A combined mechanism for UAV explorative path planning, task allocation and predictive placement

The use of Unmanned Aerial Vehicles (UAVs) is becoming ever more common by people or organisations who wish to get information about an area quickly and without a human presence. As a result, there has been a concerted effort to develop systems that allow the deployment of UAVs in disaster scenarios, in order to aid first responders with collecting imagery and other sensory data without putting human lives at risk. In particular, work has focused on developing autonomous systems to minimise the involvement of overstretched first responder personnel, and to ensure action can be taken by the UAVs quickly, co-operatively, and with close to optimal results. Key to this work, is the idea of enabling coordinated UAVs to explore a disaster space to discover incidents and then to allow more detailed examination, imagery, or sensing of these locations.Consequently, in this thesis we examine the challenge of coordinating exploratory and task-responsive UAVs in the presence of prior (but uncertain) beliefs about incident locations, and the combination of their roles together. To do this, we first identify the key components of such a system as: path planning, task allocation, and using belief data for predictive UAV placement. Subsequently, we introduce our contributions in the form of a complete, decentralised system for a single explorative path planner to minimise the time to identify incidents, to allocate incidents to UAVs as tasks, and to place UAVs prior to new tasks being found.Having demonstrated the efficacy of this solution in experimental scenarios, we extend the formulation of our explorative path-planning problem to multiple UAVs by constructing a coordinated, factored Monte-Carlo Tree Search algorithm for use in a discretised space representation of a disaster area. Subsequently, we detail the performance of our new algorithm against uncoordinated alternatives using real data from the 2010 Haiti earthquake. We demonstrate the performance benefits of our method via the metric of people discovered in the simulation; showing improvements of up to 23% in cases with ten UAVs. This is the first application of this technique to very large action spaces of the type encountered in realistic disaster scenarios.Finally, we modify our coordinated exploration algorithm to function in a continuous action space. This represents the first example of a continuous factored coordinated Monte-Carlo Tree Search algorithm. We evaluated this algorithm on the same Haiti dataset as the discretised version, but with a new sensor model simulating mobile phone signal detection to represent the types of sensors deployed by first responders. In addition to the benefits of a more realistic model of the environment, we found improvements in survivor localisation times of up to 20% over the discrete algorithm; demonstrating the value in our approach.As such, the contributions presented in this thesis advance the state of the art in UAV coordination algorithms, and represent a progression towards the widespread deployment of autonomous platforms that can aid rescue workers in disaster situations and—ultimately—save lives