Autor(es):
Bicho, E. 
Data: 2000
Identificador Persistente: http://hdl.handle.net/1822/2173
Origem: RepositóriUM - Universidade do Minho
Assunto(s): Autonomous robotics; Behavior based robotics; Dynamical systems; Moving attractors; Reppelers; Target acquisition; Obstacle avoidance; Wall following; Target representation; Wall representation; Dynamic fields; Dynamic neural fields
Descrição
Tese de doutoramento em Electrónica Industrial, ramo de Automação e Controlo. Guimarães : Universidade do Minho, 1999. The aim of this thesis is to investigate how behavior-based robots can be modeled
by non-linear dynamical systems. Taking the example of navigation as a case
study, dynamic control architectures are developed and implemented on low-level
vehicles. These architectures combine a number of behaviors and lead to flexible
and smooth overt behavior which is stably coupled in closed loop with sensory information.
Moreover, these architectures also comprise dynamical representations of
information which enable the vehicles to exhibit cognitive behaviors such as decision
making, memory, forgetting and robustness against noisy sensory information. The
design of the individual behaviors, of the representations of particular types of information,
as well as of their coupling is based on the qualitative theory of dynamical
systems and dynamic field theory. These provide a general theoretical language in
which autonomous robot architectures can be built. Dynamical systems theory was used as a theoretical language and tool to design,
specify, analyze, simulate and implement behavior-based control architectures. The
architectures were fully formulated in terms of dynamics and implemented on computationally
modest vehicle platforms based on very low-level sensory information.
The main ideas and achievements are the following:
1. Attractor dynamics can be used to control motion based on low-level
sensors
(a) Robot action can be generated in the manner of control systems, by
assigning values to planning (i.e. behavioral) variables continuously in
time. A process that can be formalized through dynamical systems. It
was shown how an intelligent choice of such planning variables makes it
possible to generate flexible behavior from asymptotically stable states
(attractors) of such dynamical systems.
(b) Generation of behavior is an intrinsically non-linear problem. Behavioral
situations exist, in which a minor change in the configuration or sensory
situation must elicit a qualitative change in the behavior. This is a simple
form of decision making. This non-linearity poses a problem for the design
of dynamical systems, as no general theory exists for such systems.
(c) Bifurcation theory is one branch of the theory of non-linear dynamical
systems that is very structured by powerful theorems. Local bifurcation
theory helps to describe how attractors and repellers of dynamical systems
may annihilate by collision or emerge through splitting at critical
parameter values. Because the approach in this thesis made use of attractor
solutions only, local bifurcation theory may be employed to design
the dynamical systems such that an appropriate bifurcation is obtained
when a behavioral decision must be made. Additional benefits are the contribution of the behavioral dynamics to the overall control theoretic
stability of the autonomous robot even when decision making takes place.
(d) Sensory information enters into the dynamical system by defining attractive
or repulsive values of the planning variables, determining the
strength of attraction or repulsion as well as the range of values over
which theses forces act. Although the contribution of each individual
sensor to the dynamical system is not invariant under change of the values
of the planning variables, and neither necessarily generates the right
postulated functional form nor necessarily generates the right attractors
and repellers, the superposition of all contributions of all sensors does
exhibit that invariance and thus generates the designed dynamics. This
is ultimately true because the environment is invariant, for instance, under
rotation of the vehicles on the spot, and the summed contributions
sample that environment.
(e) We made a detailed presentation of how individual motion behaviors can
be designed, how they can be integrated and moreover how they can be
implemented on autonomous vehicles equipped with low-levels sensors
like infra-red sensors, sonars, photo-resistors or microphones.
(f) Vehicle motion toward targets while simultaneously avoiding obstacles
and/or following walls was generated from attractors of dynamical systems
of heading direction and path velocity.
(g) Vehicle motion toward targets and avoidance of perceived obstructions
also was generated from attractors of dynamical systems of angular and
path velocities.
2. Dynamic fields can endow robots with sub-symbolic representations
based on low-level sensory information
(a) We showed how the ideas of attractor dynamics employed to control the
motion of the robots can be extended to the level of representation by
using dynamic fields to interpolate sensory information.
(b) We have demonstrated how dynamic fields can provide robotic systems
with sub-symbolic representations that rely on low-level sensory information.
These representations enabled a robotic vehicle to exhibit the
simplest forms of cognitive abilities. For instance:
• A dynamic field model for target representation based on low-level
sound sensors was built. This permitted the robot to exhibit skills
such as detecting targets only if sensory information was consistent,
estimating direction to a target through interpolation, deciding which
target to track when multiple targets were presented and stabilization
of such decision, maintaining targets in short-term memory during
momentary absence of pertinent sensory information, and deleting a memory item after a characteristic delay to clear the memory from
obsolete information.
• A dynamic field model for wall representation based on low-level
distance sensors (infra-red sensors and sonars) was also built that
supported wall detection, robust wall orientation estimation and wall
selection.
3. Representations can be integrated with stable action planning and
control
(a) The pattern of activation in a dynamic field shapes continuously in time
the vector field of the dynamics of the behavior to which it is dedicated:
i. The value over which a peak of positive activation is centered de-
fines an attractive value of the planning variable used to design the
behavioral dynamics.
ii. The amount of total positive activation in the field determines the
attraction strength of that force.
iii. Because the field dynamics is invariant under rotations of the vehicle
on the spot so is the dynamics that brings about the motion of the
vehicle.
(b) The amount of positive activation in a dynamic field may also inhibit the
contribution of other behaviors to the complete behavioral dynamics.
(c) The time scale of the field dynamics is set much faster than that of
the planning variables so that the field has typically relaxed to a stable
pattern on the time scale on which the movement plans of the robot
evolve. The resulting behavior is therefore stable.
(d) Two examples are:
• The dynamic field for target representation specifies the particular
form of the target acquisition dynamics.
• Analogously, the dynamic field for wall representation specifies the
wall-following dynamics.
• The dynamic field for wall representation inhibits the contribution of
obstacle avoidance to the planning dynamics when a wall is detected.
4. Navigation in non-engineered environments
(a) The smooth overt behavior generated when the systems are set to work
in non-structured environments was documented.
(b) The implementation of the dynamic field for target representation on a
small autonomous vehicle enabled it to find sound sources while avoiding
obstacles. Memory helped to keep it moving toward the target during the periods when sensory information was not available. Decision making
enabled the vehicle to track only one sound source. Hysteresis in the
field dynamics enabled the vehicle to continue moving toward the selected
sound target even when it approach a second sound source of equal intensity
due to obstacle avoidance. Detection of a sound target with an
intensity near the sensor threshold is stabilized through the cooperative
forces within the field.
(c) The implementation of the dynamic field for wall representation on a
small autonomous vehicle enabled it to follow walls with various shapes
(e.g. planar, circular and concave corners). Decision making allowed to
select a wall and hysteresis permitted to maintain the decision stable even
when new walls were encountered during the motion course.
