A STUDY OF A CONSCIOUS ROBOT

An Attempt to Perceive the Unknown

Soichiro Akimoto and Junichi Takeno

Robot Science Laboratory, Meiji University, 1-1-1 Higashimita, Tama-ku, Kawasaki-shi, Kanagawa, 214-8571, Japan

Keywords: Cognition and Behavior, Consciousness, Pain of the heart, Detecting unknown, Neural Networks, Robotics.

Abstract: The authors are developing a robot that has consciousness, emotions and feelings like humans. As we make

progress in this study, we look forward to deepening our understanding of human consciousness and

feelings. So far, we have succeeded in representing consciousness in a robot, evolved this conscious system

by adding the functions of emotions and feelings, and successfully performed mirror image cognition

experiments using the robot. Emotions and feelings in a robot are, like those of humans, basic functions that

can enable a robot to avoid life-threatening situations. We believe that consistency of cognition and

behavior generates consciousness in a robot. If we can detect what happens in the robot when this

consistency is lost, we may be able to develop a robot that is capable of discriminating between what it has

learned and what it has not learned. Furthermore, anticipate that the robot may eventually be able to feel a

“pain of the heart.” This paper reports on autonomous detection by a robot of non-experienced phenomena,

or awareness of the unknown, using the function of consciousness embedded in the robot. If the robot is

capable of detecting unknown phenomena, it may be able to continually accumulate experiences by itself.

1 INTRODUCTION

We have already defined our belief that

consciousness is generated from consistency of

cognition and behavior. Based on this definition, we

have developed a conscious system using recurrent

neural networks (Suzuki, 2005). We then

incorporated this conscious system into a robot, and

the robot performed successfully in experiments on

imitation behavior and mirror image cognition, or

so-called mirror tests (Igarashi, 2007) ,(Takeno,

2008).

The present paper reports on our development of

a robot that is capable of detecting a situation that it

has never experienced before and representing a

“heartfelt discomfort” or anxiety such as

unpleasantness (a feeling) or pain (an emotion)

according to the stimuli occurring in such a

situation.

By developing a robot having these functions

and creating a new model of the human brain based

on the robot, the authors would like to demonstrate

that it is possible to provide medical information on

treatment of human brain diseases and develop a

robot having a function very close to that of human

consciousness in the near future.

2 PAIN OF THE HEART

This chapter first introduces phantom pain. A patient

who has had a limb amputated due to injury or

disease may feel as though the limb still existed as it

was before amputation and may feel pain in the

nonexistent limb. (

Ramachandran, 1999)

The authors believe that the patient feels pain at

the invisible and nonexistent limb because the brain

cannot cognize the actual situation correctly, or what

is cognized by the brain is different from the real

world. The authors hypothesized that this cognitive

gap generates unpleasant, and leads to a feeling of

pain. If this hypothesis is true, we might then say

that phantom pain—a feeling of unpleasant—occurs

when the consistency of cognition and behavior (the

definition of consciousness used by the authors used

throughout our study) is lost.

The authors define the unpleasant and pain that

occur when consistency of cognition and behavior is

lost as “pain of the heart” for the purpose of

elucidating phantom pain. We call this “pain of the

heart” because it is the working of the reason

subsystem whereas bodily pain is a result of physical

stimuli, such as when the body collides with some

external object.

498

Akimoto S. and Takeno J. (2010).

A STUDY OF A CONSCIOUS ROBOT - An Attempt to Perceive the Unknown.

In Proceedings of the 2nd International Conference on Agents and Artiﬁcial Intelligence - Artiﬁcial Intelligence, pages 498-501

DOI: 10.5220/0002704904980501

 SciTePress

Unpleasant arising from perception of the

unknown, as discussed in this paper, is one of the

“pain of the heart.”

3 STRUCTURE OF A MONAD

A MoNAD is a conscious module developed by the

authors. It is a computational model for studying

consciousness using neural networks (NNs).

A MoNAD consists of a cognition system (a),

behavior system (b), primitive representation region

(c), symbolic representation region (d) which is a

common area shared by the cognition and behavior

systems, and input/output units (S and M) as shown

in Fig. 1.

External information (S) enters the MoNAD at

the input unit and normally goes through p1,

cognition system (a), p5 and symbolic representation

region (d). The symbolic representation region (d)

represents the state of the self and the other as

determined by the input information and language

labels RL and BL. RL is a language label for

information that is currently being cognized while

BL is the label for the behavior that is to be

performed next. RL is neuro-calculated from

information S, cognition information BL one-step

prior and behavior information M’ one-step prior.

Information of the cognition representation RL is

constantly being copied to the behavior

representation BL in the absence of any other

information input from the higher-level modules into

BL. In the symbolic representation region, the state

of the self and the other is labeled separately. At

present, all neural networks used in the study are

top-down networks, and supervised learning is used

because the main topic of this present study is the

function of human consciousness.

The behavior command from BL basically passes

through p6, behavior system (b), p2 and output unit

M. M is the information output from the MoNAD. It

is also transmitted to the lower-level MoNAD

modules. Information M is neuro-calculated from

information passing through BL and p6, input

information S and behavior information M’ one-step

prior.

One of the features of our MoNAD is that the

cognition system and behavior system share the

primitive representation region (c). Thanks to this

primitive representation region, the system learns

behaviors when cognizing and learns cognition

when behaving. The closed loop information circuit

passing through the primitive representation region

and symbolic representation region makes it possible

for the MoNAD to generate artificial inner thoughts

Figure 1: Structure of MoNAD.

and expectations in the robot. Another feature of the

MoNAD is that the state of the self is understood by

somatic sensation M’.

4 STRUCTURE OF THE

CONSCIOUS SYSTEM

The conscious system comprises three subsystems:

reason, emotion-feelings and association. All

subsystems have a similar MoNAD structure. The

reason and emotion-feelings subsystems have a

hierarchical structure formed of MoNADs. The

reason subsystem uses input information to cognize

the outer world and the state of the self, and outputs

behavior from the output unit. The emotion-feelings

subsystem employs information from the condition

of the robot’s body and represents emotions and

feelings. Two feeling MoNADs, one representing

pleasant and one unpleasant, form the top layer of

the emotion subsystem. Information from the reason

subsystem (cognized representation) is also used in

the final determination of a state of pleasant or

unpleasant.

The association subsystem uses information from

the reason and emotion-feelings subsystems and

integrates the representations of both subsystems.

The association subsystem learns, by using the result

of integration, and outputs a behavior capable of

eventually realizing pleasant for the self from the

reason subsystem.

Figure 2: A Model of the Conscious System.

A STUDY OF A CONSCIOUS ROBOT - An Attempt to Perceive the Unknown

499

5 PERCEPTION OF THE

UNKNOWN

To reproduce the situation where consistency of

cognition and behavior is lost, or when there is a gap

between the cognition and behavior, the authors

conducted color identification experiments with the

robot. The robot learns colors beforehand and

discriminates between learned and non-learned

colors. Two situations may be expected of the robot:

one is where the pre-learned colors are cognized by

the MoNADs smoothly and the other is where the

non-learned colors are not cognized smoothly.

The robot pre-learns green, red and blue colors.

In the experiment, the robot is shown black,

which is a non-learned information. We selected

black because the robot would easily identify it to be

unknown since black has no color information at all.

We intend to try other complex colors later.

Images are taken by the camera embedded in the

robot. The color is analyzed using the RGB values

of the image. One image is taken about every

second. This image-taking continues until the robot

identifies the color. The robot represents pleasant

from its feeling subsystem when it succeeds in

identifying the color. Unpleasant is represented from

the feeling subsystem when the robot fails to

identify the color.

The feelings of the robot are classified into 4

types: much unpleasant, unpleasant, pleasant, much

pleasant. Pleasant is determined when the robot

identifies the color while unpleasant is determined

when it fails.

Figure 3: Development of the Conscious System with a

MoNAD Structured Neural Network.

6 DETECTION OF GAPS

BETWEEN COGNITION AND

BEHAVIOR (DIFFERENCES

BETWEEN RL AND BL)

To detect the gaps between cognition and behavior,

we calculated the gaps in bit patterns between RL

and BL (to be discussed later) in the primitive

representation region. We expected that a gap

between cognition and behavior would be present in

the primitive representation region when consistency

between cognition and behavior was lost because

information on both cognition and behavior (inputs

of external conditions, somatic sensations, and the

result of the behavior one-step prior that is used to

determine behavior) coexists in this region.

Calculating the bit pattern gap (or simply, the P-

error) involves finding the sum of the mean squared

error of the bits of RL and BL (note that each RL

and BL have a 5-bit pattern). The first 3 bits

represent the identified color and the last 2 bits are

the behavior of the self.

We determine that no gap exists between

cognition and behavior when the P-error is less than

a certain value (0.0002).

Figure 4: Description of RL and BL Nodes.

7 ROBOT EXPERIMENTS

An e-puck commercial mini-robot was used in our

experiments. We installed color LEDs on the

standard robot to be able to visualize its internal

state.

The experiment was conducted in the following

manner. First, colors are presented to the robot

which is equipped with a camera. The robot views

the color and behaves as it has learned. The robot

then responds in one of four ways: advances (upon

seeing green), stops (red), backs up (blue), or

oscillates back and forth (non-learned color).

First, we placed a color in front of the camera of

the robot to have the robot identify the color and

respond according to what it has learned. We

checked the feeling (pleasant or unpleasant) and

representation of the robot as it identified the color

successfully. Another color was then shown to the

robot to continue the experiment.

The robot advances when green is shown. Green

is a learned color and the robot cognizes it relatively

quickly. When successful, the robot feels pleasant

and the relevant LEDs light up.

The robot stops upon seeing red. Red is also a

learned color like green and the robot cognizes it

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

500

relatively quickly. When successful, the robot feels

pleasant and the relevant LEDs light up.

The result is the same when blue is shown to the

robot, and the robot backs up and feels pleasant.

When a non-learned color (black in this case) is

shown, the robot oscillates back and forth. The robot

performs many more computations than when

viewing a learned color, as it attempts to cognize the

unknown color. The robot eventually represents

unpleasant in its emotion subsystem as the number

of computations increases while it tries to identify

the unknown color. The ‘pain’ LED lights up on the

robot as it feels very unpleasant. The robot identifies

the color.

Figure 5: Views of Experiments.

8 RESULTS OF EXPERIMENTS

AND OBSERVATIONS

For the three learned colors—green, red and blue—

the consistency of cognition and behavior was

established in the robot after 3 or 4 recursive

computations with the MoNAD, and the colors were

then successfully identified. The robot showed a

pleasant state when successfully identifying colors.

When presented with non-learned information

(black), the robot required 5 to 10 recursive

computations to identify the information, and the

robot showed a very unpleasant state.

Figure 6 shows the P-error between RL and BL

when the color was changed from red to green, and

from blue to black, respectively, as well as from

green, blue and black to red.

Thanks to our embedded conscious system used

in these experiments, the robot showed, by

representing unpleasant, that it was capable of

perceiving non-learned information or a condition

that it had never experienced before.

This was our demonstration of perceiving the

unknown using a robot.

Figure 6: RL, BL and P-error.

9 SUMMARY

This paper discussed consciousness, emotions and

feelings. Based on a belief that consistency between

cognition and behavior generates consciousness, the

authors believe that pain and unpleasant, which

occur when consistency between cognition and

behavior is lost, comprise “pain of the heart.”

REFERENCES

V, S, Ramachandran., Sandra, Blakeslee., 1999. Phantoms

in the Brain: Probing the Mysteries of the Human

Mind. HARPER PERENNIAL, pp.3-328.

Suzuki, T., Inaba, K., Takeno, J., 2005. Conscious Robot

that Distinguishes between Self and Others and

implements Imitation Behavior. IEA/AIE, pp.101-110.

Igarashi, R, Suzuki, T, Shirakura, Y, Takeno, J., 2007.

Realization of an emotional robot and imitation

behavior. The 13th IASTEND International

Conference on Robotics and Applications, RA2007,

pp.135-140.

Takeno, J., 2008. A Robot Succeeds in 100% Mirror

Image Cognition. INTERNATIONAL JOURNAL ON

SMART SENSING AND INTELLIGENT SYSTEMS,

VOL. 1, NO. 4, DECEMBER 2008, pp.891-911.

Torigoe, S, Igarashi, R, Komatsu, T, Takeno, J., 2009.

Creation of a Robot that is Conscious of Its

Experiences. ICONS2009 September 21-23, 2009

Istanbul, Turkey.

A STUDY OF A CONSCIOUS ROBOT - An Attempt to Perceive the Unknown

501