Situated Robotics

Arya MirAI and Robotics

Oct 14, 2011 (5 years and 10 months ago)

777 views

Situated robotics is the study of robots embedded in complex, often dynamically changing environments. The complexity of the robot control problem is directly related to how unpredictable and unstable the environment is, to how quickly the robot must react to it, and to how complex the task is.

This article appears in the Encyclopedia of
Cognitive Science, Nature Publishers Group,
Macmillian Reference Ltd., 2002.
Level 2
Situated Robotics

Maja J Matarić, University of Southern California, Los Angeles, CA, USA


CONTENTS

Introduction Comparison and discussion

Types of robot control
______________________________________________________________________________


Situated robotics is the study of robots embedded in and stability of the environment largely
complex, often dynamically changing environments. The determines the complexity of the robot that
complexity of the robot control problem is directly related must exist in it; situated robots present a
to how unpredictable and unstable the environment is, to significant challenge for the designer.
how quickly the robot must react to it, and to how Embodiment is a concept related to
complex the task is. situatedness. It refers to having a physical body
interacting with the environment through that
body. Thus, embodiment is a form of
INTRODUCTION
situatedness: an agent operating within a body is

situated within it, since the agent s actions are
Robotics, like any concept that has grown and
directly and strongly affected by it. Robots are
evolved over time, has eluded a single, unifying
embodied: they must possess a physical body in
definition. What once used to be thought of as
order to sense their environment, and act and
a replacement for repetitive, manual labor, has
move in it. Thus, in principle every robot is
grown into a large field that includes
situated. But if the robot s body must exist in a
applications as diverse as automated car
complex, changing environment, the
assembly, space exploration and robtic soccer.
situatedness, and thus the control problem, are
Although robotics includes teleoperation, in
correspondingly complex.
which the robot itself may be merely a remotely-

operated body, in most interesting cases the
system exists in the physical world, typically in TYPES OF ROBOT CONTROL
ways involving movement. Situated robotics,
focuses specifically on robots that are embedded Robot control is the process of taking
in complex, challenging, often dynamically information about the environment, through the
changing environments. Situatedness refers to robot’s sensors, processing it as necessary in
existing in, and having one’s behavior strongly order to make decisions about how to act, and
affected by such an environment. Examples of then executing those actions in the
situated robots include autonomous robotic cars environment. The complexity of the
on the highway or on city streets (Pomerleau environment, i.e., the level of situatedness,
1989), teams of interacting mobile robots clearly has a direct relation to the complexity of
(Mataric’ 1995), a mobile robot in a museum full the control (which is directly related to the task
of people (Burgard et al, 2000). Examples of of the robot): if the task requires the robot to
unsituated robots, which exist in fixed, react quickly yet intelligently in a dynamic,
unchanging environments, include assembly challenging environment, the control problem is
robots operating in highly structured, strongly very hard. If the robot need not respond
predictable environments. The predictability quickly, the required complexity of control is
Situated Robotics 2
reduced. The amount of time the robot has to
Reactive Control
respond, which is directly related to its level of

situatedness and its task, influences what kind of
Don’t think, react! Reactive control is a
controller the robot will need.
technique for tightly coupling sensory inputs
While there are infinitely many possible robot
and effector outputs, to allow the robot to
control programs, there is a finite and small set
respond very quickly to changing and
of fundamentally different classes of robot
unstructured environments (Brooks, 1986).
control methodologies, usually embodied in
Reactive control is often described as its
specific robot control architectures. The four
biological equivalent: stimulus-response . This
fundamental classes are: reactive control ( don t
is a powerful control method: many animals are
think, react ), deliberative control ( think, then
largely reactive. Thus, this is a popular approach
act ), hybrid control ( think and act
to situated robot control. Its limitations include
independently in parallel ), and behavior-based
the robot’s inability to keep much information,
control ( think the way you act ).
form internal representations of the world
Each of the approaches above has its
(Brooks 1991a), or learn over time. The
strengths and weaknesses, and all play important
tradeoff is made in favor of fast reaction time
and successful roles in certain problems and
and against complexity of reasoning. Formal
applications. Different approaches are suitable
analysis has shown that for environments and
for different levels situatedness, the nature of
tasks that can be characterized a priori, reactive
the task, and the capabilities of the robot, in
controllers can be shown to be highly powerful,
terms of both hardware and computation.
and, if properly structured, capable of optimal
Robot control involves the following
performance in particular classes of problems
unavoidable trade-offs:
(Schoppers 1987; Agre and Chapman 1990).

But in other types of environments and tasks,
Thinking is slow, but reaction must often be
where internal models, memory, and learning are
fast.
required, reactive control is not sufficient.


Thinking allows looking ahead (planning) to
Deliberative Control
avoid bad actions. But thinking too long can

be dangerous (e.g., falling off a cliff, being
Think, then act. In deliberative control, the
run over).
robot uses all of the available sensory

information, and all of the internally stored
To think, the robot needs potentially a great
knowledge, to reason about what actions to take
deal of accurate information. Information
next. The reasoning is typically in the form of
must therefore actively be kept up to date.
planning, requiring a search of possible state-
But the world keeps changing as the robot is
action sequences and their outcomes. Planning,
thinking, so the longer it thinks, the more
a major component of artificial intelligence, is
inaccurate its solutions.
known to be a computationally complex

problem. The robot must construct and then
Some robots do not think at all, but just
evaluate potentially all possible plans until it
execute preprogrammed reactions, while others
finds one that will tell it how to reach the goal,
think a lot and act very little. Most lie between
solve the problem, or decide on a trajectory to
these two extremes, and many use both thinking
execute. Planning requires the existence of an
and reaction. Let us review each of the four
internal representation of the world, which
major approaches to robot control, in turn.
allows the robot to look ahead into the future,

to predict, the outcomes of possible actions in
various states, so as to generate plans. The
Situated Robotics 3
internal model, thus, must be kept accurate and al.,1983; Firby, 1987; Arkin, 1989; Malcolm and
up-to-date. When there is sufficient time to Smithers, T., 1990; Connell, 1991; Gat, 1992).
generate a plan, and the world model is accurate,
this approach allows the robot to act
Behavior-Based Control
strategically, selecting the best course of action

for a given situation. However, being situated in
Think the way you act. Behavior-based control
a noisy, dynamic world usually makes this
draws inspiration from biology, and tries to
impossible. Thus, few situated robots are purely
model how animals deal with their complex
deliberative.
environments. The components of behavior-

based systems are called behaviors: these are
Hybrid Control
observable patterns of activity emerging from
interactions between the robot and its
Think and act independently in parallel. environment. Such systems are constructed in a
Hybrid control combines the best aspects of bottom-up fashion, starting with a set of survival
reactive and deliberative control: it attempts to behaviors, such as collision-avoidance, which
combine the real-time response of reactivity couple sensory inputs to robot actions.
with the rationality and efficiency of Behaviors are added to provide more complex
deliberation. The control system contains both capabilities, such as wall following, target
a reactive and a deliberative component, and chasing, exploration, and homing. New
these must interact in order to produce a behaviors are introduced into the system
coherent output. This is difficult: the reactive incrementally, from the simple to the more
component deals with the robot’s immediate complex, until their interaction results in the
needs, such as avoiding obstacles, and thus desired overall capabilities of the robot. Like
operates on a very short time-scale and uses hybrid systems, behavior-based systems may be
direct external sensory data and signals; while organized in layers, but unlike hybrid systems,
the deliberative component uses highly the layers do not differ from each other greatly
abstracted, symbolic, internal representations of in terms of time-scale and representation used.
the world, and operates on a longer time-scale. All the layers are encoded as behaviors,
As long as the outputs of the two components processes that take inputs and send outputs to
are not in conflict, the system requires no each other.
further coordination. However, the two parts of Behavior-based systems and reactive systems
the system must interact if they are to benefit share some similar properties: both are built
from each other. Thus, the reactive system incrementally, from the bottom up, and consist
must override the deliberative one if the world of distributed modules. However, behavior-
presents some unexpected and immediate based systems are fundamentally more powerful,
challenge; and the deliberative component must because they can store representations (Matarić,
inform the reactive one in order to guide the 1992), while reactive systems cannot do so.
robot toward more efficient trajectories and Representations in behavior-based systems are
goals. The interaction of the two parts of the stored in a distributed fashion, so as to best
system requires an intermediate component, match the underlying behavior structure that
whose construction is typically the greatest causes the robot to act. Thus if a robot needs to
challenge of hybrid design. Thus, hybrid plan ahead, it does so in a network of
systems are often called three layer systems , communicating behaviors, rather than a single
consisting of the reactive, intermediate, and centralized planner. If a robot needs to store a
deliberative layers. A great deal of research has large map, the map is likely to be distributed
been conducted on how to designing these over multiple behavior modules representing its
components and their interactions (Giralt et components, like a network of landmarks, as in
Situated Robotics 4
(Matarić, 1990), so that reasoning about the map demanding very fast responses; this capability
can be done in an active fashion, for example comes at the price of not looking into the past
using message passing within the landmark or the future. Reactive systems are also a
network. Thus, the planning and reasoning popular choice in highly stochastic
components of the behavior-based system use environments, and environments that can be
the same mechanisms as the sensing and action- properly characterized so as to be encoded in a
oriented behaviors, and so operate on a similar reactive input-output mapping. Deliberative
time-scale and representation. In this sense, systems, on the other hand, are the best choice
thinking is organized in much the same way as for domains that require a great deal of strategy
acting . and optimization, and in turn search and
Because of their capability to embed planning. Such domains, however, are not
representation and plan, behavior-based control typical of situated robotics, but more so of
systems are not an instance of behaviorism as scheduling, game playing, and system
the term is used in psychology: behaviorist configuration, for instance. Hybrid systems are
models of animal cognition involved no internal well suited for environments and tasks where
representations. Some argue that behavior-based internal models and planning can be employed,
systems are more difficult to design than hybrid and the real-time demands are few, or
systems, because the designer must directly take sufficiently independent of the higher-level
advantage of the dynamics of interaction rather reasoning. Thus, these systems think while they
than minimize interactions through traditional act. Behavior-based systems, in contrast, are
system modularity. However, as the field is best suited for environments with significant
maturing, expertise in complex system design is dynamic changes, where fast response and
growing, and principled methods of distributed adaptivity are necessary, but the ability to do
modularity are becoming available, along with some looking ahead and avoid past mistakes is
behavior libraries. Much research has been required. Those capabilities are spread over the
conducted in behavior-based robot control. active behaviors, using active representations if
necessary (Matarić, 1997). Thus, these systems
COMPARISION AND DISCUSSION think the way they act.
We have largely treated the notion of
Behavior-based systems and hybrid systems situated robotics here as a problem: the need
have the same expressive and computational for a robot to deal with a dynamic and
capabilities: both can store representations and challenging environment it is situated in.
look ahead. But they work in very different However, it has also come to mean a particular
ways, and the two approaches have found class of approaches to robot control, driven by
different niches in mobile robotics problem and the requirements of situatedness. These
application domains. For example, hybrid approaches are typically behavior-based,
systems dominate the domain of single robot involving biologically-inspired, distributed, and
control, unless the domain is so time-demanding scalable controllers that take advantage of a
that a reactive system must be used. Behavior- dynamic interaction with the environment rather
based systems dominate the domain of multi- than of explicit reasoning and planning. This
robot control, because the notion of collections overall body of work has included research and
of behaviors within the system scales well to contributions in single-robot control for
collections of such robots, resulting in robust, navigation (Connell, 1990; Matarić, 1990),
adaptive group behavior. models of biological systems ranging from
In many ways, the amount of time the robot sensors to drives to complete behavior patterns
has (or does not have) determines what type of (Beer, 1990; Cliff, 1990; Maes, 1990; Webb,
controller will be most appropriate. Reactive 1994; Blumberg, 1996), robot soccer (Asada et
systems are the best choice for environments al., 1994; Werger, 1999; Asada et al., 1998),
Situated Robotics 5
cooperative robotics (Matarić, 1995; Kube, Intelligent Robots and Systems, pp.917-924.
1992; Krieger et al., 2000; Gerkey and Matarić, Munich: IEEE Computer Society Press.
2000), and humanoid robotics (Brooks and Beer R, Chiel H and Sterling L (1990) A
Stein, 1994; Scassellati, 2000; Matarić, 2000). In biological perspective on autonomous agent
all of these examples, the demands of being design. Robotics and Autonomous Systems 6: 169-
situated within a challenging environment while 186.
attempting to safely perform a task (ranging Blumberg B (1996) Old Tricks, New Dogs: Ethology
from survival, to achieving the goal, to winning and Interactive Creatures. PhD thesis, MIT.
a soccer match) present a set of challenges that Brooks A (1991a) Intelligence without
require the robot controller to be real-time, representation. Artificial Intelligence 47: 139-
adaptive, and robust. 160.
The ability to improve performance over Brooks A (1991b) Intelligence without reason.
time, in the context of a changing and dynamic In: Proceedings, International Joint Conference on
environment, is also an important area of Artificial Intelligence Sydney, Australia, pp.569-
research in situated robotics. Unlike in classical 595. Cambridge, MA. MIT Press.
learning, where the goal is to optimize Brooks R (1986) A robust layered control
performance over a typically long period of system for a mobile robot. IEEE Journal of
time, in situated learning the aim is to adapt Robotics and Automation 2: 14-23.
relatively quickly, achieving greater efficiency in Brooks R and Stein L (1994) Building brains for
the light of uncertainty. Models from biology bodies. Autonomous Robots 1: 7-25.
are often considered, and reinforcement learning Burgard W, Cremers A, Fox D et al. (2000)
models are particularly popular, given their Experiences with an interactive museum tour-
ability to learn directly from environmental guide robot. Artificial Intelligence 114: 32-149.
feedback. Cliff D (1990) The computational hoverfly; a
This area continues to expand and address study in computational neuroethology. In:
increasingly complex robot control problems. Meyer J-A and Wilson S (eds) Proceedings,
There are several good surveys on situated Simulation of Adaptive Behavior, pp. 87-96.
robotics which provide more detail and Cambridge, MA: MIT Press.
references (e.g. Brooks, 1991b; Matarić, 1998). Connell J (1990) Minimalist Mobile Robotics: A
Colony Architecture for an Artificial Creature.
References Boston, MA: Academic Press.
Connell J (1991) SSS: a hybrid architecture
Agre P and Chapman D (1990) What are plans applied to robot navigation. In: Proceedings,
for? In: Maes P (ed) Designing Autonomous International Conference on Robotics and
Agents, pp.17-34. Cambridge, MA: MIT Press. Automation, Nice, France, pp. 2719-2724. Los
Arkin R (1989) Towards the unification of Alamitos, CA: AAAI/MIT Press.
navigational planning and reactive control. In: Firby J (1987) An investigation into reactive
Proceedings, American Association for Artificial planning in complex domains. In: Proceedings of
Intelligence Spring Symposium on Robot Navigation, the Sixth National Conference of the American
pp.1-5. Palo Alto, CA: AAAI/MIT Press. Association for Artificial Intelligence Conference, pp.
Asada M, Stone P, Kitano H et al. (1998) The 202-206 Seattle, WA: AAAI/MIT Press.
RoboCup physical agent challenge: Phase I. Gat E (1998) On three-layer architectures. In:
Applied Artificial Intelligence 12: 251-263. Kortenkamp D, Bonnasso R and Murphy R
Asada M, Uchibe E, Noda S, Tawaratsumida S (eds) Artifical Intelligence and Mobile Robotics.
and Hosoda K (1994) Coordination of AAAI Press.
multiple behaviors acquired by a vision-based Gerkey B and Matarić M (2002) Principled
reinforcement learning. In: Proceedings, communication for dynamic multi-robot task
IEEE/RSJ/GI International Conference on allocation. In: Rus D and Singh S (eds)
Situated Robotics 6
Proceedings of the International Symposium on Pomerleau D (1989) ALVINN: an autonomous
Experimental Robotics 2000, Waikiki, Hawaii, pp. land vehicle in a neural network. In: Touretzky
341-352. Berlin Heidelberg: Springer-Verlag. D (ed) Advances in Neural Information Processing
Giralt G, Chatila R and Vaisset M (1983) An Systems 1, pp. 305-313. San Mateo, CA:
integrated navigation and motion control Morgan Kaufmann.
system for autonomous multisensory mobile Scassellati B (2001) Investingating models of
robots. In: Proceedings of the First International social development using a humanoid robot.
Symposium on Robotics Research, pp. 191-214. In Webb B and Consi T (eds) Biorobotics,
Cambridge, MA: MIT Press. pp.145-168. Cambridge, MA: MIT Press.
Krieger M, Billeter J-B and Keller L (2000) Ant- Schoppers M (1987) Universal plans for reactive
like task allocation and recuirtiment in robots in unpredictable domains. In:
cooperative robots. Nature 406: 992. Proceedings, IJCAI-87, pp. 1039-1046. Menlo
Kube R and Zhang H (1992) Collective robotic Park, CA: Morgan Kaufman
intelligence. In: Proceedings, Simulation of Webb B (1994) Robotic experiments in cricket
Adaptive Behavior, pp. 460-468. Cambridge, phonotaxis. In: Proceedings of the Third
MA: MIT Press. International Conference on the Simulation of
Maes P (1990) Situated agents can have Adaptive Behavior, pp. 45-54. Cambridge, CA:
goals. Robotics and Autonomous Systems 6: 49-70. MIT Press.
Malcolm C and Smithers T (1990) Symbol Werger B (1999) Cooperation without
grounding via a hybrid architecture in an deliberation: a minimal behavior-based
autonomous assembly system. Robotics and approach to multi-robot teams. Artificial
Autonomous Systems 6: 145-168. Intelligence 110: 293-320.
Matarić M (1990) Navigating with a rat brain: a
neurobiologically-inspired model for robot Further Readings
spatial representation. In: Meyer J-A and
Wilson S (eds) Proceedings, From Animals to Arkin R (1998) Behavior-Based Robotics.
Animats 1, First International Conference on Cambridge, MA: MIT Press.
Simulation of Adaptive Behavior. pp. 169-175. Brooks R (1999) Cambrian Intelligence.
Cambridge,MA: MIT Press. Cambridge,MA: MIT Press.
Matarić M (1992) Integration of representation Maes P (1994) Modeling adaptive autonomous
into goal-driven behavior-based robots. IEEE agents. Atificial Life 2(2): 135-162.
Transactions on Robotics and Automation 8 (3): Russell S and Norvig P (1995) Artificial
304-312. Intelligence: A Mondern Approach. Englewood
Matarić M (1995) Designing and understanding Cliffs, NJ: Prentice Hall.
adaptive group behavior. Adaptive Behavior
4(1): 51-80. Glossary
Matarić M (1997) Behavior-based control:
examples from navigation, learning, and group Autonomous robot A robot capable of
behavior. Journal of Experimental and Theoretical performing without any external user or
Artificial Intelligence 9: 323-336. operator intervention.
Matarić M (1998) Behavior-based robotics as a Behavior-based robot control Using
tool for synthesis of artificial behavior and collections of behaviors (which may be
analysis of natural behavior. Trends in Cognitive reactive or may contain state and internal
Science 2(3): 82-87. representations) to structure robot control.
Matarić M (2000) Getting humanoids to move Deliberative robot control The use of
and imitate. IEEE Intelligent Systems 15(4): 18- centralized representations and planning
24. methods for generating a sequence of actions
for the robot to perform.
Situated Robotics 7
Embodiment A form of situatedness, having a takes sensory inputs from its environment,
body and having one’s actions directly and processes them, and acts on its environment
strongly affected and constrained by that through its effectors in order to achieve a set
body. of goals.
Hybrid robot control Using a combination of Robot control The process of taking
methods, typically a combination of information about the environment, through
deliberative and reactive control, to control a the robot’s sensors, processing it as necessary
robot. in order to make decisions about how to act,
Learning robots Robots capable of improving and then executing those actions in the
their performance over time, based on past environment.
experience. Situated robotics The field of research that
Reactive robot control The use of only focuses on robots that are embedded in
reactive rules, and no internal memory or complex, challenging, often dynamically
planning, in order to enable the robot to changing environments.
quickly react to its environment and task. Situatedness Existing in, and having one’s
Robot A physical system equipped with sensors behavior strongly affected by a complex
(e.g., cameras, whiskers, microphones, sonars) environment.
and effectors (e.g., arms, legs, wheels) that


Keywords: (Check)

Robotics; situatedness; embodiment; learning; autonomy