In this presentation, I address two major questions:
1. What is Intelligent Agent Perception?
2. How do Intelligent Agents Perceive?
In addressing the meaning of agent perception, I highlight impediments to the perceptual process and the process of situation assessment.
In addressing how agents perceive, I highlight traditional approaches to robotic perception and then the next step after sensor input, which is how sensor data (vision sensor, tactile, olfactory, haptic, etc.) is interpreted. Methods for interpretation include solutions based on Bayes' Theorem, the underpinning of many robotics algorithms; Probabilistic Algorithms; and Artificial Neural Networks.
I also discuss a current system for robotic perception, designed to accommodate more robust and complex robotic needs: using sensors in tandem with machine learning. This method is closer to mimicking the human perceptual process than previous methods. I discuss some examples of this: 1) a study in which researchers used visual sensors and an artificial neural network (ANN) for robotic perception; 2) a study in which researchers used haptic sensors and a classification algorithm, called a boosting algorithm for robotic perception; and 3) a study in which researchers used pressure sensors, an ANN, and intended to add pre-programmed models in order to facilitate robotic perception.
2. What is Intelligent Agent Perception?
Perception is the process by which intelligent agents
sense, perceive, and interpret their external worlds [8],
called domains [4]. Agents use built-in sensors in order
to obtain sensory data [2].
Intelligent Agent Perception | Molly Maymar | 7.10.2015
3. During the perceptual process, an intelligent agent
assigns meaning to incoming visual, auditory,
olfactory, and/or tactile sensory data [2]. The
assignment of meaning can be bottoms-up, by
which data significance is drawn directly from
sensation, or top-down, by which previously
perceived or programmed meanings are recalled
and assigned to current data [2]. Perception can be
top-down within a single cycle, in that the agent
relies on recent recollections of the same dataâin
the same domain and within the same problem
spaceâto inform data significance in real-time [2].
I. What is Intelligent Agent Perception?
Intelligent Agent Perception | Molly Maymar | 7.10.2015
4. Perception entails the filtering of sensory information
[2]. Through a process called attentionâby which
the agent determines how to direct its perceptual
resources [4] and select the most relevant, urgent,
or important information in its worldâthe agent
ignores some sensory data while processing others
[2].
The output of the perceptual process is a perceptâ
a categorical judgment or concept that the agent
has created as a result of the perceptual process
[11,2].
I. What is Intelligent Agent Perception?
Intelligent Agent Perception | Molly Maymar | 7.10.2015
5. âą âPerception tends to be ambiguous and noisy.â [9]
âą Noisy sensors interfere with agentsâ abilities to
perceive their worlds completely and accurately [4]
âą Dynamic worlds [4] mean that agents cannot rely on
pre-programmed models or sensory information [5]
and must undertake more complex data filtering
processes [2]
Impediments to the Perceptual Process
I. What is Intelligent Agent Perception?
Intelligent Agent Perception | Molly Maymar | 7.10.2015
6. The agent combines perceptual data; recognition
and categorization of familiar patterns to build a
model of its domain and/or the dynamics of that
domain, in order that the agent can best sense and
act [4,9,2]
An example is an agent using vision to detect
features of its world and then using these features to
determine its position within the world, as well as the
obstacles it will encounter [9]
Situation Assessment
I. What is Intelligent Agent Perception?
Intelligent Agent Perception | Molly Maymar | 7.10.2015
7. Intelligent agents sense through simple, inexpensive modalities
like temperature measurement and/or complex, expensive
modalities [4] like stereo visionâseeing in 3D by using two
cameras as âeyesâ [5]. The input from these modalities can be
combined and output as a single assessment [4,11].
How Do Intelligent Agents Perceive?
Intelligent Agent Perception | Molly Maymar | 7.10.2015
8. The most common method for acquiring sensory
data is through sensors.
Sensor data comes in two varieties [10]:
âą Data that characterizes the momentary situation
EX: camera images, laser range scans
âą Data related to a change in the situation
EX: controls, odometer readings
II. How Do Intelligent Agents Perceive?
Intelligent Agent Perception | Molly Maymar | 7.10.2015
9. Model-based robotics rely on complete and accurate,
pre-programmed models of the robot and its
environment in order for the robot to perceive and act.
This method presents challenges in dynamic or noisy
environments, which do not accurately reflect the
models [10].
Traditional Approaches
to Robotic Perception
II. How Do Intelligent Agents Perceive?
Intelligent Agent Perception | Molly Maymar | 7.10.2015
10. Behavior-based robotics rely on immediate sensor
feedback for determining a robotâs action. This method
places physical limits on perceptual process and poses
boundaries on the types of task that can be handled
using this approach [10].
II. How Do Intelligent Agents Perceive? :: Traditional Approaches
Intelligent Agent Perception | Molly Maymar | 7.10.2015
11. Bayesâ Theorem
Bayesâ Theorem is a rule that is used to determine
how an agent filters new data, based on previous
knowledge [9,8]. Bayesâ rule is an equation for
showing the relationship between a conditional
probability (the probability of event1, given that
event2 has already occurred) and its reverse form
(the probability of event2, given event1) [9].
II. How Do Intelligent Agents Perceive?
How is Sensor Data Interpreted?
Intelligent Agent Perception | Molly Maymar | 7.10.2015
12. Bayes Theorem is as follows [8]:
P(A|B) = P(B|A) * P(A)/P(B)
P(A) is the probability of A, independent of B
P(B) is the probability of B, independent of A
P(A|B) is the probability of A, given that B is true
P(B|A) is the probability of B, given that A is true
II. How Do Intelligent Agents Perceive? :: Sensor Data :: Bayesâ Rule
Intelligent Agent Perception | Molly Maymar | 7.10.2015
13. Mollyâs Example:
An agent needs to know if an object is a cat
P(cat) is the probability of an object being a cat, whether or
not it has pointy ears
P(ears) is the probability of an object having pointy ears,
whether or not it is a cat
P(cat|ears) is the probability the object is a cat, given it has
pointy ears
P(ears|cat) is the probability an object has pointy ears, given
that it is a cat
II. How Do Intelligent Agents Perceive? :: Sensor Data :: Bayesâ Rule
Intelligent Agent Perception | Molly Maymar | 7.10.2015
14. Mollyâs Example:
The agent would determine the probability of whether or not
the object is a cat using the following formula:
P(cat|ears) = P(ears|cat) * P(cat)/P(ears)
II. How Do Intelligent Agents Perceive? :: Sensor Data :: Bayesâ Rule
Intelligent Agent Perception | Molly Maymar | 7.10.2015
15. Probabilistic Algorithms
Probabilistic algorithms, also called randomized
algorithms, account for the inherent uncertainty in
intelligent agent perception [10].
Agents use sensors to input data from their
environments. As a result of sensor limitations, noise,
and environmental dynamism and unpredictability, their
input is uncertain. Probabilistic algorithms process
sensor data in a way that computes a probability
distribution over what might be the case in the agentâs
world overall, versus generating a single best guess [10].
II. How Do Intelligent Agents Perceive? :: Sensor Data
Intelligent Agent Perception | Molly Maymar | 7.10.2015
16. Probabilistic algorithms have proven successful with
mobile robot localizationâthe problem of finding out the
robotâs coordinates relative to its environmentâand
mappingâthe problem of generating maps from sensor
measurements [10].
II. How Do Intelligent Agents Perceive? :: Sensor Data :: Probabilistic Algorithms
Intelligent Agent Perception | Molly Maymar | 7.10.2015
17. Probabilistic algorithms are preferable to traditional
approaches to robotic perception, like methods
associated with model- and behavior-based
robotics, in that they are more robust in the face
of sensor limitations, noise, and environmental
dynamics. They are computationally inefficient in
that they consider entire probability densities, so
they generate approximate outputs [10].
II. How Do Intelligent Agents Perceive? :: Sensor Data :: Probabilistic Algorithms
Intelligent Agent Perception | Molly Maymar | 7.10.2015
18. Artificial Neural Networks
An artificial neural network (ANN) is an information
processing technique inspired by the way that the
human nervous system processes information. In
human brains, neurons receive signalsâelectric
impulses transmitted by chemical processesâfrom
other neurons, modify the signals, and forward them
to their connections. An artificial neuron is a
computational unit with many inputs and one output. It
forwards signals, or âfiresâ, based on the input pattern
received [5].
II. How Do Intelligent Agents Perceive? :: Sensor Data
Intelligent Agent Perception | Molly Maymar | 7.10.2015
19. ANNs are comprised of interconnected firing
neurons capable of being trained. ANNs consist of
three layers: an input layer, a hidden layer, and an
output layer. Each sensory input arrives in an input
node in the input layer. This input is then provided to
every node in the hidden layer, where each
connection is assigned a weight. The output value of
every hidden node is then provided to the output
node, where connections are again weighted and
then output as a percept [5,11,2].
II. How Do Intelligent Agents Perceive? :: Sensor Data :: Artificial Neural Networks
Intelligent Agent Perception | Molly Maymar | 7.10.2015
20. II. How Do Intelligent Agents Perceive?
Machine learning is the science of pattern
recognition and computational learning theory in
artificial intelligence. Machine learning focuses on
the construction and study of algorithms that learn
from and make predictions on data [6].
In order to ensure more robust robot perception in
light of unpredictable, dynamic, and noisy real-
world scenarios, scientists pair machine learning
approaches with biologically-inspired sensor
systems [5].
Sensors + Machine Learning
Intelligent Agent Perception | Molly Maymar | 7.10.2015
21. Some Examples:
Visual Sensors + ANN
Researchers working with the iCub humanoid robot
sought to show that artificial neural networks can be
trained to allow humanoid robot spatial perception in
Cartesian (3D) coordinates. Spatial perception is the
detection and localization of objects in the robotâs
world [5].
II. How Do Intelligent Agents Perceive? :: Sensors + Machine Learning
Intelligent Agent Perception | Molly Maymar | 7.10.2015
22. The iCub, moving autonomously in a world, obtained visual
information from two cameras. The ANN provided the robot
with the ability to estimate the position of camera-identified
objects relative to itself in this 3D space. The ANN was
comprised of three separate neural networksâone trained
to predict object position and robot poses along the X axis,
one along the Y, and one along the Z. Visual input regarding
image coordinates and robot poses arrived in each input
node and each output layer provided an estimate of the
object and robotâs position along one Cartesian axis [5].
The iCub showed that by training ANNs, artificial agents can
learn the skill of estimating object locations based on visual
perception [5].
II. How Do Intelligent Agents Perceive? :: Sensors + Machine Learning :: Visual + ANN
Intelligent Agent Perception | Molly Maymar | 7.10.2015
23. Haptic Sensors + Classification Algorithms
Researchers working with the ALoF robot
(autonomous legged robot on four legs) used a
classification algorithmâan algorithm that selects
the best possible hypothesis based on observed
data [1]âto identify the geometric texture shape
and surface properties of a robotâs selected
footholds based on haptic information, or tactile
feedback [3].
II. How Do Intelligent Agents Perceive? :: Sensors + Machine Learning
Intelligent Agent Perception | Molly Maymar | 7.10.2015
24. Researchers bisected each of ALoFâs legs and placed
force sensor resisters at the bisection in order to
capture haptic feedback from ALoFâs knee joint
movement when it was trodding on different surfaces.
The resistors registered the normal force (the force
supporting the weight of ALoFâs âfemurâ on its âtibiaâ, or
tibia on the floor) and moments (the force necessary to
rotate the tibia on the knee/axis) in the joint. ALoF
âwalkedâ and âscratchedâ on artificial terrain textured
separately in four geometrical shapes, ranging from a
Teflon-like surface to a rocky-feeling one [3].
II. How Do Intelligent Agents Perceive? :: Sensors + Machine Learning :: Haptic + Classification
Intelligent Agent Perception | Molly Maymar | 7.10.2015
25. When trained in real-world walking force and contact
force measurement data samples, a special classification
algorithm, called a boosting algorithmâwhich combines
a weighted ensemble of weak classifiers into a stronger
one, and is slightly more accurate than a probabilistic
algorithmâwas able to identify surface types based on
the haptic feedback [3].
II. How Do Intelligent Agents Perceive? :: Sensors + Machine Learning :: Haptic + Classification
Intelligent Agent Perception | Molly Maymar | 7.10.2015
26. ALoF showed that robots can use haptics, or tactile
sensory information, to identify local surface properties.
This will help robots characterize their environment
while walking through it, especially when detailed
environmental models have not been supplied in
advance [3].
II. How Do Intelligent Agents Perceive? :: Sensors + Machine Learning :: Haptic + Classification
Intelligent Agent Perception | Molly Maymar | 7.10.2015
27. Pressure Sensors + ANN + Offline Planning
Through the integration of sensor-based
techniques, offline planning, and an ANN,
researchers sought to inform the development of a
new breed of safe and intelligent robots that could
share a common workspace with humans [7].
II. How Do Intelligent Agents Perceive? :: Sensors + Machine Learning
Intelligent Agent Perception | Molly Maymar | 7.10.2015
28. Conventional offline planning techniquesâ
preprogrammed automation tasks reliant on
complete knowledge of the static environmentâ
require complete knowledge about the static
environment. They do not account for human
motion, which is unpredictable, and unstructured or
time-varying interactive environments [7].
II. How Do Intelligent Agents Perceive? :: Sensors + Machine Learning :: Sensors + ANN + Offline
Intelligent Agent Perception | Molly Maymar | 7.10.2015
29. In this study, humans stood on mats consisting of
pressure-activated nodes. Data collected by the
nodes was clustered into subsets corresponding to
each foot using a self-organizing map (SOM)âa
special class of ANN that discovers salient features
in time-varying data sets without any prior training.
Using these subsets, each humanâs body
orientation and location were derived and paired
with average human body dimensions in order to
obtain a model of each person [7].
II. How Do Intelligent Agents Perceive? :: Sensors + Machine Learning :: Sensors + ANN + Offline
Intelligent Agent Perception | Molly Maymar | 7.10.2015
30. Researchers will use online information sensed by
the mats and interpreted by the SOM to modify
robotsâ predefined motion paths, in order that the
robots can reach their intended goals, or preform
their automated functions, while avoiding obstacles.
This will allow robots to collaborate safely and
amicably with humans [7].
II. How Do Intelligent Agents Perceive? :: Sensors + Machine Learning :: Sensors + ANN + Offline
Intelligent Agent Perception | Molly Maymar | 7.10.2015
31. References
[1] Castelli, V. (2005). Lecture 2. Retrieved from: http://www.ee.columbia.edu/~vittorio/Lecture-2.pdf
[2] Goertzel, B., & Wang, P. (2007). A foundational architecture for artificial
general intelligence. Advances in Artificial General Intelligence: Concepts, Architectures and Algorithms, 6, 36.
[3] Höpflinger, M., Remy, C. D., Hutter, M., Spinello, L., & Siegwart, R. (2010, May). Haptic terrain classification for
legged robots. In Robotics and Automation (ICRA), 2010 IEEE International Conference on (pp. 2828-2833). IEEE.
[4] Langley, P., Laird, J. E., & Rogers, S. (2009). Cognitive architectures: Research issues and
challenges. Cognitive Systems Research, 10(2), 141-160
[5] Leitner, J., Harding, S., Frank, M., Forster, A., & Schmidhuber, J. (n.d.). Artificial neural networks for spatial
perception: Towards visual object localisation in humanoid robots. The 2013 International Joint Conference on
Neural Networks (IJCNN).
[6] Machine Learning (2015, July 3). Wikipedia. Retrieved from:
https://en.wikipedia.org/wiki/Machine_learning
[7] Najmaei, N., & Kermani, M. R. (2011). Applications of artificial intelligence in safe humanârobot
interactions. Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on, 41(2) 448-459.
[8] Nilsson, N. J. (2009). The quest for artificial intelligence. Cambridge University Press.
[9] Poole, D. L., & Mackworth, A. K. (2010). Artificial Intelligence: foundations of computational agents. Cambridge
University Press.
[10] Thrun, S. (2000). Probabilistic algorithms in robotics. Ai Magazine, 21(4), 93.
[11] Wang, P. (2014). Perception. Introduction to artificial intelligence [online lecture notes]. Retrieved from:
http://www.cis.temple.edu/~pwang/3203-AI/Lecture/IO-2.htm