OF MULTI-ROBOTS

THROUGH A SINGLE EYE

George Tho, Douglas Tavares, Pablo Alsina, Luiz Gonc¸alves

Universidade Federal do Rio Grande do Norte

DCA-CT-UFRN, Campus Universit

ario, Lagoa Nova, 59072-970, Natal, RN, Brasil

Keywords:

Multi-robots, landmarks, visual servor, visual control

Abstract:

We propose a way to control several robots at a glance by using visual data provided by a single vision camera.

Position and orientation of the group is given by a main robot (a master) which has a very simple camera on

its top. This robot also has two encoders for determining its positioning related to the origin. The other (slave)

robots have no position/orientation sensors as camera/angle-sensors. Their only local sensing capabilities are

light (infra-red) and bump sensors. They have a top mark that allows the master discovering their orientation

and positioning through the vision software. All robots can be controled by using a control software running

in a host computer, through a communication protocol also developed in this work. The robots have some

software tools running inside a local processor and memory so they can perform some tasks autonomously, as

obstacle avoidance. The main goal is to control all slave robots to stay inside the ﬁeld of view of the master.

With this behavior, the master can have control over the slaves position and can eventually send some of them

to perform usefull tasks in certain places, inside its ﬁeld of view.

1 INTRODUCTION

We propose a way to control several robots at a glance

through a vision system architecture composed by a

single camera. The camera is mounted on the top of

a mobile robotics platform, a master robot, allowing

to acquire images in real-time. Computer vision tech-

niques are used to get position and orientation of other

robots (slaves), inside the ﬁeld of view of the cam-

era, in relation to the master robot. The slave robots

have a sphere on its top allowing its detection by the

vision system. All robots have local sensing capabil-

ities as infra-red light sensors, for obstacle distance,

and bump (touch) sensors. The master robot has en-

coders (angle sensors) that are used in combination

with landmarks for determining its position related to

the environment origin. All robots have a process-

ing unit that can run, locally, some small behaviors.

That is, some software tools (tasks) can be running

in its local processor and memory so they can even-

tually perform some tasks autonomously, as obstacle

avoidance. The robots can communicate to each other

and to a main computer via an infra-red port. All

robots have two motors which can be remotely con-

troled through a communication protocol also devel-

oped in this work. That is, a control software runs in

a host computer, which can repeatedly transmit con-

trol commands, in real-time, to all robots through the

communication protocol. The main goal is to con-

trol all slave robots to stay inside the ﬁeld of view of

the master (as the hen and young chikens biological

behavior). With this behavior, the master can have

control over the slaves position and can eventually

send some of them to perform usefull tasks in certain

places of its ﬁeld of view, individually or colectively.

It is possible that each robot has a specialized behav-

ior so that, if a mission is given to the group, each

robot can perform a part of it or some speciﬁc tasks in

parallel or sequential way.

2 RELATED WORK

Several works discuss use of visual features and

visual architectures for attention control (Treisman,

1985; Tsotsos et al., 1995). Vision is also used for

recognition purposes (Rao and Ballard, 1995; Rybak

et al., 1998). In previous work, researchers have used

vision for attention control and pattern recognition

(Gonc¸alves et al., 2000; Milanese et al., 1995). Sev-

339

Tho G., Tavares D., Alsina P. and Gonçalves L. (2004).

CYCLOP SYSTEM: MOTION CONTROL OF MULTI-ROBOTS THROUGH A SINGLE EYE.

In Proceedings of the First International Conference on Informatics in Control, Automation and Robotics, pages 339-342

DOI: 10.5220/0001133403390342

 SciTePress

eral works discuss robotics vision applications as a

mean to provide localization and cognition to robots,

mainly in the context of landmark selection (Milanese

et al., 1995). We used in previous work as a basis

for controling multi-user agents in a mixed reality en-

vironment, mainly given as a feedbak for the robots

controllers and also for determining their positioning

in the environment(Tavares et al., ). Several tools or

behaviors work in a bidirecional channel linking the

real platform and the virtual side of the application.

Control can be alternated betwenn real and virtual en-

tities, that is, the virtual can follow the real one, based

on its position and orientation given by the vision sys-

tem, and vice-versa, depending on a requested user

behavior. In the current work, we adopted a simi-

lar philosophy, but with the camera ﬁeld of view in

horizontal positioning, that is, with camera gaze par-

allel, approximately, to the ground plane (see Figure

1). The environment is composed by different types

of objects as obstacles and robots. The robots are

autonomous vehicles implemented by using LEGO

hardware prototyping technology. However, we have

developed/upgraded our own software tools for con-

trol based on the BricOS operating system, once the

software provided by the fabricant of LEGO kits is

very limited. Obstacles are colored cubes or cylin-

ders. The environment can have more than one robot

so a robot could be considered as a dynamic obstacle

for another one. With this hardware setup, we can per-

form many kind of multi-robot tasks as: (1) Moving

around and avoiding obstacles; (2) Looking for spe-

cial objects (as colored cubes or cylinders); (3) Iden-

tifying and catching special objects; (4) manipulating

objects.

Figure 1: Hardware setup.

3 THE PROPOSED SYSTEM

Basically, the system has three low-level working

modules and a high-level control module as seen in

Figure 2. The ﬁrst low-level module (image acquisi-

tion) is responsible for taking pictures from the cam-

era posted on the master robot and passing them to

the computer. This module runs on the host computer

and its coding depends on the hardware used for im-

age acquisition. Its output must be an array of chars

(bytes) encoding the image. This is transparent to the

vision processing, the second low-level module. This

module takes an image as input and returns position

and orientation of each robot inside the ﬁeld of view.

The third low-level module (robot control) is respon-

sible for packing commands and sending them to all

robots, that is, it provides communication between the

computer and the robots.

High−level

control

Image

Acquisition

Vision

processing

Robot

control

Figure 2: System architecture.

The vision module acquires images through a digi-

tal camera posted on the top of the master robot. This

camera is capable of capturing color images up to 352

× 264 pixels in size at a rate of 30 frames per second.

It is connected to the host computer through the USB

port. The vision software developed runs under Linux

Operating System. Images are captured on the RGB

format and stored in the computer memory as an uni-

dimensional array of bytes. Each pixel is simply rep-

resented by three consecutive bytes, corresponding to

its R, G, and B component values. Vision processing

provides in real time position and orientation for each

robot inside the ﬁeld of view of the robot containing

the camera. This information is determined by using

some image processing techniques to segmentate the

image followed by an algorithm for determining the

position and orientation of each detected robot.

We use a visual mark to identify each slave robot,

a sphere painted with two colors on the top of each

robot as seen in Figure 3. By ﬁnding this sphere,

we can get position and orientation of a robot. The

ball has 5 cm of diameter and is positioned approxi-

mately in the same height, in relation to the ground, as

the projection center of the camera. We consider that

the ground is a plane with same height level along it.

With that, the projection of the sphere in the image

lays, approximately, in a horizontal line passing near

the image center. This line is tracked from image to

image so its current position can be easily recovered.

The above key restrictions substantially reduce com-

putational costs.

We apply band-pass ﬁlters in the image, enhanc-

ICINCO 2004 - ROBOTICS AND AUTOMATION

340

Figure 3: Two colors (red/blue and white) ball.

ing red and/or blue and white pixels. A treshold is

applied, at this time asking for pixels in two bynary

images that attend an “or” between the colors, red or

white and then blue or white. Then, we jump from 6

to 6 pixels in the image central line in order to ﬁnd

a pixel inside the ball. We have, then, just to ﬁnd

the edges or separations between red/blue and white.

This is done after the determination of the ball center,

by scanning the center horizontal line for an abrupt

change, in the original image. This scheme for getting

orientation makes the algorithm a little slow, running

at a 15 fps. We are currently trying to improve this

performance.

The control module is responsible for mapping the

real world into an internal geometric representation

and is responsible for generating trajectories and pass-

ing motion commands for the robots. With the ori-

entation and position of each robot been informed at

each 0.033 of a second, the control module can be

devised. We implemented a robot controller using

only position and orientation, a simple proportional

(P) one. This resulting P controller is actually com-

posed of two other P controllers working in parallel.

One is responsible for position and another for orien-

tation. At the end, the angular velocity to be applied in

each motor (wheel) is the summation of the velocity

calculated by the position controller and the velocity

calculated by the orientation controller. To control po-

sition, the reference angle θ

ref

is is the simple differ-

ence between the desired position and the actual one.

For orientation, the reference signal is given by sub-

tracting the actual value from the desired one. Finally,

having calculated the value to be applied in each of

the wheels, it is necessary to send the signal through

the infra-red tower. Several values are sent as the left

and right velocity values, and some control variables.

These values must be in the interval [0, 255] (unsigned

char). That means a minimum change for velocity can

be of 1/255 of velocity unit at each time interval.

4 EXPERIMENTS AND RESULTS

We initially conducted experiments in order to setup

the transformation between (R, X)− > (z, x) (cal-

ibration of the system). Basically, we have 12 un-

knows and each picture gives 2 equations plus one re-

striction (an orthonormal tranform matrix). So, 4 pic-

tures would be enough for ﬁnding all coefﬁcients. As

we have taken 9 pictures, we used a minimum square

approach to perform this calculation. Table 4 summa-

rizes the taken data.

Image R X z(cm)

1 105.0 0 11

2 55.0 0 15

3 39.0 0 33

4 24.5 0 69

5 15.0 0 114

6 12.0 0 147

7 28.5 160 47

8 16.5 290 86

9 11.0 500 143

Figure 4: Data taken by hand.

Next, we tested the determination of the ball radius

in pixels by the vision system. The error in determi-

nation of the radius of the bal was up to 10%. That

means a maximum linear error up to 15 cm in the de-

termination of the robot positioning. Note that the er-

ror increases as a robot gets far from the master. Then,

we conducted some other experiments on determining

radius. Table 5 summarizes the corresponding calcu-

lated and measured by hand data. We have had no

more than 10% of error in the calculation of the robot

position.

Image Xc Yc Rc zh xh zc xc

a42 167 149 23 42 0 45 1

a43 32 149 21 42 15 41 14

a31 264 152 18 57 -15 53 -14

a32 171 151 19 57 0 56 0

a33 63 151 17 57 15 59 16

a22 170 152 16 72 0 70 2

a23 81 151 15 72 15 67 18

a11 233 154 14 87 -15 88 -13

a12 173 152 14 87 0 88 -1

a11 95 152 12 87 15 85 14

Figure 5: Data taken by hand and calculated by the system.

We conducted experiments for determining orien-

tation using two color balls. In these, the error in de-

termination of the angle was greather than 10 degrees

when the robot was close to zero or 180 degrees (re-

member origin for angle is gaze of master robot).

CYCLOP SYSTEM: MOTION CONTROL OF MULTI-ROBOTS THROUGH A SINGLE EYE

341

We also tested system performance in terms of time

spent for all calculations. For the system determining

only position (white ball), it can run in real time, up

to 30 fps. With the red/blue scheme, orientation and

position can be determined at a 10 times per minute

rate. This is also the rate in which the current version

of the communication protocol can operate in order

for the control part, so, this is ﬁne.

5 CONCLUSIONS AND FUTURE

WORK

The proposed techique have allowed to deﬁne po-

sition and orientation of the robots using only one

single camera and a ball with two colors posted on

the top of the robot. This is a good rewsult and

this scheme can be implemented by using very sim-

ple hardware interfaces. We are currently developing

a very robust robotic platform. We will change the

LEGO robots by this one in short term. This new mas-

ter robot will have embedded processing (a PC-104)

and a stereo vision system. With this platform, we

may send the robots to other places as rooms since it is

not connected to any other computer. We also intend

to specialize the slave robots so each of them pos-

sesses a native behavior as lift objects, follow lines,

and others. Depending on the problem proposed to

the master, this one could send a specialized robot to

certain positions. In future versions, we intend to put

odometry on the slave robots so one can estimate their

positions in the images, decreasing computations nec-

essary for ﬁnding them.

We are currently developing a way to reduce un-

certainty present in the odometry system of the mas-

ter robot. A common way to do that is to make the

robot learn visually discernible marks with known po-

sition in the environment. The known coordinates of

these visually discernible marks can then be used to

correct the odometer systematic and non-systematic

errors. We have a currently implemented approach in

which we consider lines separating tiles on the ground

with known sizes to constantly update the odome-

try. Another way is to determine or to learn some

features in the environment that can be retrieved by

image ﬁltering processes and use these feature read-

ings to try to determine a current positioning of the

robot. In principle, we intend to test with some ex-

isting techniques for landmark selection. Then to dis-

cover the best ones that can be used for detection of

useful marks in the environment. As the current posi-

tioning is supposed to be known, the robot can search

for some determined features in determined places of

its ﬁeld of view, trying to identify those.

REFERENCES

Gonc¸alves, L. M. G., Grupen, R. A., Oliveira, A. A.,

Wheeler, D., and Fagg, A. (2000). Tracing patterns

and attention: Humanoid robot cognition. The Intelli-

gent Systems and their Applications, 15(4):70–77.

Milanese, R., Gil, S., and Pun, T. (1995). Attentive mecha-

nisms for dynamic and static scene analysis. Optical

Engineering, 34(8).

Rao, R. P. N. and Ballard, D. (1995). An active vision ar-

chitecture based on iconic representations. Artiﬁcial

Intelligence Magazine, 78(1-2):461–505.

Rybak, I. A., Gusakova, V. I., Golovan, A. V., Podlad-

chikova, L. N., and Shevtsova, N. A. (1998). A model

of attention-guided visual perception and recognition.

Vision Research, 38(2):387–400.

Tavares, T., Lemos, G., Medeiros, A., Tho, G., Antunes, V.,

and Gonc¸alves.

Treisman, A. (1985). Preattentive processing in vision.

Computer Graphics and Image Processing, (31):156–

177.

Tsotsos, J., Culhane, S., Wai, W., Lai, Y., Davis, N., and Nu-

ﬂo, F. (1995). Modeling visual attention via selective

tuning. Artiﬁcial Intelligence Magazine, 78(1-2):507–

547.

ICINCO 2004 - ROBOTICS AND AUTOMATION

342