| Autonomous systems are increasingly prevalent in many application areas such as military Command and Control, industrial manufacturing, and medical systems. However, analysis of such systems becomes intractable when modelled by classical methods like differential/difference equations. Discrete-event systems are increasingly being used to model and control such systems and a body of literature called supervisory control theory (SCT) has emerged. Traditionally, SCT uses a qualitative method called maximal permissiveness to arrive at optimal supervisors. A quantitative performance index based on a measure of formal languages has been proposed to overcome the shortcomings of qualitative design in SCT. The research reported in this thesis has adopted the concept of language measure to develop a comprehensive framework for decision & control of autonomous systems. A consistent framework based on language theory is built to address sensing, planning and action aspects of autonomous systems. The proposed framework is demonstrated by its application to well known problems in robotics such as obstacle avoidance and path planning. The discrete-event models are constructed and pertinent parameters have been identified from the physical dynamics of the system via a system identification. This approach creates a seamless transition from the continuous dynamics to the discrete-event control by eliminating the need to manually design the discrete-event models of the plant dynamics and control specifications. In the language-theoretic setting, autonomous perception is also addressed using a tool called symbolic dynamic filtering (SDF). Application of SDF as a feature extractor in a domain independent manner is demonstrated and pattern classifiers are constructed to recognize behaviours autonomously. A new type of pattern classifier based on the language measure is proposed and is shown to be faster than traditional Bayesian classifier with comparable robustness. The proposed feature extraction and classification methods are validated on a distributed network of piezoelectric sensors to classify the behaviors of two types of robots in real time. |
