Abstract
Planning of spacecraft operations has historically been performed manually by ground operations staff, although there is an increasing trend towards supporting these activities with more and more sophisticated software tools. This has now reached the stage where automated planning systems are used at some ground stations to generate complete schedules for batches of operations that are simply reviewed and adjusted by manual operators. The increased automation of planning promises to reduce the cost of operating spacecraft. This batch planning approach relies on the spacecraft being able to execute the entire schedule as planned. In practice, however, models used to support planning cannot fully predict the actual behaviour of the system and other factors that will affect the outcome of the plan. Images taken of the Earth, for example, may be obscured by cloud, while a Martian rover may take longer than expected to cover a piece of terrain. A spacecraft that is able to modify its own operations schedule may be able to respond intelligently to adverse outcomes rather than simply abandoning its plans until ground operators provide new instructions. This would also enable it to take advantage of unforeseen opportunities to achieve goals that could not be identified when planning far in advance. This 'schedule repair' approach to spacecraft autonomy has been a popular line of research, whereby a spacecraft is able to make incremental changes to its operations schedule in light of an unexpected outcome. The most notable practical example is the Autonomous Science Experiment. This thesis presents a new approach to spacecraft autonomy that gives the vehicle greater freedom to make its own decisions. Here, an evolutionary algorithm is used on board to continuously search for the optimal operations plan in light of the most recent information available. As circumstances change and as new goals arrive the optimal solution to the planning problem will move within the search space and the evolving population of plans will follow it. The required effort from ground station operators is further reduced by removing the need to group goals into batches to be performed over set planning intervals. Instead goals can be sent to the on board planner in an ad hoc manner. A key feature of the search algorithm used here is that it is able to make immediate short-term decisions using the fittest member of the current population whilst allowing later plans to evolve further. In this way the operations schedule is continually optimised in parallel with execution. An implementation of this planning approach was tested on two case studies, scheduling imaging operations on the UK-DMC Earth observation satellite and planning a planetary rover's route around a set of science targets. In both cases dynamic aspects of the planning problem were simulated such that outcomes were non-deterministic. The planning approach developed in this work was found to outperform a pre-optimised batch planning approach in these situations. The algorithm was also tested on a goal set taken from the imaging history of UK-DMC and found to yield a higher rate of return of science data. Furthermore, a model rover was used to demonstrate this planning approach in real-time practical experiments.