ROBOT BEHAVIOUR CONTROL:

embarrassedlopsidedAI and Robotics

Nov 14, 2013 (3 years and 9 months ago)

65 views

Ministry of Education and Science of the Russian Federation

State Educational Institution of Higher Professional Training

National Research Tomsk Polytechnic University






ROBOT

BEHAVIOUR CONTROL:

SUCCESSFUL
TRIAL
OF


MARKERLESS

MOTION

CAPTURE

TECHNOLOG
Y







Student


E.E.

Shelomentsev

Group

8
Е
00

Scientific supervisor
Т
.
V. Alexandrova

Language supervisor

T.I.Butakova






Tomsk 2012

2



Contents


1.

Introduction

................................
................................
................................
.....

3

2.

Methodology

................................
................................
................................
...

4

2.1. Markerless Motion
Capture Technology

................................
....................

5

2.2.

Hierarchical Attentive Multiple Models for Execution and Recognition
(HAMMER)

................................
................................
................................
.......

6

3.

Results

................................
................................
................................
.............

8

4.

Conclusion

................................
................................
................................
......

9

5.

References

................................
................................
................................
.....

10










3


1.

Introduction

The truth of the matter is

that
a

human

b
ehavio
r has been studied from
diverse
perspectives
by a number of disciplines.
Psychology

and

S
ociology,
A
nthropology
and
N
euroscience
apply various approaches,
use different
methodologies, scope
s
, and
also
evaluation criteria to understand
the

aspects of
such a

behavior.
Not to mention the fact
,

that
Computer
S
cience, and in
particular
R
obotics, offer a complementary perspective on the study
of
this
topic
.
In

other

words
,

it

researches the recogni
tion

of the human behavior.

The given

research focuses on
construc
ting

embodied computational models
of
a social
human behavior, especially the developmental pro
gression
, a process
of developing gradually towards a more advanced state,

of early social
human
skills such as
communication, interaction
,

etc.
Furthermore, the

research
embraces

the principles of
computational modeling

in different ways. For
example, computational modeling is used to predict future human actions and
fix
a
state of the robot.

Nowadays
,

a lot of

people are
certain

that a

new generation of robots w
ill

be
mostly
human
-
centered

[1] as opposed to
the
robots
,

which are
currently
being
used

by manufacturers during the
manufacturing

stage

in factories
,

plants
,
enterprises, etc.
The present
-
day robots
mechanically repeat
the commands that

they have been

pr
ogrammed to do

before.


That is to say
,

they
do

n
o
t have any
complex adaptive functionality.

In addition,

they cannot be
applied in society
without reflex reactions on chang
es

of the environment.

That is a real problem
for human
-
cent
e
red robots.
However,
a
s we mentioned above,
the main concept
of human
-
centered

robotics is
to understand or recognize

the intentions and
cap
abilities of the user
,

a human
[2]. This understanding

is fundamental for a
social robot
ready

to react and respond

in accordance with

the

human user

behavior
.


Consequently, t
he main goals of our research

are as follows:

4


-

Firstly,
to
develop
a new method of
recogni
zing human motions
and try
it
in practice
in
the laboratory conditions
;

-

Secondly,
to
create
a
software for the robot
needed to bu
ild the

appropriate
model of the robot’s behavior
,

using the new method of human motions
recogni
tion;

-

Lastly,
to present the results of our trial
in this research paper
.

The next section introduces
and logically describes
possible ways
to fulfill the
task
of modeling and simulation of

the

human
-
oriented robot
,

which were

attempted by us
at
Imperial College London

in the summer 2012
.


2.

Methodology

We take it for granted that different
human motions

are traditionally
recognized with
the help of
the Mechanical
Motion Capture or Marker Motion
Capture technologies. Both of these methods require additional hardware like
exoskeletons
for Mechanical technology
or special suits

with markers (lamps)
for Marker technology

to capture the motion of the human.
That means t
hat w
e
need to provide special conditions in
the
environment for
a
robot.

Nevertheless,
t
he

solution

of

our

problem

requires

a

new

method

of

human

recognizing

in

different

environments.

The application of
Markerless Motion Capture

(MMC)
technology allows r
ecognizing different movements and poses of an operator’s
body.

[3].
Apart from

other

technologies
,

MMC

technology
has the following
advantages: first of all, it
does not
acquire
any additional equipment
,
and

then
,

it
does not need

any
special environmenta
l conditions
for the robot. That is very
important for our
problem
,
since

there is no special equipment in surroundings
for the robot when it
works

in a real social environment.
Besides
, the method
allows realization of computer vision principles in the ro
bot sensor system. That
means the robot can
capture

a video image from sensors and use it in the control
system.

5


What is more, t
here
exist

two
sides

to

our
complex
problem.
The f
irst
one

deals with the
recogni
tion

of the human in the environment. The secon
d
one is
concerned with
the robot reaction form
ation
,

that is,

determination of the human
intentions and s
ubsequent action
s
.


The next two sections
introduce

not only
the description of the

above

problem
sides
but also

a
sequential
solution of
our

complex
problem.


2.1.
Markerless

Motion

Capture

T
echnology


The solution
of

the first
side
to

the
problem
is the application of the
M
arkerless Motion

Capture technology.

Emerging techniques and research in
computer vision are leading to the rapid development of t
he markerless
approach to motion capture. Markerless systems such as those developed at
Stanford, University of Maryland, MIT, and Max Planck Institute, do not require
subjects to wear special equipment for tracking. Special computer algorithms are
designe
d to allow the system to analyze multiple streams of optical input and
identify human forms, breaking them down into constituent parts for tracking.

The

technology

uses

special

Red Green Blue
-

Distance

(
RGB
-
D
)

sensors

that

receive

information

about

the

sp
atial

arrangement

of

the

objects

which

have

been

detected

in

the

work
ing

area

of

the

sensors
[
4
].
The

use

of

computer

vision

algorithms

allows

the robot to identify or recognize
the

human in the working
area of the sensors

(
F
ig
.1).

The received

information

about

the human

body

position

in

space

must

be

converted

to

a

more

suitable

structure for

further processing. This

structure

contains

the

numerical

values

of

the

fold

limbs

angles in every joint. These

angles

can

be

found

as

the

angles

between

the

skeleto
n’s vectors of the human
that can be read with RGB
-
D sensors software. Consequently, w
e get
the
information about
a

current state of
the person
, which ca
n be used as a control
action for

the robot.

6



Fig
.1.

Data obtained by
RGB
-
D
sensors

at

Personal Robot
ics Lab
,


Imperial College London



As a result, we have an instrument to get information and recognize
the
human in the environment
.
The former

will be used

as a data receiver to
determine the states and intentions of the human in the next section.


2.2.



Hie
rarchical Attentive Multiple Models for Execution and
Recognition

(HAMMER)

To determine th
e intentions of
the
human and
to form the robot reactions to
various user actions
,
we use

the

algorithm
of
HAMMER architecture,
which was
developed
by Y. Demiris and
his colleagues,

at Imperial College London (Fig.
2) [5,6].

Th
e

algorithm can generate many behaviors,
which are
implemented on the
basis of the current condition

of the
human.
Moreover
, this algorithm can
calculate the subsequent actions of the person and

decide on the need of any
robot’s action
.

HAMMER

has been

successfully used in many different
contexts; for
instance
,

to recogni
z
e and imitate a human
,
moving an object
between

two tables [7], to recogniz
e comp
lex

and single actions of

multiple
7


robots [8
]

and to predict the intention of opponents

in a real
-
time strategy game
as well [9
].


Fig
.2.
The a
lgorithm of

HAMMER

architecture

As a general
rule,
HAMMER architecture
comprises
:


1.

State of the world, for every time step
;


2.

Inverse
m
odels
;


3.

Forward
m
odels
;


4.

Action
s
i
gnals, which will be used to send commands

between inverse
and forward models
;

5.

Confidence
e
valuation
f
unction
.


An

inverse model

is a function which takes the state of the

world as input, and
an optional explicit target state. It outputs

the acti
on signals to reach the target
state, which could be

implicitly hard coded in the model or explicitly passed as
an

input parameter.

A
forward model

is defined as
a function that takes an action
signal and
outputs the predict
ed state of the world after the
signal has been executed.
The
term
forward model

has been
used in the literature to represe
nt many different
concepts.
We consider it as
an output predictor,
following the analysis in

the
research paper of Karniel

[
10
].
Pairing

together an
inv
erse model

an
d
forward
8


model
,
we obtain a system that genera
tes a hypothesis about the next
state of the
world. By combining

several inverse
-
forward pairs,
which
are

normally run in
pa
rallel, numerous hypotheses are
proposed. These hypotheses
are compared
with

the actu
al
state of the world at the nex
t time
-
step. A confidence value
is
computed for every inverse
-
forward pair, representing the
confidence on that
inverse
-
f
orward
pair being the behavior that is being currently executed.

By
repe
ating this process iteratively,

confidences for diffe
rent
behaviors

can be
observed
over time
.

The next section introduces
our

trial

results

and some
effective
applications of
the described technologies.


3.

Results

As
it
has been mentioned earlier, t
h
e

researched
solution
concerned with o
ur
complex problem
w
as

tested this summer at the Imperial College

London.
We
attempted to try

MMC technology and HAMMER architecture

in a real situation
as

a simulation of
a
human motion
,

detected
before
by

RGB
-
D sensor (Fig. 3).

The
anthropomorphous
r
obot

Nao

was programmed
with the help of

this method

by us
,

and the
whole
task
completed
successfully
.


All

motions

of

the

human

were

simulated

by

the

robot
, because of well configured Confidence evaluation
function of HAMMER
.
The configuration of this functio
n determine
d

efficiency
and appropriateness of all subsequent actions. When we
saw
how
success
ful
the
method

was,

the robot was reprogrammed to use this method as an instrument in
teacher’s activity.
The robot
Nao

taught

some
children

to

dance

some

simple

motions
.

The c
hildren were
so
excited by a

new small teacher


that

performed
dancing
movements more diligently
.
That

fact
prove
d

once again

that the social
robotics is
a
very perspective area f
or

research.
However,
th
at

simple example is
not the only poss
ible application of the method

described by us
.

9


The application of the
various

computer vision algorithms
in the method
allow
s

the recognition of the different tools, which can be used by the robot

[11]
.
It g
i
ve
s

the chance
to realize different interaction
s

with human and
environment
.

This idea will be a part of the further research.




Fig
.3.
Robot simulates movements of
an

operator



The successfully completed

work helped
us
to draw
the
conclusions, which
are outlined in the

next

section
.



4.

Conclusion

Ta
king everything into account
,
we can draw the following conclusions:

In the first place
,
the problem of motion
recognition

can be

successfully
solved by
the
application

of M
arkerless Motion

Capture technology

with
the
help of
the
computer vision algorithms

and HAMMER

architecture
.

Next, t
he a
pplication
of
MMC

technology
enables us to

form the flow of
information about the person

and
his/her

actions
.
The s
ensors
,

required to
implement th
is

technology, allow

obtaining

the
additional information about
objects

in the environment of the robot
through various recognition methods
,

10


which are
related to the field of
computer
vision.

HAMMER architecture
permits

the robot to carry out the analysis of
people

behavior, to draw
the right
conclusions about
their
status
es

and intentions,
to predict future human actions.


On top of that
, the application of the
robot control
system
,

described
in this
research paper,

helps
to build an appropriate model of
the
robot
’s

behavior.

Consequently
, t
hat

means that
the
described method

can be used as reflex
system of the robot.



5.

References

1.

S. Schaal,
The New Robotics
-
towards human
-
centered machines
,

HFSP
journal, vol. 1, no. 2, pp. 115

26, 2007.

2.

Y. Demiris,
Prediction of intent in robotics and multi
-
agent

systems
,

Cognitive processing,

vol. 8, no. 3, pp. 151

158, 2007.

3.

http://en.wikipedia.org/wiki/Motion_captue

4.

Arnaud Ramey, Víctor González
-
Pacheco, Miguel A Salichs.
Integration
of a Low
-
Cost RGB
-
D Sensor in a Social Robot for Gesture Recognition
.

6th international conference on Humanro
bot interaction HRI 11, 2011

5.

Miguel Sarabia, Raquel Ros, Yiannis Demiris.
Towards an open
-
source
social middleware for humanoid robots
, 11th IEEE
-
RAS International
Conference on Humanoid Robots, 2011

6.

Y. Demiris and B. Khadhouri,
Hierarchical Attentive Mult
iple Models for
Execution and Recognition (HAMMER)
, Robotics and Autonomous
Systems,
vol. 54, no. 5, pp. 361

369,
2006

7.

Abstraction in Recognition to Solve the Correspondence Problem for
Robot Imitation,

in Proc. of the Conf. Towards Autonomous Robotics

Syst
ems, 2004, pp. 63

70.

8.

M. F. Martins and Y. Demiris,
Learning multirobot joint action plans
from simultaneous task execution demonstrations
, in Proc. of the Intl.

11


Conf. on Autonomous Agents and Multiagent Systems, vol. 1, 2010, pp.

931

938.

9.

S. Butler and Y.

Demiris,
Partial Observability During Predictions of the
Opponent’s Movements in an RTS Game
, in Proc. of the Conf. on

Computational Intelligence and Games, 2010, pp. 46

53.

10.

A. Karniel,
Three creatures named ‘forward model’
,
Neural Networks,

vol. 15, no. 3, pp. 305

7, 2002.

11.

Y. Wu, Y. Demiris,
Learning Dynamical Representations of Tools for
Tool
-
Use Recognition
, IEEE International Conference on Robotics and
Biomimetics, 2011