A Neural Schema Architecture for Autonomous Robots
Abstract
As autonomous robots become more complex in their behavior, more sophisticated software
architectures are required to support the ever more sophisticated robotics software. These
software architectures must support complex behaviors involving adaptation and learning,
implemented, in particular, by neural networks. We present in this paper a neural based schema
[2] software architecture for the development and execution of autonomous robots in both
simulated and real worlds. This architecture has been developed in the context of adaptive
robotic agents, ecological robots [6], cooperating and competing with each other in adapting to
their environment. The architecture is the result of integrating a number of development and
execution systems: NSL, a neural simulation language; ASL, an abstract schema language; and
MissionLab, a schema-based mission-oriented simulation and robot system. This work
contributes to modeling in Brain Theory (BT) and Cognitive Psychology, with applications in
Distributed Artificial Intelligence (DAI), Autonomous Agents and Robotics.
Areas: Robotics, Agent-oriented programming, Neural Nets
Keywords: Autonomous Robots, Autonomous Agents, Schemas, Neural Networks,
Architecture
1 Introduction
To enable the development and execution of complex behaviors in autonomous robots
involving adaptation and learning, sophisticated software architectures are required.
The neural schema architecture provides such a system, supporting the development
and execution of complex behaviors, or schemas [3][2], in a hierarchical and layered
fashion [9] integrating with neural network processing.
In general, schema theory helps define brain functionality in terms of concurrent
activity of interacting behavioral units called schemas. Schema-based modeling may
be specified purely on behavioral data (ethology), while becoming part of a neural
based approach to adaptive behavior when constrained by data provided by, e.g., the
effects of brain lesions upon animal behavior (neuroethology). Schema modeling
provides a framework for modeling at the purely behavioral level, at the neural
network level or even below [28]. In terms of neural networks, neural schema theory
provides a functional/structural decomposition, in strong contrast with models which
employ learning rules to train a single, otherwise undifferentiated, neural network to
respond as specified by some training set. Neural schema-based modeling proceeds
at two levels: (1) model behavior in terms of schemas, interacting functional units;
(2) implementation of schemas as neural networks based on neuroanatomical and
neurophysiological studies. What makes the linking of structure and function so
challenging is that, in general, a functional analysis proceeding "top-down" from
some overall behavior need not map directly into a "bottom up" analysis proceeding
upwards from the neural circuitry.
The work described in this paper is the product of a collaborative research depicted in
Figure 1.
Neural Simulation
Language (NSL)
Abstract Schema
Language (ASL)
Perceptual-Motor
Schema Model
MissionLab (Mlab)
Predictions
Results
Biological Data Simulations Robot Experiments
Common Language
Figure 1. Collaboration Map
Biological data from behavioral studies in the praying mantis "Chantitlaxia" [11] and
the frog and toad prey acquisition and predator avoidance behaviors [12][14], are
used to generate neural schema models: perceptual schemas, dealing with sensory
input or perceptions; motor schemas, dealing with motor action; and sensorimotor
schemas, integrating between sensory input and motor action. These studies are
modeled in terms of computational schemas in the Abstract Schema Language (ASL)
[25], implemented as neural networks in the Neural Simulation Language (NSL) [27],
and simulated in a virtual world or executed in the real world with the MissionLab
(Mlab) robotic system [23].
2 Schemas, Neural Networks and Autonomous Robots
The neural schema architecture for autonomous robots comprises the integration of
three separately developed architectures, each built to support a different aspect of
schema modeling.
2.1 Schemas
As a computational model, schemas define a hierarchical and distributed architecture
for the development of complex adaptive systems. A number of schema-based
architectures have been developed for different application domains, e.g. VISIONS
[18], in vision; RS (Robot Schemas) [22] and MissionLab [3], in robotics. Based on
these domain specific architectures, a unified schema computational model, ASL
(Abstract Schema Language) [25], was designed with the ability to integrate with
neural networks processing across different domains as well. Schemas in ASL are
hierarchical and distributed autonomous agents, where ASL integrates concurrent
object-oriented programming methodologies [29] with agent modeling [8]. As a
language ASL corresponds more to a specification language rather than to an explicit
programming language. The detailed implementation is left unspecified, only
specifying what is to be achieved. Different implementations may correspond to a
single schema, where implementation are in terms of neural networks or other schema
process. The ASL computational model is shown in Figure 2.
At the top of Figure 2, a schema is shown decomposed into other schemas. This
decomposition gives rise to schema aggregation, or schema assemblages. Schemas
are specified and implemented either through wrapping, which enables static
integration of external programs, or through task delegation, which enables dynamic
integration of schemas as separate specification and implementation tasks. (Solid
arrows between boxes represent connections between objects, while dashed arrows
represent task delegation.)
schema
instance
schema
instance
data in
data out
schema
instance
neural schema other process
Figure 2. Schema Hierarchy
Schema interfaces consists of multiple unidirectional control or data, input and output
ports, and a method section where schema behavior is specified. Communication is in
the form of asynchronous message passing, hierarchically managed, internally,
through anonymous port reading and writing, and externally, through dynamic port
connections and relabelings. When doing connections, output ports from one schema
are connected to input ports from other schemas, and ports from schemas at different
hierarchies are linked to each other when doing relabelings. The hierarchical port
management methodology enables the development of distributed systems where
objects may be designed and implemented independently and without prior
knowledge of their final execution environment, encouraging model reusability. This
supports both top-down and bottom-up system designs as required by neural schema
modeling.
In order to support complex schema modeling, ASL is design as a distributed
multithreaded system architecture, executing on different platforms [10], as shown in
Figure 3.
Figure 3. Abstract Schema Language (ASL) System Architecture
2.2 Neural Networks
Neural networks serve as the underlying implementation for neural schemas. Lower
level neural network components integrate with higher level schemas, as shown in
Figure 4:
schema
instance
data in
data out
schema
instance
neural
schema
instance
complex simple
neural
network
Figure 4. Neural Schema Hierarchy
The Neural Schema Language (NSL) [27] provides the linkage to ASL, enabling the
integration of neural networks as schema implementations. The ability to implement
schemas through different neural networks results in the added benefit of enabling the
construction of distributed neural networks. Mapping between schemas and neural
networks may not only be 1 to 1, but also many to many. The neural schema model
not only enables the encapsulation of neural networks into schemas, but also provides
an extended model where neurons themselves may have their task delegated by neural
implementations of different levels of detail, from the very simple neuron models to
the very complex ones [26].
Operating System
User Models
NSL Tcl/Tk Scrippting Language
NSL Model Class Library
NSL Graphics Interface
NSL Runtime System
NSL Java/C++ Compiled Language
Figure 5. NSL System Architecture
The NSL system architecture is shown in Figure 5. Models are described via a
compiled language, where graphics displays and a scripting language provide the
interfacing mechanisms between the model and the user. Two implementations of the
system currently exist: NSLC in C++ and NSLJ in Java.
2.3 Schema-based control for autonomous robots
In robotics, schemas have been used to provide the underlying software control
mechanisms for a number of systems, e.g. MissionLab [3] and RS [22]. In particular,
in the control of autonomous robots, such as with MissionLab, motor schemas have
been encoded as a variant of the potential field methodology [21]. In this context,
schemas have the following characteristics:
1. Each is an independent asynchronous computational agent executing in parallel
with other schemas.
2. Sensing is directly tied to motor control following the action-oriented perception
paradigm, where information is obtained via sensing on a need-to-know basis [4].
3. Each active schema produces a vector that encodes the behavioral response for a
given stimulus.
4. Coordination of schemas is typically conducted via behavioral fusion: vector
summation and normalization of the individual schemas outputs.
5. Schemas can be aggregated into assemblages, which provide a higher level of
abstraction.
6. Their use is rooted in neuroscientific and psychological theory.
This particular form of behavioral control has been tested on a wide range of robotic
systems: from teams of small robots used for competitions to military sized vehicles
[5], as shown in the Figure 6.
Figure 6. Collection of schema-based robots
MissionLab [23] is a tool that has been recently developed for the testing and
deployment of schema-based reactive controllers for autonomous robots. It
incorporates a graphical user interface, reusable software libraries, a simulation
facility, and the capability to download executable robot code for a range of real
mobile platforms. MissionLab serves as the testbed for the results in this project. The
architecture of MissionLab is shown in Figure 7.
Graphic
Designer
[CDL]
CDL Compiler
syntax semantics
Compiler
Interface
SAUSAGES
Architecture
Descriptions
Robot
Descriptions
architecturej
Unix
Executable
Behavior
library
Maintenance
Interface
Architecture
binder
interface
Architecture specific
and robot specific
representations
USE
R
behaviors
abstractbehaviors
Architecture
binding
Robot
binding
Requirements
checking
Requirements
checking roboti
Code generator
for Code generator
for UGV
Code generator
for AuRA
architecture
ParseTree
C++ Code
executeon
matching
simulation
executeon
matching
robot
CNL Code
behaviorimplementations
Figure 7. MissionLab System Architecture
2.4 Integrated Architecture
In order to enable the described schema modeling, the three architectures: ASL, NSL
and Missionlab, were integrated under a single system environment. ASL was first
integrated to NSL [10], and then the ASL/NSL system to MissionLab [24]. The
integrated architecture is shown in Figure 8.
Sensor Input
Motor Output
Perceptual and
Sensorimotor
Neural
Schemas
Motor
Schemas
MissionLab ASL/NSL
Figure 8. ASL/NSL/MissionLab Integrated Architecture
Integration is carried out through binding points between ASL/NSL and MisionLab.
Sensor input from MissionLab, simulated data or real world data from actual robots,
is read by the perceptual neural schemas in the ASL/NSL system. Sensorimotor
neural schemas in ASL/NSL generate output to the motor schemas executing in
MissionLab, either in the simulated or real world.
3 Computational Neuroethology
Neuroethology, the study of the nervous system and animal behavior, has inspired a
number of computational models, such as Rana Computatrix, the computational frog
[1], the computational cockroach [7], and the computational hoverfly [13]. Such
computational models involve a rich number of neural based behaviors, such as the
Chantlitaxia, searching from a proper habitat, taken from the Praying Mantis behavior
[11], as described in the ethogram in Figure 9.
Exploration Orientation
Swinging
Walk
Detour
Jump
Grasp
REST (mimesis)
Alternative conducts
Climbing
· Mating
· Hunting
· Self-cleaning
· Fly
· “Oviposicion”
· Ecdisis
· Defense
Criptosis
Deimatic
Flight
Simulated catalepsy
(The action of releasing
the ootheca with the eggs).
Figure 9. Praying Mantis' Chantlitlaxia Ethogram
Different models are currently being developed under the ASL/NSL/MissionLab
neural schema architecture. Besides the Chantlitlaxia behavior [6], we have
prototyped the adaptive toad's prey acquisition behavior due to a static barrier [14],
and developing a prey acquisition and predator avoidance behavior modulated by
learning processes in neural networks [20].
3.1 Prey Acquisition with Detour Behavior
As an example of a model developed under the neural based schema architecture we
describe the toad's detour behavior due to stationary objects on its way to a prey [14].
The experiment being modeled consists of a barrier placed between a prey and a toad,
as shown in Figure 10.
Two different barrier sizes were tried, 10 and 20 cm. Both barriers are made of
fenceposts, where each fencepost has a very small width, but tall enough not to have
the toad jump over it. The fence posts are distanced 2 cm from each other. The toad is
20 cm away from the barrier, and the prey is 10 cm away opposite the barrier.
1
0 cm
2
0 cm
2
0
cm
1
0
cm prey
toad
20cm 10cm
20cm
10cm
barrier
Figure 10. Toad's prey acquisition with detour behavior experiment
When the barrier is 10 cm wide the toad approaches directly to the barrier edges and
from there continues to the prey, as shown in Figure 10. When the barrier is 20 cm
wide, the toad advances to the middle of the barrier, more precisely to the closest gap
between the fenceposts. Not being able to go through the gap, the robot backs up,
reorients and tries again. This adaptive process continues until the edge of the barrier
is in sight. Figure 11 shows the toad's behavior with a 20 cm barrier without and with
learning. These experiments are further described in [17][15].
Toad approximation to 20cm barrier
after learning
Toad approximation to 20cm barrier
Figure 11. Toad's prey acquisition model for a 20 cm barrier, without and with learning.
Schemas
In order to reproduce these experiments we developed a schema based model with a
robotic agent taking the place of the toad. At the highest level, model behavior is
described by means of schema specifications. The complete model at this level is
described by a network of interconnected schemas as shown in Figure 12:
The model consists of visual and tactile sensory input, perceptual schemas for
recognizing stationary objects and prey moving objects, sensorimotor schemas for
prey approach and static object avoidance, and motor schemas for performing
forward, backward, sidestep and orient motions. Visual input is used to recognize
both the static barrier and the moving prey, while tactile input is triggered when the
robotic agent bumps into the barrier (not being able to go through the gap).
Rather than processing input symbols to yield output symbols, the individual schemas
have activation levels which measure their degree of confidence. In response to the
perceptual schemas input, the more active of the two sensorimotor schemas will
trigger the appropriate motor schema to yield the appropriate response.
Static Object
Avoid
Orient
Sidestep
Backward
Prey Approach Forward
Static Object
Recognizer
Prey
Recognizer
Tactile
Visual
Sensors
Sensorimotor
Schemas Actuators
Perception Action
Perceptual
Schemas
Motor
Schemas
Legs
Figure 12. Schemas for toad's prey acquisition with detour behavior.
In other words, the sensorimotor schemas compete to control the behavior of the
animal. This is a very simple example of the type of mechanisms of competition and
cooperation that can be exhibited by a network of schemas. In particular multiple
motor schemas may be coactivated to control subtle behaviors. The perceptual
schemas are not simply yes-no recognizes, being equipped with a confidence level to
provide a parametric description which can be used in tuning motor behavior
appropriately. When the toad recognizes the prey, the animal does not respond by
moving in a standard or random direction, bur rather it snaps at the position in space
where the prey is located as indicated by the "prey-recognizer" schema.
Neural Networks
Some of the schemas in the toad's prey acquisition model are implemented all the way
down to neural networks. Other schemas, for which the detailed neural circuitry is not
known or involves unnecessary computation for the range of phenomena under study,
are modeled in a simpler manner. For example, motor schemas in this model were not
implemented through neural circuitry for simplification reasons. The neural network
level implementing higher level schemas is shown in Figure 13.
R4
Visual
R1-R2
R3
Retina
T5_2 layer
TH10 layer
Motor
Heading
Map
Static Object
Recognizer
Prey
Recognizer
TH10
T5_2
Static Object
Avoidance
Prey Approach
Forward
Orient
Sidestep
Tactile Backward Schema Level
Neural Level
Figure 13. Neural schemas for toad's prey acquisition with detour behavior.
The neural level consists of a Retina corresponding to the Visual input, T5_2 and
TH10 neural layers corresponding the moving prey and static object recognizer, and a
motor heading map where the static object and prey maps integrate. The motor
heading map produces a target heading angle corresponding to the strongest map
activity; providing inhibition between the prey approach and static object avoidance.
This inhibition is important to avoid activating antagonist motor schemas
simultaneously. A tactile modulation component provides adaptation to the model by
increasing the inhibition repetitively, every time the robot hits the barrier. (The
detailed model description can be found in [14].)
Autonomous Robots
The complete autonomous robotic agent is built by integrating the perceptual and
sensorimotor schemas in ASL/NSL with the motor schemas in MissionLab, as
shown in Figure 14.
Visual
Tactile
Orient
Forward
Backward
Sidestep
Robot
Sensor
and Motor
Binding
Schema and
Neural
Network
Processing
Figure 14. Perceptual and Motor Schema Linkage for toad's prey acquisition model.
The robot provides visual and tactile input to the neural schema process. These
respond by producing appropriate forward, orient, sidestep and backward activations,
generating robot movement. The cycle continues indefinitely, terminating only when
reaching the prey. When executed in a real robot, only sensory and actuator binding is
modified in MissionLab without the need to change any of the actual model details.
4 Results
4.1 Prey Acquisition with Detour Behavior
As seen from the Missionlab simulation console, the robot (SP Frog) is initially
positioned in front of a barrier, with a prey away from it, as shown in the left of
Figure 15. The right hand side shows the resulting trajectory generated by the agent.
Figure 16, left, shows the agent's view of the barrier. Figure 16, right, the resulting
attraction field integrating the prey attraction and the barrier repulsion. The highest
value activity in the figure corresponds to the robot's preferred orientation (which
initially corresponds to the prey's direction).
As the robot bumps into the barrier, the barrier's gap inhibition gets incremented
resulting in a new attraction field in the motor heading map, Figure 17, left, producing
reorientation. Every time the frog hits the barrier, it backs down and sidesteps.
Following, the frog gets attracted by the prey again, hitting the barrier, this time on a
different gap. This process continues until the edge of the barrier is perceived,
generating a direct path to the prey, as shown in Figure 17, right.
Figure 15. MissionLab Console view of agent response to the 20cm wide barrier.
Figure 16. Attractant field integrating prey attraction and barrier repulsion.
Figure 17. Attractant field when barrier gap is in sight.
This specific trajectory was generated due to the model reorientation specifics. Other
simulated results, and more detailed results, can be found in [14].
5 Conclusions and Future Work
This paper has shown the fundamentals of the ASL/NSL/MissionLab neural schema
architecture for autonomous robots. A previous architecture is described in [19].
An important aspect of this architecture is the ability to incorporate adaptation and
learning through neural network processes in developing new behavioral architectures
for autonomous agents [16] as well as robots. This goes beyond architectures where
behaviors are described in terms of global states or architectures limited in terms of
adaptation and learning mechanisms. Furthermore, as models become more complex
in their nature, the distributed and concurrent nature of the ASL/NSL/MissonLab
architecture becomes of even greater importance. The prey acquisition model
presented in this paper reproduces one of a number of behavioral experiments with
toads. Other experiments are currently being tested under this architecture, in
particular, extensions to the toad's and praying mantis prey acquisition and predator
avoidance models as they are modulated by learning processes [17]. Furthermore, we
are also in the process of experimenting with these models with actual robots in the
real world [6].
References
[1] Arbib, M.A., Levels of Modelling of Mechanisms of Visually Guided Behavior,
Behavior Brain Science 10:407-465, 1987.
[2] Arbib, M.A., "Schema Theory", in the Encyclopedia of Artificial Intelligence, 2nd
Edition, Editor Stuart Shapiro, 2:1427-1443, Wiley, 1992.
[3] Arkin, R.C., "Motor Schema-Based Mobile Robot Navigation", International
Journal of Robotics Research, Vol. 8, No. 4, August 1989, pp. 92-112, 1989.
[4] Arkin, R.C., "The Impact of Cybernetics on the Design of a Mobile Robot
System: A Case Study", IEEE Transactions on Systems, Man, and Cybernetics,
20:6, pp. 1245-1257, Nov/Dec. 1990.
[5] Arkin, R.C. and Balch, T., "AuRA: Principles and Practice in Review", Journal of
Experimental and Theoretical Artificial Intelligence, 9:2-3, pp. 175-189, April-
Sept. 1997.
[6] Arkin, R.C., Cervantes-Perez, F., and Weitzenfeld, A., "Ecological Robotics: A
Schema-Theoretic Approach", to appear in "Intelligent Robots: Sensing,
Modelling and Planning", Editors R.C.Bolles, H.Bunke, and H.Noltemeier,
World Scientific, 1997.
[7] Beer, R. D., Intelligence as Adaptive Behavior: An Eperiment in Computational
Neuroethology, San Diego, Academic Press, 1990.
[8] Briot, J.-P., Gasser, L., "From Objects to Agents: Connections between Object-
Based Concurrent Programming and Distributed Artificial Intelligence", IJCAI
'91 Workshop on Objects and AI, 1991.
[9] Brooks, R., "A Robust Layered Control System for a Mobile Robot", IEEE
Journal of Robotics and Automation, 2:14-23, March 1986.
[10] Calderas, C., Mármol, S., "Diseño e Implementación de un Sistema Distribuido
Heterogéneo para la Extensión del Lenguaje de Modelos Teórico Esquemáticos
ASL", ITAM Undergraduate Thesis, CANNES-TR-96-01, ITAM, México,
Octubre 1996.
[11] Cervantes-Perez, F., Franco, A., Velazquez, S., Lara, N., 1993, "A Schema
Theoretic Approach to Study the 'Chantitlaxia' Behavior in the Praying Mantis",
Proceeding of the First Workshop on Neural Architectures and Distributed AI:
From Schema Assemblages to Neural Networks, USC, October 19-20, 1993.
[12] Cervantes-Perez, F., Herrera, A., and García, M., "Modulatory effects on preyrecognition
in amphibia: a theoretical 'experimental study'", in Neuroscience:
from neural networks to artificial intelligence, Editors P. Rudomian, M.A. Arbib,
F. Cervantes-Perez, and R. Romo, Springer Verlag Research Notes in Neural
Computing vol 4, pp. 426-449, 1993.
[13] Cliff, D., Neural Networks for Visual Tracking in an Artificial Fly, in Towards a
Practice of Autonomous Systems: Proc. of the First European Conference on
Artifical Life (ECAL 91), Editors, F.J., Varela and P. Bourgine, MIT Press, pp
78-87, 1992.
[14] Corbacho, F., Arbib M. "Learning to Detour", Volume 3, Number 4, pp 419-468,
Adaptive Behavior, 1995.
[15] Corbacho, F. Arbib, M. A., "Learning Internal Models to Detour" Society for
Neuroscience. Abs. 624.7, 1997.
[16] Corbacho, F. Arbib, M. A. "Schema-Based Learning: Towards a Theory of
Organization for Adaptive Autonomous Agents" Proceedings of the First
International Conference on Autonomous Agents, Marina del Rey, California,
1997.
[17] Corbacho, F., Khort, B., Lin, B. Nothis, A. & Arbib, M. A. "Learning to Detour:
Behavioral Experiments with frogs" Proceedings of the Workshop on
Sensorimotor Coordination: Amphibians, Models, and Comparative Studies.
Sedona, Arizona, 1996.
[18] Draper, B., Collins, R., Brolio, J., Hanson, A., Riseman, E., "The Schema
System", Int Journal of Computer Vision, 2:209-250, 1989.
[19] Fagg, A., King, I., Lewis, A., Liaw, J., Weitzenfeld, A., "A Testbed for
Sensorimotor Integration", Proceedings of IJCNN '92, Baltimore, MD, 1:86-91,
1992.
[20] Flores, L.R., "Modulación de Tiempo Variable de la Elicitación de la Conducta
de Captura de Presas en los Anfibios Anuros", UNAM Master Thesis, CANNESTR-
97-01, ITAM, Mexico, DF, Agosto 1997.
[21] Khatib, O., "Real-time Obstacle Avoidance for Manipulators and Mobile
Robots", Proc. IEEE International Conference Robotics and Automation, pp.
500-505, St. Louis, 1985.
[22] Lyons, D.M., Arbib, M.A, "A Formal Model of Computation for Sensory-Based
Robotics, IEEE Trans. on Robotics and Automation", 5:280-293, June, 1989.
[23] MacKenzie, D., Arkin, R.C., and Cameron, R., "Multiagent Mission
Specification and Execution", in Autonomous Robots, 4:1, pp.29-52, Jan.1997.
(Also appears in Robot Colonies, ed.R.Arkin and G.Bekey, Kluwer Academic
Publishers, 1997.)
[24] Olivares, R., "Integración del Ambiente de Simulación de Robótica
Neuroetológica: NSL/ASL/MissionLab", ITAM Undergraduate Thesis,
CANNES-TR-97-02, ITAM, México, DF, Septiembre 1997.
[25] Weitzenfeld, A., "ASL: Hierarchy, Composition, Heterogeneity, and Multi-
Granularity in Concurrent Object-Oriented Programming", Proceeding of the
Workshop on Neural Architectures and Distributed AI: From Schema
Assemblages to Neural Networks, Oct 19-20, USC, October 19-20, 1993.
[26] Weitzenfeld, A., Arbib, M., "A Concurrent Object-Oriented Framework for the
Simulation of Neural Networks", Proceedings of ECOOP/OOPSLA '90
Workshop on Object-Based Concurrent Programming, OOPS Messenger,
2(2):120-124, April 1991.
[27] Weitzenfeld, A., Arbib, M.A., "NSL - Neural Simulation Language", in Neural
Networks Simulation Environments, Editor J. Skrzypek, Kluwer, 1994.
[28] Weitzenfeld, A., Arbib, M.A., Cervantes-Perez, F., Rudomin, P., and Alexander,
A., 1998, "Multi-level Simulation Methodology: A Computational and
Experimental Approach to Neural Systems", in Proc. of 1998 NSF Design and
Manufacturing Grantees Conference, pp 651A-652A, Jan. 5-8, Monterrey,
Mexico.
[29] Yonezawa, A., Tokoro, M., Eds., 1987, Object-oriented concurrent
programming, MIT Press.
Abstract
As autonomous robots become more complex in their behavior, more sophisticated software
architectures are required to support the ever more sophisticated robotics software. These
software architectures must support complex behaviors involving adaptation and learning,
implemented, in particular, by neural networks. We present in this paper a neural based schema
[2] software architecture for the development and execution of autonomous robots in both
simulated and real worlds. This architecture has been developed in the context of adaptive
robotic agents, ecological robots [6], cooperating and competing with each other in adapting to
their environment. The architecture is the result of integrating a number of development and
execution systems: NSL, a neural simulation language; ASL, an abstract schema language; and
MissionLab, a schema-based mission-oriented simulation and robot system. This work
contributes to modeling in Brain Theory (BT) and Cognitive Psychology, with applications in
Distributed Artificial Intelligence (DAI), Autonomous Agents and Robotics.
Areas: Robotics, Agent-oriented programming, Neural Nets
Keywords: Autonomous Robots, Autonomous Agents, Schemas, Neural Networks,
Architecture
1 Introduction
To enable the development and execution of complex behaviors in autonomous robots
involving adaptation and learning, sophisticated software architectures are required.
The neural schema architecture provides such a system, supporting the development
and execution of complex behaviors, or schemas [3][2], in a hierarchical and layered
fashion [9] integrating with neural network processing.
In general, schema theory helps define brain functionality in terms of concurrent
activity of interacting behavioral units called schemas. Schema-based modeling may
be specified purely on behavioral data (ethology), while becoming part of a neural
based approach to adaptive behavior when constrained by data provided by, e.g., the
effects of brain lesions upon animal behavior (neuroethology). Schema modeling
provides a framework for modeling at the purely behavioral level, at the neural
network level or even below [28]. In terms of neural networks, neural schema theory
provides a functional/structural decomposition, in strong contrast with models which
employ learning rules to train a single, otherwise undifferentiated, neural network to
respond as specified by some training set. Neural schema-based modeling proceeds
at two levels: (1) model behavior in terms of schemas, interacting functional units;
(2) implementation of schemas as neural networks based on neuroanatomical and
neurophysiological studies. What makes the linking of structure and function so
challenging is that, in general, a functional analysis proceeding "top-down" from
some overall behavior need not map directly into a "bottom up" analysis proceeding
upwards from the neural circuitry.
The work described in this paper is the product of a collaborative research depicted in
Figure 1.
Neural Simulation
Language (NSL)
Abstract Schema
Language (ASL)
Perceptual-Motor
Schema Model
MissionLab (Mlab)
Predictions
Results
Biological Data Simulations Robot Experiments
Common Language
Figure 1. Collaboration Map
Biological data from behavioral studies in the praying mantis "Chantitlaxia" [11] and
the frog and toad prey acquisition and predator avoidance behaviors [12][14], are
used to generate neural schema models: perceptual schemas, dealing with sensory
input or perceptions; motor schemas, dealing with motor action; and sensorimotor
schemas, integrating between sensory input and motor action. These studies are
modeled in terms of computational schemas in the Abstract Schema Language (ASL)
[25], implemented as neural networks in the Neural Simulation Language (NSL) [27],
and simulated in a virtual world or executed in the real world with the MissionLab
(Mlab) robotic system [23].
2 Schemas, Neural Networks and Autonomous Robots
The neural schema architecture for autonomous robots comprises the integration of
three separately developed architectures, each built to support a different aspect of
schema modeling.
2.1 Schemas
As a computational model, schemas define a hierarchical and distributed architecture
for the development of complex adaptive systems. A number of schema-based
architectures have been developed for different application domains, e.g. VISIONS
[18], in vision; RS (Robot Schemas) [22] and MissionLab [3], in robotics. Based on
these domain specific architectures, a unified schema computational model, ASL
(Abstract Schema Language) [25], was designed with the ability to integrate with
neural networks processing across different domains as well. Schemas in ASL are
hierarchical and distributed autonomous agents, where ASL integrates concurrent
object-oriented programming methodologies [29] with agent modeling [8]. As a
language ASL corresponds more to a specification language rather than to an explicit
programming language. The detailed implementation is left unspecified, only
specifying what is to be achieved. Different implementations may correspond to a
single schema, where implementation are in terms of neural networks or other schema
process. The ASL computational model is shown in Figure 2.
At the top of Figure 2, a schema is shown decomposed into other schemas. This
decomposition gives rise to schema aggregation, or schema assemblages. Schemas
are specified and implemented either through wrapping, which enables static
integration of external programs, or through task delegation, which enables dynamic
integration of schemas as separate specification and implementation tasks. (Solid
arrows between boxes represent connections between objects, while dashed arrows
represent task delegation.)
schema
instance
schema
instance
data in
data out
schema
instance
neural schema other process
Figure 2. Schema Hierarchy
Schema interfaces consists of multiple unidirectional control or data, input and output
ports, and a method section where schema behavior is specified. Communication is in
the form of asynchronous message passing, hierarchically managed, internally,
through anonymous port reading and writing, and externally, through dynamic port
connections and relabelings. When doing connections, output ports from one schema
are connected to input ports from other schemas, and ports from schemas at different
hierarchies are linked to each other when doing relabelings. The hierarchical port
management methodology enables the development of distributed systems where
objects may be designed and implemented independently and without prior
knowledge of their final execution environment, encouraging model reusability. This
supports both top-down and bottom-up system designs as required by neural schema
modeling.
In order to support complex schema modeling, ASL is design as a distributed
multithreaded system architecture, executing on different platforms [10], as shown in
Figure 3.
Figure 3. Abstract Schema Language (ASL) System Architecture
2.2 Neural Networks
Neural networks serve as the underlying implementation for neural schemas. Lower
level neural network components integrate with higher level schemas, as shown in
Figure 4:
schema
instance
data in
data out
schema
instance
neural
schema
instance
complex simple
neural
network
Figure 4. Neural Schema Hierarchy
The Neural Schema Language (NSL) [27] provides the linkage to ASL, enabling the
integration of neural networks as schema implementations. The ability to implement
schemas through different neural networks results in the added benefit of enabling the
construction of distributed neural networks. Mapping between schemas and neural
networks may not only be 1 to 1, but also many to many. The neural schema model
not only enables the encapsulation of neural networks into schemas, but also provides
an extended model where neurons themselves may have their task delegated by neural
implementations of different levels of detail, from the very simple neuron models to
the very complex ones [26].
Operating System
User Models
NSL Tcl/Tk Scrippting Language
NSL Model Class Library
NSL Graphics Interface
NSL Runtime System
NSL Java/C++ Compiled Language
Figure 5. NSL System Architecture
The NSL system architecture is shown in Figure 5. Models are described via a
compiled language, where graphics displays and a scripting language provide the
interfacing mechanisms between the model and the user. Two implementations of the
system currently exist: NSLC in C++ and NSLJ in Java.
2.3 Schema-based control for autonomous robots
In robotics, schemas have been used to provide the underlying software control
mechanisms for a number of systems, e.g. MissionLab [3] and RS [22]. In particular,
in the control of autonomous robots, such as with MissionLab, motor schemas have
been encoded as a variant of the potential field methodology [21]. In this context,
schemas have the following characteristics:
1. Each is an independent asynchronous computational agent executing in parallel
with other schemas.
2. Sensing is directly tied to motor control following the action-oriented perception
paradigm, where information is obtained via sensing on a need-to-know basis [4].
3. Each active schema produces a vector that encodes the behavioral response for a
given stimulus.
4. Coordination of schemas is typically conducted via behavioral fusion: vector
summation and normalization of the individual schemas outputs.
5. Schemas can be aggregated into assemblages, which provide a higher level of
abstraction.
6. Their use is rooted in neuroscientific and psychological theory.
This particular form of behavioral control has been tested on a wide range of robotic
systems: from teams of small robots used for competitions to military sized vehicles
[5], as shown in the Figure 6.
Figure 6. Collection of schema-based robots
MissionLab [23] is a tool that has been recently developed for the testing and
deployment of schema-based reactive controllers for autonomous robots. It
incorporates a graphical user interface, reusable software libraries, a simulation
facility, and the capability to download executable robot code for a range of real
mobile platforms. MissionLab serves as the testbed for the results in this project. The
architecture of MissionLab is shown in Figure 7.
Graphic
Designer
[CDL]
CDL Compiler
syntax semantics
Compiler
Interface
SAUSAGES
Architecture
Descriptions
Robot
Descriptions
architecturej
Unix
Executable
Behavior
library
Maintenance
Interface
Architecture
binder
interface
Architecture specific
and robot specific
representations
USE
R
behaviors
abstractbehaviors
Architecture
binding
Robot
binding
Requirements
checking
Requirements
checking roboti
Code generator
for Code generator
for UGV
Code generator
for AuRA
architecture
ParseTree
C++ Code
executeon
matching
simulation
executeon
matching
robot
CNL Code
behaviorimplementations
Figure 7. MissionLab System Architecture
2.4 Integrated Architecture
In order to enable the described schema modeling, the three architectures: ASL, NSL
and Missionlab, were integrated under a single system environment. ASL was first
integrated to NSL [10], and then the ASL/NSL system to MissionLab [24]. The
integrated architecture is shown in Figure 8.
Sensor Input
Motor Output
Perceptual and
Sensorimotor
Neural
Schemas
Motor
Schemas
MissionLab ASL/NSL
Figure 8. ASL/NSL/MissionLab Integrated Architecture
Integration is carried out through binding points between ASL/NSL and MisionLab.
Sensor input from MissionLab, simulated data or real world data from actual robots,
is read by the perceptual neural schemas in the ASL/NSL system. Sensorimotor
neural schemas in ASL/NSL generate output to the motor schemas executing in
MissionLab, either in the simulated or real world.
3 Computational Neuroethology
Neuroethology, the study of the nervous system and animal behavior, has inspired a
number of computational models, such as Rana Computatrix, the computational frog
[1], the computational cockroach [7], and the computational hoverfly [13]. Such
computational models involve a rich number of neural based behaviors, such as the
Chantlitaxia, searching from a proper habitat, taken from the Praying Mantis behavior
[11], as described in the ethogram in Figure 9.
Exploration Orientation
Swinging
Walk
Detour
Jump
Grasp
REST (mimesis)
Alternative conducts
Climbing
· Mating
· Hunting
· Self-cleaning
· Fly
· “Oviposicion”
· Ecdisis
· Defense
Criptosis
Deimatic
Flight
Simulated catalepsy
(The action of releasing
the ootheca with the eggs).
Figure 9. Praying Mantis' Chantlitlaxia Ethogram
Different models are currently being developed under the ASL/NSL/MissionLab
neural schema architecture. Besides the Chantlitlaxia behavior [6], we have
prototyped the adaptive toad's prey acquisition behavior due to a static barrier [14],
and developing a prey acquisition and predator avoidance behavior modulated by
learning processes in neural networks [20].
3.1 Prey Acquisition with Detour Behavior
As an example of a model developed under the neural based schema architecture we
describe the toad's detour behavior due to stationary objects on its way to a prey [14].
The experiment being modeled consists of a barrier placed between a prey and a toad,
as shown in Figure 10.
Two different barrier sizes were tried, 10 and 20 cm. Both barriers are made of
fenceposts, where each fencepost has a very small width, but tall enough not to have
the toad jump over it. The fence posts are distanced 2 cm from each other. The toad is
20 cm away from the barrier, and the prey is 10 cm away opposite the barrier.
1
0 cm
2
0 cm
2
0
cm
1
0
cm prey
toad
20cm 10cm
20cm
10cm
barrier
Figure 10. Toad's prey acquisition with detour behavior experiment
When the barrier is 10 cm wide the toad approaches directly to the barrier edges and
from there continues to the prey, as shown in Figure 10. When the barrier is 20 cm
wide, the toad advances to the middle of the barrier, more precisely to the closest gap
between the fenceposts. Not being able to go through the gap, the robot backs up,
reorients and tries again. This adaptive process continues until the edge of the barrier
is in sight. Figure 11 shows the toad's behavior with a 20 cm barrier without and with
learning. These experiments are further described in [17][15].
Toad approximation to 20cm barrier
after learning
Toad approximation to 20cm barrier
Figure 11. Toad's prey acquisition model for a 20 cm barrier, without and with learning.
Schemas
In order to reproduce these experiments we developed a schema based model with a
robotic agent taking the place of the toad. At the highest level, model behavior is
described by means of schema specifications. The complete model at this level is
described by a network of interconnected schemas as shown in Figure 12:
The model consists of visual and tactile sensory input, perceptual schemas for
recognizing stationary objects and prey moving objects, sensorimotor schemas for
prey approach and static object avoidance, and motor schemas for performing
forward, backward, sidestep and orient motions. Visual input is used to recognize
both the static barrier and the moving prey, while tactile input is triggered when the
robotic agent bumps into the barrier (not being able to go through the gap).
Rather than processing input symbols to yield output symbols, the individual schemas
have activation levels which measure their degree of confidence. In response to the
perceptual schemas input, the more active of the two sensorimotor schemas will
trigger the appropriate motor schema to yield the appropriate response.
Static Object
Avoid
Orient
Sidestep
Backward
Prey Approach Forward
Static Object
Recognizer
Prey
Recognizer
Tactile
Visual
Sensors
Sensorimotor
Schemas Actuators
Perception Action
Perceptual
Schemas
Motor
Schemas
Legs
Figure 12. Schemas for toad's prey acquisition with detour behavior.
In other words, the sensorimotor schemas compete to control the behavior of the
animal. This is a very simple example of the type of mechanisms of competition and
cooperation that can be exhibited by a network of schemas. In particular multiple
motor schemas may be coactivated to control subtle behaviors. The perceptual
schemas are not simply yes-no recognizes, being equipped with a confidence level to
provide a parametric description which can be used in tuning motor behavior
appropriately. When the toad recognizes the prey, the animal does not respond by
moving in a standard or random direction, bur rather it snaps at the position in space
where the prey is located as indicated by the "prey-recognizer" schema.
Neural Networks
Some of the schemas in the toad's prey acquisition model are implemented all the way
down to neural networks. Other schemas, for which the detailed neural circuitry is not
known or involves unnecessary computation for the range of phenomena under study,
are modeled in a simpler manner. For example, motor schemas in this model were not
implemented through neural circuitry for simplification reasons. The neural network
level implementing higher level schemas is shown in Figure 13.
R4
Visual
R1-R2
R3
Retina
T5_2 layer
TH10 layer
Motor
Heading
Map
Static Object
Recognizer
Prey
Recognizer
TH10
T5_2
Static Object
Avoidance
Prey Approach
Forward
Orient
Sidestep
Tactile Backward Schema Level
Neural Level
Figure 13. Neural schemas for toad's prey acquisition with detour behavior.
The neural level consists of a Retina corresponding to the Visual input, T5_2 and
TH10 neural layers corresponding the moving prey and static object recognizer, and a
motor heading map where the static object and prey maps integrate. The motor
heading map produces a target heading angle corresponding to the strongest map
activity; providing inhibition between the prey approach and static object avoidance.
This inhibition is important to avoid activating antagonist motor schemas
simultaneously. A tactile modulation component provides adaptation to the model by
increasing the inhibition repetitively, every time the robot hits the barrier. (The
detailed model description can be found in [14].)
Autonomous Robots
The complete autonomous robotic agent is built by integrating the perceptual and
sensorimotor schemas in ASL/NSL with the motor schemas in MissionLab, as
shown in Figure 14.
Visual
Tactile
Orient
Forward
Backward
Sidestep
Robot
Sensor
and Motor
Binding
Schema and
Neural
Network
Processing
Figure 14. Perceptual and Motor Schema Linkage for toad's prey acquisition model.
The robot provides visual and tactile input to the neural schema process. These
respond by producing appropriate forward, orient, sidestep and backward activations,
generating robot movement. The cycle continues indefinitely, terminating only when
reaching the prey. When executed in a real robot, only sensory and actuator binding is
modified in MissionLab without the need to change any of the actual model details.
4 Results
4.1 Prey Acquisition with Detour Behavior
As seen from the Missionlab simulation console, the robot (SP Frog) is initially
positioned in front of a barrier, with a prey away from it, as shown in the left of
Figure 15. The right hand side shows the resulting trajectory generated by the agent.
Figure 16, left, shows the agent's view of the barrier. Figure 16, right, the resulting
attraction field integrating the prey attraction and the barrier repulsion. The highest
value activity in the figure corresponds to the robot's preferred orientation (which
initially corresponds to the prey's direction).
As the robot bumps into the barrier, the barrier's gap inhibition gets incremented
resulting in a new attraction field in the motor heading map, Figure 17, left, producing
reorientation. Every time the frog hits the barrier, it backs down and sidesteps.
Following, the frog gets attracted by the prey again, hitting the barrier, this time on a
different gap. This process continues until the edge of the barrier is perceived,
generating a direct path to the prey, as shown in Figure 17, right.
Figure 15. MissionLab Console view of agent response to the 20cm wide barrier.
Figure 16. Attractant field integrating prey attraction and barrier repulsion.
Figure 17. Attractant field when barrier gap is in sight.
This specific trajectory was generated due to the model reorientation specifics. Other
simulated results, and more detailed results, can be found in [14].
5 Conclusions and Future Work
This paper has shown the fundamentals of the ASL/NSL/MissionLab neural schema
architecture for autonomous robots. A previous architecture is described in [19].
An important aspect of this architecture is the ability to incorporate adaptation and
learning through neural network processes in developing new behavioral architectures
for autonomous agents [16] as well as robots. This goes beyond architectures where
behaviors are described in terms of global states or architectures limited in terms of
adaptation and learning mechanisms. Furthermore, as models become more complex
in their nature, the distributed and concurrent nature of the ASL/NSL/MissonLab
architecture becomes of even greater importance. The prey acquisition model
presented in this paper reproduces one of a number of behavioral experiments with
toads. Other experiments are currently being tested under this architecture, in
particular, extensions to the toad's and praying mantis prey acquisition and predator
avoidance models as they are modulated by learning processes [17]. Furthermore, we
are also in the process of experimenting with these models with actual robots in the
real world [6].
References
[1] Arbib, M.A., Levels of Modelling of Mechanisms of Visually Guided Behavior,
Behavior Brain Science 10:407-465, 1987.
[2] Arbib, M.A., "Schema Theory", in the Encyclopedia of Artificial Intelligence, 2nd
Edition, Editor Stuart Shapiro, 2:1427-1443, Wiley, 1992.
[3] Arkin, R.C., "Motor Schema-Based Mobile Robot Navigation", International
Journal of Robotics Research, Vol. 8, No. 4, August 1989, pp. 92-112, 1989.
[4] Arkin, R.C., "The Impact of Cybernetics on the Design of a Mobile Robot
System: A Case Study", IEEE Transactions on Systems, Man, and Cybernetics,
20:6, pp. 1245-1257, Nov/Dec. 1990.
[5] Arkin, R.C. and Balch, T., "AuRA: Principles and Practice in Review", Journal of
Experimental and Theoretical Artificial Intelligence, 9:2-3, pp. 175-189, April-
Sept. 1997.
[6] Arkin, R.C., Cervantes-Perez, F., and Weitzenfeld, A., "Ecological Robotics: A
Schema-Theoretic Approach", to appear in "Intelligent Robots: Sensing,
Modelling and Planning", Editors R.C.Bolles, H.Bunke, and H.Noltemeier,
World Scientific, 1997.
[7] Beer, R. D., Intelligence as Adaptive Behavior: An Eperiment in Computational
Neuroethology, San Diego, Academic Press, 1990.
[8] Briot, J.-P., Gasser, L., "From Objects to Agents: Connections between Object-
Based Concurrent Programming and Distributed Artificial Intelligence", IJCAI
'91 Workshop on Objects and AI, 1991.
[9] Brooks, R., "A Robust Layered Control System for a Mobile Robot", IEEE
Journal of Robotics and Automation, 2:14-23, March 1986.
[10] Calderas, C., Mármol, S., "Diseño e Implementación de un Sistema Distribuido
Heterogéneo para la Extensión del Lenguaje de Modelos Teórico Esquemáticos
ASL", ITAM Undergraduate Thesis, CANNES-TR-96-01, ITAM, México,
Octubre 1996.
[11] Cervantes-Perez, F., Franco, A., Velazquez, S., Lara, N., 1993, "A Schema
Theoretic Approach to Study the 'Chantitlaxia' Behavior in the Praying Mantis",
Proceeding of the First Workshop on Neural Architectures and Distributed AI:
From Schema Assemblages to Neural Networks, USC, October 19-20, 1993.
[12] Cervantes-Perez, F., Herrera, A., and García, M., "Modulatory effects on preyrecognition
in amphibia: a theoretical 'experimental study'", in Neuroscience:
from neural networks to artificial intelligence, Editors P. Rudomian, M.A. Arbib,
F. Cervantes-Perez, and R. Romo, Springer Verlag Research Notes in Neural
Computing vol 4, pp. 426-449, 1993.
[13] Cliff, D., Neural Networks for Visual Tracking in an Artificial Fly, in Towards a
Practice of Autonomous Systems: Proc. of the First European Conference on
Artifical Life (ECAL 91), Editors, F.J., Varela and P. Bourgine, MIT Press, pp
78-87, 1992.
[14] Corbacho, F., Arbib M. "Learning to Detour", Volume 3, Number 4, pp 419-468,
Adaptive Behavior, 1995.
[15] Corbacho, F. Arbib, M. A., "Learning Internal Models to Detour" Society for
Neuroscience. Abs. 624.7, 1997.
[16] Corbacho, F. Arbib, M. A. "Schema-Based Learning: Towards a Theory of
Organization for Adaptive Autonomous Agents" Proceedings of the First
International Conference on Autonomous Agents, Marina del Rey, California,
1997.
[17] Corbacho, F., Khort, B., Lin, B. Nothis, A. & Arbib, M. A. "Learning to Detour:
Behavioral Experiments with frogs" Proceedings of the Workshop on
Sensorimotor Coordination: Amphibians, Models, and Comparative Studies.
Sedona, Arizona, 1996.
[18] Draper, B., Collins, R., Brolio, J., Hanson, A., Riseman, E., "The Schema
System", Int Journal of Computer Vision, 2:209-250, 1989.
[19] Fagg, A., King, I., Lewis, A., Liaw, J., Weitzenfeld, A., "A Testbed for
Sensorimotor Integration", Proceedings of IJCNN '92, Baltimore, MD, 1:86-91,
1992.
[20] Flores, L.R., "Modulación de Tiempo Variable de la Elicitación de la Conducta
de Captura de Presas en los Anfibios Anuros", UNAM Master Thesis, CANNESTR-
97-01, ITAM, Mexico, DF, Agosto 1997.
[21] Khatib, O., "Real-time Obstacle Avoidance for Manipulators and Mobile
Robots", Proc. IEEE International Conference Robotics and Automation, pp.
500-505, St. Louis, 1985.
[22] Lyons, D.M., Arbib, M.A, "A Formal Model of Computation for Sensory-Based
Robotics, IEEE Trans. on Robotics and Automation", 5:280-293, June, 1989.
[23] MacKenzie, D., Arkin, R.C., and Cameron, R., "Multiagent Mission
Specification and Execution", in Autonomous Robots, 4:1, pp.29-52, Jan.1997.
(Also appears in Robot Colonies, ed.R.Arkin and G.Bekey, Kluwer Academic
Publishers, 1997.)
[24] Olivares, R., "Integración del Ambiente de Simulación de Robótica
Neuroetológica: NSL/ASL/MissionLab", ITAM Undergraduate Thesis,
CANNES-TR-97-02, ITAM, México, DF, Septiembre 1997.
[25] Weitzenfeld, A., "ASL: Hierarchy, Composition, Heterogeneity, and Multi-
Granularity in Concurrent Object-Oriented Programming", Proceeding of the
Workshop on Neural Architectures and Distributed AI: From Schema
Assemblages to Neural Networks, Oct 19-20, USC, October 19-20, 1993.
[26] Weitzenfeld, A., Arbib, M., "A Concurrent Object-Oriented Framework for the
Simulation of Neural Networks", Proceedings of ECOOP/OOPSLA '90
Workshop on Object-Based Concurrent Programming, OOPS Messenger,
2(2):120-124, April 1991.
[27] Weitzenfeld, A., Arbib, M.A., "NSL - Neural Simulation Language", in Neural
Networks Simulation Environments, Editor J. Skrzypek, Kluwer, 1994.
[28] Weitzenfeld, A., Arbib, M.A., Cervantes-Perez, F., Rudomin, P., and Alexander,
A., 1998, "Multi-level Simulation Methodology: A Computational and
Experimental Approach to Neural Systems", in Proc. of 1998 NSF Design and
Manufacturing Grantees Conference, pp 651A-652A, Jan. 5-8, Monterrey,
Mexico.
[29] Yonezawa, A., Tokoro, M., Eds., 1987, Object-oriented concurrent
programming, MIT Press.
Posts
No comments:
Post a Comment