Autonomous robotics for identifying the dispersion characteristics of gas sources in complex environments

When an uncontrolled gas release is first detected, an immediate response must first be elicited in order to identify the extent of the hazard. This can be in the form of parameterising the gas leak in terms of key information such as release rate, location, diffusivity, amongst other source parameters. This is formally known as source term estimation (STE) and is useful to first responders so that remedial action can be taken to stop the release. Source identification can also be performed in the form of gas distribution mapping (GDM), which characterises the dispersion of the hazard and provides a concentration distribution over space. GDM can be used to dictate exclusion zones, plume boundaries and also be used to identify likely source locations. To attain the data required to generate this information, current techniques require either a preassembled static sensor network or human first responders with handheld chemical sensors. To forgo the need for prebuilt infrastructure and in order to minimise the risk to human operators, the obvious next step is to develop robotic sensing platforms that can operate in complex environments and collect the required data autonomously. This thesis investigates techniques for both STE and GDM with mobile sensors, as well as the required path planning considerations involved with making these systems autonomous.

To aid the autonomous STE task, two new path planning techniques are developed that improve the searching efficiency of an autonomous agent. The first addresses an informative path planning solution for a searching agent operating in cluttered environments, namely “informed tree planning”. This method uses a batch sampling approach to first generate preferential sampling locations within the environment, upon which a tree is then expanded that grows towards the estimated source location. An information theoretic utility function is then used to evaluate the benefit of these potential sampling locations and the trajectory with the highest reward is executed. By performing the search in this manner, the agent inherently moves towards a recursively updating source location estimate whilst taking informative manoeuvres around obstacles. This method is shown to greatly outperform current state-of-the-art deterministic methods in terms of both searching efficiency and reliability. The second algorithm developed investigates performing STE in a-priori unknown environments. To address this operational problem, a rapidly exploring random tree search (RRT*) is performed within a visible radius and the dual control technique is leveraged to drive the agent towards the source location whilst simultaneously choosing locations of high information theoretic importance. 

For GDM, the current state-of-the-art mapping techniques are first compared for their performance in both open and cluttered environments. In the case of a cluttered environment, the Gaussian Markov random field technique is found to be capable of resolving gas distribution maps whilst also accounting for the influence of obstacles on said distribution. This method is brought forward for use within an informative path planning framework, and several utility functions are proposed that automate the GDM acquisition procedure. It is found that considering the posterior estimates of concentration, variance and travel cost into a single metric dramatically speeds up the acquisition of distribution maps compared to the standard lawnmower sweep method of data collection. Furthermore, a new method of resolving GDM, leveraging Gaussian belief propagation, is developed that increases the computational speed of acquiring these maps by several orders of magnitude and also addresses the problem as a local optimisation (as opposed to a global optimisation), in turn making the solver both efficient and scalable to 3D estimation tasks.