2.5 Architectures for autonomous systems
2.5.4 Hybrid and biologically-inspired architectures
To overcome the problems with both hierarchical top-down and behavioural-based bottom-up approaches, architectures such as AuRA (Autonomous Robot Architecture) (Arkin and Balch, 1997) have emerged (Figure 2.6). AuRA is a hybrid architecture with both hierarchical, layered deliberative components that performs mission-planning, spa- tial reasoning and the scheduling of plans, as well as reactive, behavioural capabilities based on the notion ofmotor-schemas(Arkin, 1989).
Apart from hybrid architectures, system architects are obtaining more insights from biological systems in nature. Self-organization, emergent intelligence, and artificial life are just some of these. The success of biological entities in survival can be emulated by forming computational models of such processes. From nature, several aspects of autonomous behaviour can be observed. Self-organization in social animals such as swarming behaviour is an example of emergent intelligence. Emergence refers to the holistic capability arising from a collection of components interacting through the interactions the agent has with its environment (Arkin, 1998). The
N
S W E
perception mission
10 s
1 s
0.1 s
Modules
communication
OPERATOR reports
state
requests reports
control
proprioceptive sensors proximetric
sensors effectors exteroceptive
sensors
ENVIRONMENT Decisional LevelExecution Control LevelFonctional LevelLogical SystemPhysical System
Executive
servo−control control reflex actions Plan Supervisor
Task Supervisor Task
Refinment Mission Planner
Sensor and Effector Interface
monitoring modelling requests
Figure 2.4: The LAAS Architecture for Autonomy (Alami et al., 1998)
criteria for being deemed emergent is the element of surprise (Ronald and Sipper, 2001). Control architectures designed on these principles have gained ground in recent years. A demonstrably successful architecture known as 4-D/RCS is currently used as the reference model for the Demo III XUV Programme (Albus and Meystel, 2001). It comprise a hierarchy of computational nodes each comprising behaviour generation (BG), world modelling (WM), sensory processing (SP) and value judgement (VJ) processes, with intentional provision for emergent intelligence to arise (Figure 2.7).
While fundamentally different from the 4-D/RCS Reference Model, the EMIB Compu- tational Architecture (Michaud, 2002) is similarly designed for exploiting such emergent properties (Figure 2.8). EMIB incorporates motivations and emotions for intentional selection and configuration of behaviour-producing modules. It is a hybrid architecture with both biologically-inspired and cognitive science motivations, promoting not just
Figure 1-2 : DAMN Arbiter and Behaviors.
MAINTAIN HEADING AVOID
OBSTACLES AVOID
TIP-OVER ARBITERDAMN
GOALSEEK
FOLLOW ROAD
commands weights
votes MODE
MANAGER VEHICLE
CONTROLLER
Figure 2.5: The DAMN Architecture (Rosenblatt, 1997), showing behaviours and arbiters.
Schema Controller Motor Perceptual
R E P R E S E N T A T I O N Mission Planner
Spatial Reasoner
Plan Sequencer
Actuation Sensing
Hierarchical Component
Reactive Component Learning
Plan Recognition User Profile
Spatial Learning
Opportunism
On-line Adaptation
User Input
User Intentions
Spatial Goals
Mission Alterations
Teleautonomy
Figure 2.6: The Autonomous Robot Architecture (AuRA) (Arkin and Balch, 1997). This diagram depicts its high-level schematic.
emergent intelligence at the behavioural level but at higher levels of abstraction. This has been explored in separate PhD theses by Michaud (1995) and Allen (2001), the latter which deals with the concept ofconcerns– dispositions for the occurrence or non- occurrence of a given situation – as applied towards the achievement of autonomy in intelligent agents. The tenet of both research accepts the relevance of human emotional states as key components for decision making.
The close relationship between motivation and emotions, and the behaviour of a system have similarly been exploited in many architectures (Allen, 2001; Arkin, 2005; Bonarini et al., 2003; Malfaz and Salichs, 2004; Sloman, 2001; Stoytchev and Arkin, 2001). These are theoretically grounded in affect and cognition (Savage, 2003).
Figure 2.7: The 4D-RCS Reference Architecture (Albus et al., 2002) as implemented for the Demo III Cross-country Unmanned Vehicle (XUV) platform. The architecture comprises Sensory Processing, World Modelling, Value Judgement and Behaviour Generation elements.
EMOTIONS
Behaviour-Producing Module ...
Recommendation Level Motivational
Level
Behavioural Level Behaviour-Producing Module b
Behaviour-Producing Module a IMPLICIT EGOISTICAL RATIONAL
SELECTION
ARBITRATION MOTIVES
Behaviour.
Activation d/u
Behaviour.
Exploitation configuration data parameters
Commands Sensory
inputs
Figure 2.8: The EMIB Computational Architecture (Michaud, 2002).
Unfortunately, a prevalent problem in many such architectures is that emotions are often introduced into the system architecture in an arbitrary manner, ie “externally defined and observer-dependent rather than being grounded in or emergent from actual homeostatic processes (Ziemke, 2008).” This includes the assignment of emotions (namely: happy, confident, concerned, frustrated) to behaviour scripts (Murphy et al., 2002), or patterns of variation to emotions (namely linear increase of curiosity and homesickness, versus linear decrease of frustration and anger) (Stoytchev and Arkin, 2001). In the rest of this thesis, it will be shown how this issue is addressed by grounding emotionsto “homeostatic processes(Damásio, 1994; Ziemke et al., 2005)” that are based on the levels ofneeds, and the extent and rate at which theseneedschange. Therefore, theneedsof a system directly accounts for the emergence of theseemotions.