The Crucial Role of Robot Self Awareness in HRI by MikeJenny

VIEWS: 4 PAGES: 2

									            The Crucial Role of Robot Self-Awareness in HRI

                              Manuel Birlo                                         Adriana Tapus
                     Cognitive Robotics Lab/UEI                              Cognitive Robotics Lab/UEI
                          ENSTA-ParisTech                                         ENSTA-ParisTech
                 32 Blvd Victor, 75015, Paris, France                    32 Blvd Victor, 75015, Paris, France
               manuel.birlo@ensta-paristech.fr                          adriana.tapus@ensta-paristech.fr


ABSTRACT
In this paper, we present the first steps towards a new con-
cept of robot self-awareness that can be implemented into
embodied robot systems. Our concept of “the self” is inspired
by already existing approaches and aims to provide a cog-
nitive system with meta-cognitive capabilities. We believe
that robot self-awareness is a crucial factor in the improve-
ment of HRI.

Categories and Subject Descriptors
H.1.2 [Information Systems]: Models and Principles—
User/Machine Systems
                                                                                  (a)                        (b)
General Terms
Theory                                                             Figure 1: Our approach: (a) self concept and (b) perspective
                                                                   taking

1. INTRODUCTION
   Recent developments in the field of Human-Robot Interac-         thors in [2] consider a robot to be self-aware if it has the
tion(HRI) aim to focus on tasks that require more and more         ability to focus attention to the representation of its inter-
robotic cognition capabilities. However, given the fact that       nal states. They have created a framework called ASMO
a human has an infinite number of possible interaction capa-        and showed first results by using a small humanoid robot in
bilities as well as an infinite set of desires and needs, current   a simple interaction scenario. In their approach, the robot
solutions in terms of robot’s cognition capabilities look quite    has the ability to deliberate and re-plan its behavior based
poor. None of the existing cognitive architectures are ade-        on its intentions and its concept of “the self”. However,
quate for complex and dynamic environments in which the            as far as we know, they tested their approach only for a
robot has the capability to take an infinite number of possi-       specific application whose interaction possibilities are quite
ble human-sided actions and behaviors as well as dynamical         limited. Another approach was presented in [3]. The pre-
environment changes into account.                                  sented architecture which is called GMU BICA follows a
   Let’s imagine a system that’s aware of itself and thus able     concept called schema that represents concepts in one uni-
to monitor its own capabilities and limitations. Such a sys-       versal format. Their work has been validated only in simula-
tem should also recognize situations in which its own capa-        tion. Furthermore, a promising approach named ACT-R/E
bilities are not suitable for unforseen environmental changes      (ACT-R/Embodied, an extension of ACT-R [1]) has been
and, as a consequence, adapt its own code in order to be           described in [4]. Based upon the existing cognitive architec-
able to handle those new situations. Therefore, a system           ture ACT-R [1], the authors have developed modules that
that is self-aware could theoretically be able to overcome         are able to deal with real word information such as vision
current limitations of applications comprising a finite num-        and sound. Nevertheless, to the best of our knowledge, there
ber of foreseen inputs and corresponding outputs. This is          is no concept of self-awareness in the sense of having a sys-
the main focus of this paper.                                      tem acting on a meta-level that is able to deliberate and
                                                                   re-plan its performance according to its intentions and self-
2. RELATED WORK                                                    conception.
  There is no unique definition of self-awareness and of the
way it could be transposed in robot’s behavior. The au-            3.    METHOLOGY
                                                                      Our approach of robot self-awareness has the concept of
                                                                   meta-cognition in common with the work of [2] and [3] (i.e.,
Copyright is held by the author/owner(s).
HRI’11, March 6–9, 2011, Lausanne, Switzerland.                    it was proven in the psychological and behavioral literature
ACM 978-1-4503-0561-7/11/03.                                       that human self-awareness is something that happens on a
                                                                 module determines the content of its self buffer. The con-
                                                                 tent of the self buffer represents the focus of the system’s
                                                                 attention on a meta-level. By having all the other buffer
                                                                 contents as well as ACT-R/E’s current focus of attention
                                                                 (represented by its focus buffer) “in mind”, the self is able to
                                                                 interfere about what’s happening inside ACT-R/E’s proce-
                                                                 dural module. The procedural module termines the system’s
                                                                 behavior and sets the current focus of attention (the focus
                                                                 buffer). As the self module has the capability to interfere
                                                                 in the processeses of the procedural it is able to deliberate
                                                                 and re-plan ACT-R/E’s behavior; this corresponds to our
                                                                 previous explained conception of self-awareness.
Figure 2: Combination of the ACT-R/E architecture and
our self module (figure of ACT-R/E architecture is taken          4.   EXPERIMENTAL DESIGN
from http://newhri.org/Presentations/trafton newhri.pdf)            In order to test our approach we plan on using the NAO
                                                                 humanoid robot from Aldebaran Robotics and Python ACT-
                                                                 R from Carleton Cognitive Modelling Lab (http://ccmsuite.
meta-level). Our work also focuses on the representation of      ccmlab.ca/). As a first testing environment, we create a
internal states since a simple interaction between the robot’s   small interaction scenario in which the human and the robot
external states in terms of in- and outputs via the external     are sitting in front of each other at a table (see Fig.1b). The
world is not sufficient in order for the robot to be able to act   aim of this scenario is to enable the robot to take the per-
self-aware. The robot has to create its own interpretation of    spective of the human in order to understand verbal com-
what it perceives and it has to connect this information to      mands that are ambiguous correctly. As the ability of per-
its current internal state as well as to its previous states.    spective taking leads to advanced human-robot interaction
   The main question here is how to create a concept for         [4], it is a good example to test our self concept. Fig.1b
these internal representations and a meta-cognition and how      shows a table with two balls and a wall. One of the balls
to use it with a humanoid robot. The ACT-R architecture          is hidden by a wall from the human’s perspective while the
[1] that aim to model cognitive features of the human brain      robot can see both of them. When the human ask the robot
is used as a foundation for our approach. We use the chunk       “Give me that ball” this command is ambiguous at a first
and production based concept of ACT-R as a first step to-         glance since the robot doesn’t know which ball the human is
wards self-awareness within a humanoid robot as ACT-R            speaking about because there a two identical balls. However,
offers a quite easy way to use general concepts that are in-      if the robot is able to take the perspective of the human it
spired by the human brain. Buffers in ACT-R represent the         will automatically know that the human can only see one ball
current focus of attention of a human brain, and they can be     and therefore, it will know which ball the user is referring
connected to a memory. Let’s imagine a human that thinks         to. We try to achieve this goal by using our meta-cognition
about something that he perceives right now; this informa-       concept. This is work is on progress and some preliminary
tion is stored in his “buffer”. If he thinks about something      results will be available by the time of the conference.
right now that is connected to an experience in the past,
he has to recall information from his memory. This corre-        5.   CONCLUSION
sponds to ACT-R buffers that are connected to memories.             Our current aim is to create a concept of self-awareness
We propose to represent a mental state as a chunk (i.e., a       in a humanoid robot based on the existing ACT-R cognitive
string) that consists of all the required information about      architecture. To this end, we try to develop an approach
the robot’s perception, internal state, and connections to       of meta-cognition that is able to deliberate and re-plan the
previous states in an abstract way. This also requires an ap-    robot’s behavior by focusing on representations of mental
proach that reduces the real world perception of the robot,      states while using the chunk system of ACT-R. In our future
for example its vision and speech recognition, to abstract in-   work, we will try to implement more complex HRI scenarios
formation inside its internal world (see Fig.1a). Abstraction    that focus on important needs within HRI.
is neccessary because the chuck system in ACT-R allows only
a rather simple way to store information. Meta-cognition,
which represents the self, is instantiated as an independent
                                                                 6.   REFERENCES
                                                                 [1] J. R. Anderson. ACT-R ACT-R Research Group,
unit that looks over all buffers and memory contents and
                                                                     CMU, http://act-r.psy.cmu.edu.
decides on which of these buffers it will pay attention to.
   Finally, our concept of robot self-awareness looks like in    [2] R. Novianto and M.-A. Williams. The role of attention
Fig.2 that shows a combination of the ACT-R/E architec-              in robot self-awareness. In Proc. RO-MAN 2009,
ture and our self-module (in red). Our self module has access        Toyama, Japan, 2009.
to every module and buffer of ACT-R/E so as to be able to         [3] A. Samsonovich. Universal learner as an embryo of
“be aware of” everything that is happening in the robot’s            computational consciousness. In Association for the
“brain”. On one hand, access to the ACT-R/E modules is               Advancement of Artificial Intelligence, 2007.
required in order for “the self” to retrieve information about   [4] J. G. Trafton, N. L. Cassimatis, M. D. Bugajska, D. P.
memories and possible actions, and on the other hand, access         Brock, F. E. Mintz, and A. C. Schultz. Enabling
to the ACT-R/E buffers is neccessary since the self should            effective human-robot interaction using
be able to be aware of what’s currently inside the robot’s           perspective-taking in robots. IEEE Transactions on
working memory. Based on all this information, the self              Systems, Man, and Cybernetics, 35(4):460–470, 2005.

								
To top