Applications of Autonomous Systems to Rapid Environmental Assessments

Abstract

A Rapid Environmental Assessment (REA) is a methodology to quickly analyse a location’s environment and generate an operational view to inform subsequent decision making. Distinct from scientific surveys, REAs facilitate follow–on life–critical operations including amphibious landings and disaster responses, however, existing REA methods are resource intensive, predominantly manual, and risky. Potential solutions centred on autonomous systems are under research, but challenges have hindered their implementation, identified as Human–Robot Collaboration (HRC) issues concerning intuitive interfaces between humans and agents and a lack of trust in autonomously generated information. This thesis seeks to make progress towards these areas. To reduce resources when planning REAs, a novel optimisation framework is formulated to automatically ingest and analyse terrain data to determine optimal REA locations. To maintain trust, the framework is derived from established REA methods and provides human–interpretable results. Validation is performed over two Australian coastal regions, testing a wide variety of planning requirements, with the system proving robust and consistently producing the desired outcomes. To further improve HRC, a coalition formation and tasking method is developed to supplement autonomous planners. The method utilises Fuzzy Sets, allowing users to intuitively describe heterogeneous vehicles and tasks with high accuracy whilst also accounting for uncertain information. A MATLAB simulation was created for validation, with multiple vehicle types and tasks used to represent and test realistic scenarios encountered during REAs. Results are compared to a baseline vector–based system, with the fuzzy method developed here showing close alignment, while providing a more intuitive interface via fuzzy representations. Finally, to quantify risk and uncertainty in the previous results to build trust in the information, a risk framework is also developed. The framework is able to capture various obstacles and uncertainties generically through 2.5D maps, before producing an aggregate map and evaluating a single metric, Conditional Value at Risk (CVaR). A novel use of the CVaR metric provides a user–tunable parameter, enabling increased risk to be traded for decreased REA mission time. Validation is again performed in simulation, successfully capturing multiple risks and demonstrating a reduction in mission time of up to 40% under high risk tolerances.