arXiv:1602.08158v1  [cs.RO]  26 Feb 2016
Associative Memories and Human-Robot Social Interaction
Gabriel J. Ferrer
Department of Mathematics and Computer Science
Hendrix College
Conway, AR, USA
ferrer@hendrix.edu
Abstract
In this position paper, we discuss how the use of a
cognitive architecture based on unsupervised cluster-
ing (the Kohonen Self-Organizing Map) enables us to
meet our goals of efficient action selection in a mobile
robot. This architecture provides several opportunities
for human-robot interaction, and we discuss how its
features facilitate these interactions.
1. Why use cognitive architectures?
1.1. Research Context
In our research, we focus on developing, imple-
menting, and assessing robot learning algorithms that
select actions in constant time on resource-constrained
mobile robot platforms. As Jeff Hawkins points out
in his book On Intelligence, in “ ... half a second,
the information entering your brain can only traverse
a chain one hundred neurons long. That is, the brain
‘computes’ solutions to problems ... in one hundred
steps or fewer, regardless of how many total neurons
might be involved.” [1, p. 66] The key to the efficiency
of action selection by the brain is that “...the brain
doesn’t ‘compute’ the answers to problems; it retrieves
the answers from memory. In essence, the answers
were stored in memory a long time ago. It only takes
a few steps to retrieve something from memory.” [1, p.
68]. As the brain itself selects actions in constant time,
from our point of view it serves as a compelling source
of biological inspiration for a computational cognitive
architecture.
1.2. Self-Organizing Maps
The cognitive architecture described by Claude
Touzet for robot behavior learning [2] provides a
means of implementing these concepts on a mobile
robot. In Touzet’s formulation, the robot’s memories
are stored in a Self-Organizing Map (SOM) [3], a
type of artificial neural network. Each output node is
associated with a vector representing the “ideal input”
for that node. The output of a SOM is determined
competitively when an input vector is presented to
it. Let the distance between an input vector and the
ideal input for an output node be defined as the square
root of the sum-of-squared-differences between the
two vectors. We then say that the activated output
node is the output node whose ideal input has the
smallest distance to the input vector. The output nodes
themselves are arranged in a Cartesian grid. On each
iteration, the ideal input for the winning output node
as well as its four cardinal neighbors in the grid is
modified according to the following formula for each
affected output node i and ideal input element j, with
learning rate α. In Touzet’s formulation of the SOM,
α = 0.9 for the winning node on a given iteration, and
α = 0.4 for its four neighbors.
wij = wij + α ∗(inputij −wij)
Touzet’s implementation employed the distances
returned by a ring of 16 ultrasonic sensors as the
robot’s representation of its environment. Each SOM
node, then, was a vector of 16 distance values. Robot
goals were expressed in the same way, as a vector of
target distance values. To select an action, the robot
first finds a path from the SOM node corresponding
to the current input to a SOM node corresponding
to the goal input. This path contains a sequence of
intermediate nodes. The key is to select an action to
transition to a neighboring node that is closer to the
goal than the current node. To this end, a second SOM
was employed for action selection. This second SOM
encoded differences between pairs of input vectors,
with an association between each output node and a
robot action that was observed to cause the desired
transition.
Each SOM has a fixed number of input nodes. This
number of input nodes provides a constant upper bound
on the amount of computation necessary to retrieve
an action from memory to move the robot towards its
goal. In light of the quotations above from Hawkins,
we have a set of data structures (a cognitive architec-
ture) that has the potential to exhibit the functionality
of an aspect of human cognition. This architecture
enables computation that meets the constraints of our
research goals, namely, constant-time action selection
based on models of the environment.
1.3. Adaptations in our prototype
In our prototype implementation of this approach,
we have made a number of changes to the original
architecture. First, we have replaced the use of a
second SOM by a Markov chain. As our actions are
discrete, this greatly simplifies the implementation. As
there is a constant bound on the number of nodes in
the SOM, applying Dijkstra’s algorithm to the Markov
chain still matches our requirement for constant-time
computation. Second, we are investigating and proto-
typing alternatives to the SOM for clustering inputs.
The alternatives we are contemplating involve a dif-
ferent update cycle for the output nodes but preserve
the constant-time learning update that is essential to
our approach. Third, rather than using distance values
as our sensory inputs, we are using images from a
webcam.
2. Should cognitive architectures for social
interaction be inspired and/or limited by
models of human cognition?
For us, the key concept for social interaction is
to make available communication primitives that are
both intuitive to a human and readily processed by
the robot’s cognitive architecture. To the extent that
models of human cognition can give us insight into
appropriate communication primitives, those models
serve as useful inspiration.
3. What are the functional requirements
for a cognitive architecture to support
social interaction?
Determining the functional requirements requires us
to have a clear idea of how the anticipated social
interactions relate to the robot’s goals. In the context
of our work, there are three key motivations for social
interaction with a human:
1) We would like a human to be able to give a robot
goals in a reasonably natural fashion.
2) We would like for the robot to request help from
a human if it is having trouble achieving its
goals.
3) We would like a human to be able to intervene
and assist a robot in achieving its goals, whether
or not the robot has requested assistance.
These
motivations
provide
three
of
our
functional requirements. We also include the
following additional functional requirements:
4) Interaction with a human should not be a disori-
enting experience for the robot. It should be able
to continue with its goal-seeking behavior in a
seamless manner after an interaction.
5) The socially-obtained feedback should be readily
incorporated into the ongoing action selection
process.
6) The robot should be able to switch between
socially-informed and autonomously-determined
actions at the granularity of a single action-
selection decision cycle.
4. How do these requirements inform the
choice of the fundamental computational
structures of the architecture?
The computational structures of the architecture
described in Sections 1.2 and 1.3 are a good match for
our functional requirements. By presenting the robot
with an image, a human is able to express a goal in a
way that is semantically meaningful to a human while
being compatible with the data structure containing
the robot’s memories. This meets our first functional
requirement.
In regard to our second functional requirement, the
robot determines that it needs to request help from a
human in two different ways:
1) The process for identifying a path through the
robot’s memories produces an implicit plan. Al-
though only the first action from the plan is used
(as a new plan is generated at each decision
cycle), the length of the generated plan is used
to provide an estimate for the number of actions
needed to achieve the goal. If the robot finds it
is exceeding this estimate, it will ask for help.
2) It is possible for the robot to find itself in a state
without being able to find a path to the goal
state. Ideally, appropriate training would resolve
this, but it remains a live possibility as our data
structures are constituted. In this situation, the
robot will also ask for help.
To keep the interaction simple, the robot simply
stops moving and asks for help. At this stage, we
transition to handling our third functional requirement.
If the human observes that the robot needs assis-
tance, either from its own request or from human
observation, the human can give the robot commands.
As our architecture is structured to give the robot a
discrete set of commands (e.g. “go forward”, “spin
left”, etc.), it is straightforward for the human to do
this. The use of a Markov chain greatly simplifies these
communication episodes.
When a human gives a robot a command, this com-
mand replaces the action the robot itself would have
selected. On the next decision cycle, the robot resumes
its original action-selection process. This is how our
last three functional requirements are facilitated by our
cognitive architecture.
5. What is the primary outstanding chal-
lenge in developing and/or applying cogni-
tive architectures to social HRI systems?
In our system, the social interaction is not very
sophisticated. Speaking for ourselves, enabling the
system to handle ambiguous feedback from a human
remains an outstanding challenge. Determining the
proper questions to ask a human in order to reduce
the ambiguity is the key. We have not given this issue
much thought in the context of our current research,
although we think it could be very helpful when the
robot requests help.
6. What is a social interaction scenario
in which current cognitive architectures
would fail, and why?
The key issue for us is the ability of humans
to identify and eliminate ambiguity in conversation.
Imagine, for example, a robot given the command
“Get me a cold drink.” The cognitive architecture
needs to decompose this in a manner that depends
heavily on human context in general, and the context
of the questioner in particular. It would need to be
able to create natural-language questions to resolve
ambiguities, assuming it could properly identify the
ambiguities.
References
[1] J. Hawkins and S. Blakeslee, On Intelligence.
Owl
Books, 2004.
[2] C. Touzet, “Modeling and simulation of elementary robot
behaviors using associative memories,” International
Journal of Advanced Robotic Systems, vol. 3, no. 2, pp.
165–170, June 2006.
[3] T. Kohonen, Self-Organizing Maps, 3rd ed.
Springer,
2001.