Abstract
This thesis describes a framework for the high level control of an autonomous unmanned spacecraft. Greater autonomy than currently exist is required for unmanned spacecraft to enable missions to distant planets and bodies. One reason for this is that the signal return time is too long to accommodate real-time control from the ground. A second reason is that spacecraft travelling to bodies where little is known of the environment (e.g. asteroids) must have the capability to respond to unplanned events. In addition, autonomy can help reduce mission operations costs, a very important factor in the current climate where more is expected from space missions at a lower cost. The thesis proposes a novel architecture for an autonomous unmanned spacecraft, based on Distributed Artificial Intelligence (DAI), and more specifically based on the multi-agent paradigm. The proposed model for spacecraft control is decentralised. In this architecture, the spacecraft is made up of agents; the traditional ground-based controller is one agent. The spacecraft is goal-driven; it receives high level goals from the ground. The planning and scheduling of activities to achieve these goals is carried out on-board the spacecraft. The spacecraft is also event-driven; it reacts to events that occur on-board the spacecraft as well as in the environment. A DAI architecture requires a co-ordination mechanism, and a communication structure. Also, distributed versions of algorithms must be provided. In this thesis, co-ordination with and without explicit communication and distributed scheduling were investigated, and a framework proposed for both these issues. An autonomous spacecraft must have inference capability for on-board decision making to enable it to respond to unplanned events. Probabilistic reasoning in the form of Bayesian networks was used to provide the spacecraft with the capability for on-board decision making. Situations may arise where the spacecraft must make decisions with uncertain or incomplete information. The issue of decision making with uncertain or incomplete knowledge (e.g. co-ordination without explicit communication) was investigated using domain specific scenarios. Spacecraft resources are typically very limited in capacity. On-board resource management should result in more efficient use of resources. A framework for an on-board resource manager was defined and implemented using reinforcement learning. A distributed version of the scheduling algorithm using reinforcement learning was developed. Thus, this thesis describes and investigates an architectural framework for a multi-agent approach to spacecraft control.