COLLECTIVE LEARNING OF CONCEPTS USING A ROBOT TEAM

Ana Cristina Palacios-Garc´ıa, Ang´elica Mu˜noz-Mel´endez and Eduardo F. Morales

National Institute for Astrophysics, Optics and Electronics

Luis Enrique Erro No. 1, Santa Mar´ıa Tonantzintla, Puebla, Mexico

Keywords:

Robotics and automation mobile robots and autonomous systems, Vision, Recognition and reconstruction,

Network robotics.

Abstract:

Autonomous learning of objects using visual information is important to robotics as it can be used for local

and global localization problems, and for service tasks such as searching for objects in unknown places. In

a robot team, the learning process can be distributed among robots to reduce training time and produce more

accurate models. This paper introduces a new learning framework where individual representations of objects

are learned on-line by a robot team while traversing an environment without prior knowledge on the number

or nature of the objects to learn. Individual concepts are shared among robots to improve their own concepts,

combining information from other robots that saw the same object, and to acquire a new representation of an

object not seen by the robot. Since the robots do not know in advance how many objects they will encounter,

they need to decide whether they are seeing a new object or a known object. Objects are characterized by

local and global features and a Bayesian approach is used to combine them, and to recognize objects. We

empirically evaluated our approach with a real world robot team with very promising results.

1 INTRODUCTION

The design of robot teams is a very active research do-

main in the mobile robotics community. Robot teams

have effectively emerged as an alternative paradigm

for the design and control of robotic systems because

of the team’s capability to exploit redundancy in sen-

sing and actuation.

The research on robot teams has focused on de-

veloping mechanisms that enable autonomous robots

to perform collective tasks, such as strategies for co-

ordination and communication (Asada et al., 1994;

Matari´c, 1997); exploration, mapping and deploy-

ment (Howard et al., 2006); sensing, surveillance and

monitoring (Parker, 2002); and decentralized decision

making (Wessnitzer and Melhuish, 2003). In these

works, a robot team can reduce time to complete a

complex task that is allocated among its members.

Despite constant research on the design of robot

teams, very little attention has been paid so far to the

development of robot teams capable of learning from

their interaction with their environment. In addition

to their capability for accelerated learning, learning

robot teams can be used to acquire a much richer and

varied information compared to the information ac-

quired by single learning robots.

Learning is a key issue to achieve autonomy for

both, single robot and robot teams. Learning capa-

bilities can provide robots ﬂexibility and adaptation

needed to cope with complex situations. In the con-

text of robot teams, the most common machine learn-

ing approach has been reinforcement learning, where

the idea is to learn optimal policies using a set of

robots to improve the coordination of individual ac-

tions in order to reach common goals (Asada et al.,

1994; Matari´c, 1997; Parker, 2002; Fern´andez et al.,

2005).

In this work we use visual information to learn,

with a team of robots, descriptions of objects placed

in a particular environment. Learning to recognize

particular objects in an environment is important for

robotics as it can be used for local and global local-

ization tasks as well as for simple service tasks such

as searching for objects in unknown places. Contrary

to previous approaches, in our learning setting, the

robots are not told the number or nature of the objects

to be learned.

Vision is a primary source of perception in

robotics and provides different features that can be

used to classify objects. In general, using a particu-

lar set of features can be adequate for particular tasks

but inadequate for other tasks. In this work, objects

are characterized by two complementary features: (i)

SIFT features (Lowe, 2004) and (ii) informationabout

Palacios-García A., Muñoz-Meléndez A. and F. Morales E. (2010).

COLLECTIVE LEARNING OF CONCEPTS USING A ROBOT TEAM.

In Proceedings of the 7th International Conference on Informatics in Control, Automation and Robotics, pages 79-88

DOI: 10.5220/0002952800790088

 SciTePress

the silhouettes of objects. Other features could be

used as well, but the main objective in this work is to

show the different cases and possible confusions that

can arise in the recognition of objects and merging of

concepts, and how they can be addressed.

Numerous difﬁculties arise in robot teams when

learning as well as sharing concepts that represent

concrete objects. Some of these issues are discussed

by Ye and Tostsos (1996) and include, how do robots

represent their local views of the world, how is the

local knowledge updated as a consequence of the

robot’s own action, how do robots represent the local

views of other robots, and how do they organize the

knowledge about themselves and about other robots

such that new facts can be easily integrated into the

representation. This article addresses the individual

and collective representation of objects from visual

information using a team of autonomous robots.

The rest of the paper is organized as follows. Sec-

tion 2 reviews related work. Sections 3 y 4 describe,

respectively, the stages of individual learning and col-

lective learning of concepts. Section 5 describes our

experimental results, and Section 6 provides conclu-

sions and future research work.

2 RELATED WORK

Interesting experiments where physical mobile robots

learn to recognize objects from visual information

have been reported. First we review signiﬁcant work

developed for individual learning, and then we review

learning approaches developed for robot teams.

Steels and Kaplan (2001) applied an instance-

based method to train a robot for object recognition

purposes. In this work objects are represented by

color histograms. Once different representations have

been learned from different views of the same object,

the recognition is performed by classifying new views

of objects using the KNN algorithm (Mitchell, 1997).

Ekvall et al. (2006) used different learning tech-

niques to acquire automatically semantic and spatial

information of the environment in a service robot sce-

nario. In this work, a mobile robot autonomously

navigates in a domestic environment, builds a map,

localizes its position in the map, recognizes objects

and locates them in the map. Background sub-

traction techniques are applied for foreground ob-

jects segmentation. Then objects are represented

by SIFT points (Lowe, 2004) and an appearance-

based method for detecting objects named Receptive

Field Co-occurrence Histograms (Ekvall and Kragic,

2005). The authors developed a method for active ob-

ject recognition which integrates both local and global

information of objects.

In the work of Mitri et al. (2004), a scheme for

fast color invariant ball detection in the RoboCup con-

text is presented. To ensure the color-invariance of

the input images, a preprocessing stage is ﬁrst applied

for detecting edges using the Sobel ﬁlter, and speciﬁc

thresholds for color removal. Then, windows are ex-

tracted from images and predeﬁned spatial features

such as edges and lines are identiﬁed in these win-

dows. These features serve as input to an AdaBoost

learning procedure that constructs a cascade of clas-

siﬁcation and regression trees (CARTs). The sys-

tem is capable of detecting different soccer balls in

RoboCup and other environments. The resulting ap-

proach is reliable and fast enough to classify objects

in real time.

Concerning the problem of collective learning of

objects using robot teams there are, as far as we know,

very few works. Montesano and Montano (2003) ad-

dress the problem of mobile object recognition based

on kinematic information. The basic idea is that if

the same object is being tracked by two different

robots, the trajectories and therefore the kinematic in-

formation observed by each robot must be compati-

ble. Therefore, location and velocities of moving ob-

jects are the features used for object recognition in-

stead of features such as color, texture, shape and size,

more appropriate for static object recognition. Robots

build maps containing the relative position of moving

objects and their velocity at a given time. A Bayesian

approach is then applied to relate the multiple views

of an object acquired by the robots.

In the work of O’Beirne and Schukat (2004), ob-

jects are represented with Principal Components (PC)

learned from a set of global features extracted from

images of objects. An object is ﬁrst segmented and

its global features such as color, texture, and shape are

then extracted. Successive images in a sequence are

related to the same object by applying a Kalman ﬁl-

ter. Finally, a 3D reconstructed model of an object is

obtained from the multiple views acquired by robots.

For that purpose, a Shape From Silhouette based tech-

nique (Cheung et al., 2003) is applied.

In contrast to previous works, in our method each

member of the robot team learns on-line individual

representations of objects without prior knowledge on

the number or nature of the objects to learn. Indi-

vidual concepts are represented as a combination of

global and local features extracted autonomously by

the robots from the training objects. A Bayesian ap-

proach is used to combine these features and used for

classiﬁcation. Individual concepts are shared among

robots to improve their own concepts, combining in-

formation from other robots that saw the same object,

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics

and to acquire a new representation of an unnoticed

object.

3 INDIVIDUAL LEARNING OF

CONCEPTS

The individual concepts are learned on-line by a robot

team while traversing an environment without prior

knowledge on the number or nature of the objects to

learn. The individual learning of concepts consists

of tree parts: object detection, feature extraction and

individual training.

Individual concepts of objects are represented by

Principal Component (PC) over the information about

the silhouettes of objects and Scale Invariant Features

(SIFT). Learned concepts are shared among robots.

3.1 Object Detection

Robots move through an environment and learn des-

criptions of objects that they encountered during na-

vigation. Objects are detected using background

substraction. In this paper we assume a uniform and

static background. We performed morphological o-

perations (closing - erode) to achieve better segmen-

tation. Once an object is detected, it is segmented and

scaled to a ﬁxed size, to make the global PC features

robust to changes in scale and position.

3.2 Feature Extraction and Individual

Training

The segmented objects are grouped autonomously by

the robots in sets of images containing the same ob-

ject. Robots assume that they are observing to the

same object while it can be detected, and they ﬁnish

to see it when they can not detect objects in the cap-

tured images. Only one object can be detected in an

image at the same time. For each set of images, the

robot obtains an individual concept that represents the

object.

Training using Global Features. We applied Prin-

cipal Component Analysis (PCA) over the average

silhouettes that are automatically extracted from the

set of images of a particular object. The average pro-

vides a more compact representation of objects and

reduces segmentation errors. Figure 1 (a) shows an

object used in the training phase, Figure 1 (b) shows

its silhouette, and Figure 1 (c) illustrates the average

silhouette obtained from a set of images that represent

the object of Figure 1 (a). Once the robot has obtained

(a) (b) (c)

Figure 1: Examples of the silhouette (b) and average silhou-

ette (c) of an object (a).

Figure 2: Examples of the SIFT features extracted from a

set of images and the ﬁnal set of SIFT features.

an average silhouette, this is added by the robot to a

set of known average silhouettes. After that, the robot

uses PCA to reduce the dimensionality of all average

silhouettes learned to get the PC features that repre-

sent them.

Training using Local Features. Each robot ex-

tracts local SIFT features of each image of the set of

images, and groups them in a ﬁnal set which contains

all the different SIFT features that represent an object.

In Figure 2 we show an example of the SIFT points

obtained from a set of images of a vase and the ﬁnal

set of SIFT points obtained. The PC features and the

SIFT features represent the individual concept of the

observed object.

3.3 Sharing Concepts

The concepts learned by robots are shared among

them to achieve collective learning. This can be done

off-line or on-line. In the case of collective off-line

learning the robots share their individual concepts

once they have learned all the training objects. On

the other hand, in the collective on-line learning the

robots share their individual concept as soon as a new

object is learned.

4 COLLECTIVE LEARNING OF

CONCEPTS

Collective learning of concepts enables robots to

improve individual concepts combining information

COLLECTIVE LEARNING OF CONCEPTS USING A ROBOT TEAM

from other robots that saw the same object, and to ac-

quire a new representation of an object not seen by the

robot. Therefore, a robot can learn to recognize more

objects of what it saw and can improve their own con-

cepts with additional evidence from other robots.

A robot has to decide whether the concept shared

by another robot is of a new object or of a previously

learned concept. A robot can face three possibilities:

coincident, complementary or confused information.

The shared concepts are fused depending on the kind

of information detected, as described below.

4.1 Pre-analysis of Individual Concepts

The concept learned by a robot is deﬁned as follows:



Sil

, SIFT



(1)

where C

is the concept k learned by robot i, Sil

the average silhouette, and SIFT

is the set of SIFT

features that form the concept k.

In order to determine if a shared concept is pre-

viously known or not to a robot, it evaluates the pro-

babilities that the PC features and SIFT features are

previously known by the robot. The probability vec-

tors of PC features calculated by robot i, v

, indi-

cate the probability that a concept shared by robot

j, C

, is similar to the concepts known by robot i,

, . . . ,C

numObjs

, given the global features. numObjs

is the number of concepts of objects known by robot

i. The process to obtain the probability vector PC is

described as follows:

- A temporal training set of silhouettes is formed by

adding the average silhouettes of concepts known by

robot i or actual robot, Sil

, ..., Sil

numObjs

, and the a-

verage silhouette of the shared concept Sil

- The PCA is trained using the temporal set of average

silhouettes. The projection of the average silhouettes

know by robot i is obtained as a matrix of projections,

matProys. The projection of the average silhouette

Sil

is obtained in a vector, vectProys.

- The Euclidean distance (dE) is calculated between

each vector of the matrix matProys and the vector

vectProys as shown in formula 2, i.e, we obtain the

distance between all the projections already computed

and the projection of the new silhouette.

nEigens

∑

r=1

(matProys

(l,r)

− vectProys

(1,r)

)

(2)

where nEigens is the number of eigenvectors used

during the PCA training (nEigens = numOb js

− 1),

and l is the index of the distance vector, where the

maximum size of the vector dE

is numObjs

- The distance value dE

is divided by a maximum dis-

tance value, ThresholdMax, determined experimen-

tally to obtain a similarity metric also called the pro-

bability vector v

as shown in formula 3.

= 1 −

ThresholdMax

(3)

If dE

is bigger than the ThresholdMax value, then

the probability will be ﬁxed as shown in formula 4,

which indicates that the projections of the object j and

the one of the object i are completely different.

numObjs

(4)

The value of the SIFT similarity metric also called

the probability vector SIFT at the position v

, is

obtained calculating the number of coincident SIFT,

coin

, between the individual SIFT concept SIFT

learned by robot i, and the individual SIFT concept

SIFT

shared by robot j. If the number n

coin

is big-

ger than an average of coincidences determined ex-

perimentally, AverageCoin, then the probability will

be ﬁxed to v

= 1.0, which means that both concepts

contain the same local features SIFT. In other case,

the probability will be calculated using formula 5.

coin

AverageCoin

(5)

The constant AverageCoin represents the average

of coincidences between two sets of SIFT points of

the same object from different perspectives.

4.2 Analysis and Fusion of Individual

Concepts

This section describes how to detect if the shared con-

cept is coincident, complementary or confused, and

how the individual concepts are fused to form collec-

tive concepts depending on the kind of detected con-

cept.

4.2.1 Coincident Concepts

A coincident concept is detected when two or more

robots of the robot team learned individual concepts

from similar views of the same object. A shared con-

cept is classiﬁed as coincident if v

≥ α and v

≥ α.

That is, if both probabilities (PC and SIFT) of a pre-

viously learned concept are greater than a predeﬁned

threshold value (α). If a shared concept is determined

as coincident it is merged with the most similar known

concept as follows:

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics

PCA Fusion. It is obtained by evaluating a new a-

verage silhouette from the average of the known Sil

and new Sil

silhouettes. After that, it is necessary to

re-train the PCA substituting the concept Sil

with the

new average silhouette which contains information of

the concept learned by robot j.

SIFT Fusion. It is obtained by adding the comple-

mentary SIFT points of concept SIFT

to the set of

SIFT points of concept SIFT

. Also, each pair of

coincident SIFT points of both concepts is averaged

in terms of position and their correspondingSIFT des-

criptors.

The main idea to fuse individual concepts is to im-

prove their representation.

4.2.2 Complementary Concepts

A concept C

contains complementary information

if it differs with all known concepts by robot i, i.e.,

if both shape and local features are different to all

known concepts by robot i, C

, . . . ,C

numObjs

. That is,

if v

< α and v

< α.

A complementary concept C

is fused with the

collective concepts known by the robot i as follows:

PCA Fusion. The new average silhouette is added

and the new PC concepts are obtained by re-training

the PCA using the updated set of average silhouettes.

SIFT Fusion. The new SIFT features are simple

added to the current set of SIFT concepts known by

the robot i.

4.2.3 Confused Concepts

There are two types of confusion that can occur bet-

ween concepts:

Different Shape and Similar Local Features (type

1). This type of confusion occurs when the new con-

cept C

is complementary by shape, Sil

, to all the

concepts known by the robot i, Sil

, ..., Sil

numObjs

but

it is coincident by local SIFT features, SIFT

, with

at least one concept known by the robot i. That is,

≥ α and if v

< α.

Similar Shape and Different Local Features (type

2). This type of information occurs when concept

is coincident by shape, Sil

, to at least one concept

known by the robot i, but it is complementary using

its local SIFT features, SIFT

. That is, if v

≥ α and

< α.

In both types of confusion, type 1 or type 2, there

can be two options:

a) Different Objects. Both concepts correspond to

different objects.

b) Same Object. Both concepts correspond to the

same object but they were learned by robots from

different points of view.

In our current approach, both types of confusions

are solved as complementary objects. The reason is

that robot i cannot distinguish with its current infor-

mation between both, differentobjects or same object,

using only the individual and the shared concepts. To

solve the ambiguity, as future work each robot should

build autonomously a map and locate its position in

the map. In addition, for each learned object, robots

will locate them in the map. For confused objects a

robot can move to the position of the object marked

in the map to see the object from different perspec-

tives in order to solve the conﬂict.

5 EXPERIMENTS AND RESULTS

We performed several experiments to demonstrate the

proposed algorithm. In section 5.1, we show the re-

sults of a general experiment that demonstrates the

main features of the proposed approach. In section

5.2 we present the accuracy of the collective concepts

versus the individual concepts.

In these experiments we used a robot team con-

sisting of two homogeneous Koala robots equipped

with a video camera of 320× 240 pixels. For more

than two robots our method can be applied straight-

forward. The only difference is that robots will need

to consider the information from more than one robot,

possibly reducing confused concepts.

5.1 Concept Acquisition and Testing

The mobile robots learn on-line a representation of

several objects while following a predeﬁned trajec-

tory without prior knowledge on the number or nature

of the objects to learn. The idea of using pre-planned

trajectories instead of making the robots wandering

randomly, is that we can control the experimental

conditions to show different aspects of the proposed

methodology.

Each robot shares its individual concept as soon as

it is learned to improve the representation of this con-

cept or to include a new concept in the other robot.

COLLECTIVE LEARNING OF CONCEPTS USING A ROBOT TEAM

(a) (b) (c) (d) (e) (f) (g)

Figure 3: Training objects. a) vase, b) water bottle, c) can,

d) dolphin, e) soda bottle, f) bottle and g) cone.

Figure 3 shows the training objects used in this ex-

periment. As can be seen in the ﬁgure, some objects

have the same shape but different texture, some have

the same texture but different shape, some others are

not symmetric in their shape. The objective of this

experiment is to show the performance of the system

to detect coincident, complementary and confused in-

formation under a wide variety of conditions.

Robot 1 (R1) learned during individual training

concepts for: dolphin, can, water bottle and vase.

Robot 2 (R2) learned individual concepts for: vase,

soda bottle, bottle and cone. Note that some ob-

jects are learned by both robots while others are only

learned by one robot.

While learning a new concept, each robot has to

decide whether to fuse the current concept with a pre-

viously known concept or include it as a new one. Ta-

bles 1 and 2 show the probability vectors of the PC

features based on shape (v

) and of the SIFT features

) obtained by Robot 1. Tables 3 and 4 show the

probability vectors of the PC (v

) and SIFT (v

) fea-

tures obtained by Robot 2. In these tables the coinci-

dent information is represented in bold.

Table 1: Probability vectors PC (v

) obtained by R1.

New (collective concepts R1)

Objects Dol-

phin

Vase Can Soda

bot-

tle

Water

bot-

tle

Bot-

tle

Co-

Dol-

phin

- - - - - - -

Vase

0.19 - - - - - -

Can

0.31 0.26 - - - - -

Soda

bottle

0.36 0.28 0.58 - - - -

Water

bottle

0.43 0.28 0.53 0.73 - - -

Bottle

0.31 0.17 0, 56 0.61 0.58 - -

Vase

0.25 0.69 0.42 0.43 0.41 0.32 -

Cone

0.31 0.01 0.28 0.28 0.33 0.43 -

We used the deﬁned criteria in Section 4.2 to

recognize coincident, complement or confused con-

Table 2: Probability vectors SIFT (v

) obtained by R1.

New (collective concepts R1)

Objects Dol-

phin

Vase Can Soda

bot-

tle

Water

bot-

tle

Bot-

tle

Co-

Dol-

phin

- - - - - - -

Vase

0.09 - - - - - -

Can

0.12 0.12 - - - - -

Soda

Bottle

0.28 0.11 0.40 - - - -

Water

bottle

0.15 0.59 0.20 0.20 - - -

Bottle

0.08 0.15 0.65 0.04 0.12 - -

Vase

0.16 1.00 0.23 0.10 0.08 0.09 -

Cone

0.05 0.28 0.43 0.10 0.14 0.09 -

Table 3: Probability vectors PC (v

) obtained by R2.

New (collective concepts R2)

Objects Vase Dol-

phin

Soda

bot-

tle

Can Bot-

tle

Water

bot-

tle

Co-

Vase

- - - - - - -

Dol-

phin

0.19 - - - - - -

Soda

bottle

0.28 0.36 - - - - -

Can

0.26 0.31 0.58 - - - -

Bottle

0.17 0.31 0.61 0.56 - - -

Water

bottle

0.29 0.44 0.73 0.54 0.58 - -

Cone

0.01 0.31 0.28 0.28 0.43 0.33 -

Vase

0.69 0.25 0.43 0.42 0.32 0.41 0.01

cepts, with α = 0.65 as threshold value, and the

probability vectors of Tables 1, 2, 3 and 4.

Tables 5 and 6 show the results of the analysis per-

formed by each robot. As can be seen from these ta-

bles, each robot encountered the three types of possi-

ble information and fuse its concepts accordingly.

For instance, Table 1 shows how are the probabi-

lities of objects of R1 affected using only PCA over

shapes of objects, as both robots encounter and learn

concepts while traversing the environment. In the ﬁrst

row, R1 learns about the concept dolphin and acquires

it. In the second row, R2 then learns about vase and

shares this concept to R1. The probability, according

to the PCA features to be a dolphin is 0.19 (second

row). R1 learns the object can, which has a probabi-

lity of 0.31 to be a dolphin and a probability of 0.26

to be a vase, which was learned by R2 and shared to

R1 (third row). In the ﬁfth row, R1 learns about a wa-

ter bottle but it confuses with the soda bottle learned

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics

Table 4: Probability vectors SIFT (v

) obtained by R2.

New (collective concepts R2)

Objects Vase Dol-

phin

Soda

bot-

tle

Can Bot-

tle

Water

bot-

tle

Co-

Vase

- - - - - - -

Dol-

phin

0.18 - - - - - -

Soda

bottle

0.11 0.28 - - - - -

Can

0.12 0.12 0.04 - - - -

Bottle

0.15 0.08 0, 04 0.64 - - -

Water

bottle

0.59 0.15 0.20 0.20 0.12 - -

Cone

0.09 0.05 0.10 0.43 0.09 0.14 -

Vase

1.00 0.16 0.10 0.23 0.09 0.08 0.12

Table 5: Detected information by R1 for each own and

shared individual concepts.

Individual

concepts

objects

Kind of

info.

Dol-

phin

- - - Comple-

mentary

Vase

All l

(l = 1 to

numObj

)

(2,l)

0.65

(2,l)

0.65

Comple-

mentary

Can

All l v

(3,l)

0.65

(3,l)

0.65

Comple-

mentary

Soda

bottle

All l v

(4,l)

0.65

(4,l)

0.65

Comple-

mentary

Water

bottle

Soda

bottle

(5,4)

0.73

(5,i)

0.65

Confuse

type 2

Bottle

Can v

(6,l)

0.65

(6,3)

0.65

Confuse

type 1

Vase

Vase v

(7,2)

0.69

(7,2)

1.00

Coinci-

dent

Cone

All l v

(8,l)

0.65

(8,l)

0.65

Comple-

mentary

and shared before by R2. As can be seen from Fig-

ure 3, both objects have the same shape and conse-

quently the PCA features are not able to discriminate

between these two objects. This is not the case for

the SIFT features, which prevent R1 to consider it as

the same object (as explained below). In the seventh

row, R1 learns about vase which was already learned

and shared by R2, and in this case both concepts are

merged.

To test the performance of the individual concepts

and the collective concepts acquired by each robot,

the concepts were used in an object recognition task.

Each robot followed a predeﬁned trajectory to recog-

nize objects in the environment. The objects were de-

Table 6: Detected information by R2 for each own and

shared individual concepts.

Individual

concepts

objects

Kind of

info.

Vase

- - - Comple-

mentary

Dol-

phin

All k v

(2,k)

0.65

(2,k)

0.65

Comple-

mentary

Soda

bottle

All k v

(3,k)

0.65

(3,k)

0.65

Comple-

mentary

Can

All k v

(4,k)

0.65

(4,k)

0.65

Comple-

mentary

Bottle

All k v

(5,k)

0.65

(5,k)

0.65

Confuse

type 1

Water

bottle

Soda

Bottle

(6,3)

0.73

(6,k)

0.65

Confuse

type 2

Cone

All k v

(7,k)

0.65

(7,k)

0.65

Comple-

mentary

Vase

Vase v

(8,1)

0.69

(8,1)

1.00

Coinci-

dent

tected by the robot team in the following order: cone,

water bottle, vase, bottle, soda bottle and dolphin.

Once an object is detected, the robot (i) evaluates its

class using the PC (v

) and SIFT (v

) probability vec-

tors and combines both probabilities using a Bayesian

approach:

× v

× P



× v

× P





(1− v

) × (1− v

) × (1− P

)



(6)

where P

is a uniform probability distribution

numObjs

), v

= p(PC projection | Class =

i), v

= p(SIFT matching | Class = i), P

the Bayesian probability vector (p(Class = i |

PCprojection, SIFTmatching), and l is the index of

the Bayesian probability vector, where the maximum

size of the probability vector is numObjs

Figures 4, 5 and 6 show the average probabilities

obtained during the object recognition task for each

set of images of the same class, using the individual

and collective learned concepts. The dotted bars in-

dicate the classiﬁcation errors. A classiﬁcation error

is produced when a robot classiﬁes an unknown ob-

ject with a probability ≥ 0.6. The unknown objects

for robots 1 and 2 are those which were not learned

during their individual training.

The classiﬁcation errors of Robot 1 in Figure 4

occur when the objects cone, bottle and soda bottle

are classiﬁed as dolphin, water bottle and water bot-

tle, respectively. The classiﬁcation errors of Robot 2

COLLECTIVE LEARNING OF CONCEPTS USING A ROBOT TEAM

Figure 4: Average PC classiﬁcation probabilities for the ob-

ject recognition task using the individual and collective PC

concepts.

Figure 5: Average SIFT classiﬁcation probabilities for the

object recognition task using the individual and collective

SIFT concepts. Any robot makes classiﬁcation errors.

occur when the objects can, water bottle and dolphin

are classiﬁed as bottle, soda bottle and soda bottle,

respectively. For the vase there is no classiﬁcation er-

ror because both robots learn individual concepts of

it.

Although the probability bars presented in the pre-

vious ﬁgures show a higher probability for individual

concepts than for collective concepts, in reality the

collective concepts are more robust as they represent

the probabilities considering a larger number of ob-

jects. This will be discussed in Section 5.2.

Table 7: Precision in the object recognition task using the

individual and collective concepts acquired by each robot.

R1 R2 R1-R2 R2-R1

PCA 55.69 %

(100.00%)

49.98 %

(94.82 %)

86.15 % 86.16 %

SIFT 48.32 %

(86.23 %)

42.89 %

(79.11 %)

87.84 % 87.84 %

Bayes 52.59 %

(94.20 %)

51.68 %

(81.54 %)

80.73 % 80.73 %

We show in Table 7 the precision of the object

recognition task using the individual and collective

concepts. The precision is presented in two ways,

one considering the total number of objects, and the

other one taking into a count only the number of ob-

jects used during the individual training (reported in

Figure 6: Bayesian classiﬁcation probabilities which uses

the Bayesian fusion of the PCA and SIFT classiﬁcation

probabilities.

parentheses). As can be seen the collective concepts

produce a signiﬁcantly better precision.

5.2 Accuracy of the Individual and

Collective Concepts

In this section we compare the results of the indivi-

dual concepts with that of collective concepts. In each

experiment, a different set of objects was used, and

both robots learned the same set of objects. There-

fore, all the shared concepts were coincident, that is,

robots learned both individually and collectively the

same number of concepts. At the end of each ex-

periment the robots learned four concepts that were

proved by a test sequence. In Figure 7 we present

the accuracy obtained by the robots using the PC fea-

tures of the individual and collective concepts in an

object recognition task. Figure 7 shows the averages

in accuracy of the number of images well classiﬁed

under six experiments (correct), the average percent-

ages of the number of non detected or non classiﬁed

images (no detected), and the average percentage of

false positives for each concept (false +). Figure 8

and 9 show, respectively, the accuracy obtained by the

robots when using the SIFT vectors and the Bayesian

approach.

As it can be observed in Figures 7, 8 and 9, the

accuracy that indicate the quantity of well classiﬁed

images (correct) using the collective concepts for the

object recognition task, is in general better than the

accuracy using the individual concepts. For PC, SIFT

and Bayes there is an improvement in the accuracy up

to 2.56 %, 13.79 % and 20.62 %, respectively. This

demonstrates that the collective concepts have better

coverage than the individual concepts because they

contain information acquired from different points of

view, which allows a better recognition of test ob-

jects. Also, the percentages of the number of non de-

tected images of collective concepts are smaller than

the ones of the individual concepts.

In Table 8 we present the average percentages of

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics

false positives for both, the individual and the collec-

tive concepts acquired by the robots. We conclude

that the collective concepts have better quality than

the individual concepts.

In general for both, the individual and the collec-

tive concepts, we observed an improvement in the ac-

curacy when using the Bayesian approach. In Table 9

we present the average percentages of accuracy using

the individual and collective concepts.

The average proﬁt in the percentages of classiﬁca-

tion using the Bayesian approach using the collective

concepts with regard to the individual concepts is of

14.63 %.

Figure 7: Accuracy in coverage using the part PC of con-

cepts.

Figure 8: Accuracy in coverage using the part SIFT of con-

cepts.

Figure 9: Accuracy in coverage using the Bayesian ap-

proach.

Table 8: Average percentages of false positives.

PCA SIFT Bayes

Individual 14.42 % 0.64 % 0.64 %

Collective 13.14 % 0.00 % 0.00 %

Table 9: Average percentages of accuracy.

PCA SIFT Bayes

Individual 84.94 % 67.88 % 80.18 %

Collective 87.18 % 81.12 % 94.81 %

6 CONCLUSIONS AND FUTURE

WORK

In this paper we have introduced a new on-line learn-

ing framework for a team of robots. Some of the main

features of the proposed scheme are:

• The robots do not know in advance how many

objects they will encountered. This pose several

problems as the robots need to decide if a new

seen object or shared concept, is of a previously

learned concept or not.

• The representation of objects are learned on-line

while the robots are traversing a particular en-

vironment. This is relevant for constructing au-

tonomous robots.

• Three possible cases in which to merge concepts

and how to merge them were identiﬁed.

The detection of coincident concepts avoids pro-

ducing multiple concepts for the same object. The

detection of complementary concepts allows to detect

and learned unknown objects not seen by a particular

robot. The detection of confused concepts allows to

fuse information: 1) when the object have different

shape and similar SIFT features, and 2) when the ob-

jects have similar shape and different SIFT features.

These cases are particularly difﬁcult to deal with be-

cause the objects may be genuinely different or may

be the same but seen from different points of view by

the robots.

In general, the object recognition using the collec-

tive concepts had a better performance than using the

individual concepts in terms of accuracy. This occurs

because the collective concepts consider information

from multiple points of view producing more general

concepts.

As future work we propose to integrate schemes to

object segmentation for dynamic environments. For

instance, using an object segmentation based on dis-

tance as in M´endez-Polanco et al., 2009. Use a differ-

ent set of features and identify possible conﬂicts be-

COLLECTIVE LEARNING OF CONCEPTS USING A ROBOT TEAM

tween more that two kind of features. We also plan to

incorporate planning of trajectories to autonomously

allocate the environment among robots. We also plan

to add strategies to solve some confusions in shared

concepts by taking different views from these objects.

Finally, we plan to incorporate our algorithm for robot

localization and search of objects, and to test our work

for robot teams with three or more robots.

ACKNOWLEDGEMENTS

The ﬁrst author was supported by the Mexican Na-

tional Council for Science and Technology, CONA-

CYT, under the grant number 212422.

REFERENCES

Asada, M., Uchibe, E., Noda, S., Tawaratsumida, S., and

Hosoda, K. (1994). Coordination of multiple behav-

iors acquired by vision-based reinforcement learning.

Proceedings of the International Conference on Intel-

ligent Robots and Systems, pages 917–924.

Cheung, G. K. M., Baker, S., and Kanade, T. (2003). Vi-

sual hull alignment and reﬁnement across time: A 3-

D reconstruction algorithm combining Shape-From-

Silhouette with stereo. Proceedings of the IEEE Com-

puter Society Conference on Computer Vision and

Pattern Recognition (CVPR ’03), 2:375–382.

Ekvall, S., Jensfelt, P., and Kragic, D. (2006). Integrating

active mobile robot object recognition and SLAM in

natural environments. IEEE/RSJ International Con-

ference on Intelligent Robots and Systems (IROS ’06),

pages 5792–5797.

Ekvall, S. and Kragic, D. (2005). Receptive ﬁeld cooccur-

rence histograms for object detection. IEEE/RSJ In-

ternational Conference on Intelligent Robots and Sys-

tems (IROS ’05), pages 84–89.

Fern´andez, F., Borrajo, D., and Parker, L. E. (2005). A re-

inforcement learning algorithm in cooperative multi-

robot domains. Journal of Intelligent and Robotic Sys-

tems, pages 161–174.

Howard, A., Parker, L. E., and Sukhatme, G. S. (2006).

Experiments with a large heterogeneous mobile robot

team: Exploration, mapping, deployment and detec-

tion. International Journal of Robotics Research,

25:431–447.

Lowe, D. G. (2004). Distinctive image features from scale

invariant keypoints. International Journal of Com-

puter Vision, 60(2):91–110.

Matari´c, M. J. (1997). Reinforcement learning in the multi-

robot domain. Autonomous Robots, 4(1):73–83.

M´endez-Polanco, J. A., Mu˜noz-Mel´endez, A., and Morales,

E. F. (2009). People detection by a mobile robot using

stereo vision in dynamic indoor environments. Pro-

ceedings of the 8th Mexican International Conference

on Artiﬁcial Intelligence (MICAI ’09), pages 349–359.

Mitchell, T. M. (1997). Machine Learning. McGraw-Hill

Science/Enginering/Math.

Mitri, S., Pervolz, K., Surmann, H., and Nuchter, A.

(2004). Fast color independent ball detection for mo-

bile robots. Proceedings of the IEEE International

Conference Mechatronics and Robotics (MechRob

’04), pages 900–905.

Montesano, L. and Montano, L. (2003). Identiﬁcation of

moving objects by a team of robots based on kine-

matic information. Proceedings of the IEEE/RSJ In-

ternational Conference on Intelligent Robots and Sys-

tems (IROS ’03), 1:284–290.

O’Beirne, D. and Schukat, M. (2004). Exploration and ob-

ject recognition using cooperative robots. Proceedings

of the International Conference on Imaging Science,

Systems and Technology (CISST ’04), pages 592–598.

Parker, L. E. (2002). Distributed algorithms for multi-robot

observation of multiple moving targets. Autonomous

Robots, 3(12):231–255.

Steels, L. and Kaplan, F. (2001). AIBO’s ﬁrst words. The

social learning of languaje and meaning. Evolution of

Comunication, 4(1):3–32.

Wessnitzer, J. and Melhuish, C. (2003). Collective decision-

making and behaviour transitions in distributed ad hoc

wireless networks of mobile robots: Target-hunting.

Advances in Artiﬁcial Life, 7th European Conference

(ECAL ’03), pages 893–902.

Ye, Y. and Tsotsos, J. K. (1996). On the collaborative object

search team: a formulation. Distributed Artiﬁcial In-

telligence Meets Machine Learning (ECAI ’96 Work-

shop), pages 94–116.

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics