This thesis has established a framework that enhances the decision making of au- tonomous systems based on thesurvivability paradigm. It has introduced survival-like capabilities as part of the underlying architecture of autonomous systems, enhancing the effectiveness of their decision making process. The theoretical and heuristic formulation were demonstrated as a proof-of-concept and feasibility via several use cases.
Part I. Conceptual Framework and Foundations
Chapter 2 has shown that the success of existing autonomous systems depends on the integration of disparate robotic capabilities, but these are constrained by their respective limitations. To overcome them, it is proposed that by enhancing the decision making process of autonomous systems, it may be possible to maximize the capabilities found in these systems. This could be achieved by extrapolating the mechanisms that promote survivability in biological systems, and using them as principles for the design of artificial systems to enhance their autonomous operation.
The contribution of Chapter 3 is in reconciling principles ofsurvivabilityin biological systems with their potential applicability to autonomous systems. While notions of survivability have been defined in other domains (e.g. information systems), they do not directly account for the autonomous operation of a system. Nonetheless, the survivability specificationfrom the information systems domain provides a transition of these notions to the autonomous systems perspective, forming the initial basis for an architectural framework.
The investigation ofsurvivabilityin biological systems has highlighted the interplay betweenaffect,motivationandcognitionin generating self-preserving behaviours. The needsof a system are akin to the intrinsic motivations within an organism that assist self-sustaining processes in maintaining homeostasis. Survivabilityis thus understood as the ability of a system to fulfill itsneeds. While needsmotivate behaviour, analogy with neurological processes suggests that emotions can be modelled as regulatory mechanisms that alter behaviourvis-à-visgoal-prioritization and situation-assessment especially when subject to danger.
The underlying principles of survivability are formulated into the Survivability Framework in Chapter 4. This framework is directed at the design of autonomous systems and their system architectures by emulating the interactions betweenneeds, emotions, andactions. The objective is to leverage on such mechanisms and model them into forms suitable for implementation on an artificial system.
The structure of the Survivability Framework is defined by the four functional perspectives, namely ENVIRONMENT, PREFERENCES, MECHANISMS and CAPABILITIES. Defining the framework around these perspectives would allow a system developed using the framework to meet thesurvivability specificationby design. A formulation ofneeds andemotionsis provided in the same chapter, where a system’sneedsandemotionsare extensions of its internal state representation, orego-state. The emotionsof a system
are modelled from the degree of fulfilment of itsneeds, while the relationships between emotions and needs (and their fulfilment) are modelled as correlations. Mapping the state-space model of a system to itsneeds-space has revealed thatsurvivabilitycan in theory be achieved by designing the needs-state-space system to be controllable. The designer’s task is therefore to determine such correlations (betweenneeds,emotionsand actions), or subject these to determination by other techniques (e.g. machine learning).
The difficulty is that such correlations are not always obvious or straightforward to obtain. Consequently, the experimental implementation showing the feasibility of this approach is limited. Nonetheless, the proposed model makes it possible to reason about thedesignof a system based on itsneedsusing theSurvivability Framework. Together with a set of proposed survivability metrics, the effectiveness of a system can thus be evaluated.
Part II. Framework Realization and Demonstration
The feasibility of the proposed Survivability Framework has been demonstrated in Chapter 5. The realization of theSurvivability Reference Architecture in this chapter is discussed with respect to the four functional perspectives of the framework. The realization of the ENVIRONMENT perspective has shown that certainneeds(e.g. safety) can be evaluated from features obtained from the environment, showing the ease with which needs can be numerically computed using mathematical models. The identification of theneedsin the PREFERENCESphase has shown the manner in which task specifications implicate theneedsandemotionsof a system. The manner in which actionsare identified to fulfilneedsis discussed in the CAPABILITIESperspective. Finally the MECHANISMS phase has shown the realization of several system architectures for the scenarios in question, demonstrating the viability of the framework in directing the design of autonomous systems.
The proposed approach is demonstrated in Chapter 6 through experiments and simulations. The relevance of survivability as a determinant of decision making is shown with a selection of use cases to validate the framework and its premises. These use cases have shown the feasibility of usingneedsas an alternative representation of the state of a system and how they may be used to determine the actions, or behaviour of the system in both indoor and outdoor environments. The first use case has shown the use ofemotionsto monitor the progress ofneeds-fulfilmentduring the operation of the autonomous vehicle, thus demonstrating the correlations betweenneeds andemotions.
The final use case has extended the earlier results by demonstrating the interactions between needs, emotionsand actions in determining the behaviour of an autonomous surveillance vehicle as it patrols an indoor environment while detecting, tracking and following intruders. This use case has demonstrated the use of emotions as another mechanism that intensifies or diminishes the needs-actionsactivations, emulating the regulatory action of emotions on the behaviour of biological systems in safeguarding their survivability.
Contributions of this thesis
This thesis has proposed a paradigm for the design of autonomous systems, based on the introduction ofsurvivabilityas one of the factors influencing the decision making of autonomous systems. In addition, this thesis has provided a definition ofsurvivabilityas applied to autonomous systems. To the best of our knowledge, this is the first attempt at defining notions ofsurvivability for an autonomous system that is based on the ability of a system in meeting its needs. Furthermore, this thesis has provided a model for emotionsthat is defined in terms of theneedsof a system.
By centering the framework on the principles of survivability obtained from an analysis of biological systems and the mechanisms which assist in their survival, the framework offers a cognitive perspective for existing system architectures. This cognitive perspective is achieved through a transformation of concepts from human motivation and needs, affect and emotions, as mechanisms for decision making in autonomous systems.
An analysis of possible architectural realizations of the Survivability Reference Architecture, for instance, layered, behaviour-based, and subsumption-type architec- tures, has been carried out. Besides making it possible to realize these architectures, an additional contribution of this framework is in offering a cognitive explanation for system architectures in terms of the interactions between needs, emotions and actions. Consequently, this thesis has provided areference architecturefor autonomous systems, which encapsulates the interactions between the concepts of survivability, needs,emotions, andactions.
The framework has been shown to be realizable in both hardware and software implementations, whereaction-selection is performed usingneedsto determine action- activations and emotionsto influence the strength of such activations. As part of the validation process, several additional contributions have been made. This includes a
fuzzy-logic-based feature perception algorithm that integrates multi-modal information about the local environment such as its colour composition, presence of obstacles or traversable regions, and regions-of-interest. This algorithm is applied on stereovision and colour images obtained from a trinocular stereovision camera system, and a purpose- built omnidirectional camera system. In addition, a method for target-tracking and following has been developed for Use Cases 5 and 6 in Chapter 6, controlling both the vehicle’s motion and pan-tilt-zoom camera in closed-loop.
In exploring the cognitive notions of needs and emotions as representations of the internal state and observations of the autonomous system, the development of the Survivability Framework in this thesis has provided a possible way of realizing such notions onto actual systems. The objectives and contributions achieved in thesis serve as the basis for further incorporation of additional insights from biological systems to the engineering of artificial systems.