Experience based action planning for environmental manipulation in autonomous robotic systems

The ability for autonomous robots to plan action sequences in order to manipulate their environment to achieve a specific goal is of vital importance for agents which are deployed in a vast number of situations. From domestic care robots to autonomous swarms of search and rescue robots there is a need for agents to be able to study, reason about, and manipulate their environment without the oversight of human operators. As these robots are typically deployed in areas inhabited and organised by humans it is likely that they will encounter similar objects when going about their duties, and in many cases the objects encountered are likely to be arranged in similar ways relative to one another. Manipulation of the environment is an incredibly complex task requiring vast amounts of computation to generate a suitable state of actions for even the simplest of tasks. To this end we explore the application of memory based systems to environment manipulation planning. We propose new search techniques targeted at the problem of environmental manipulation for search and rescue, and recall techniques aimed at allowing more complex planning to take place with lower computational cost. We explore these ideas from the perspective of autonomous robotic systems deployed for search and rescue, however the techniques presented would be equally valid for robots in other areas, or for virtual agents interacting with cyber-physical systems.