(375ae) Policy Development for a Mobile Robot to Search for the Source of a Chemical Plume in an Urban Environment

In the event of an accidental or malicious release of a hazardous chemical in a populated environment, such as an industrial plant or urban area, we wish to quickly localize the source.

Mobile robots (e.g. aerial drones) equipped with chemical sensors are ideal for searching for the source of a hazardous chemical plume. From its history of chemical sensor measurements over its path so far, the robot must decide, under uncertainty, the direction in which to travel next. The task of the robot is analogous to an insect searching for a food source or mate based on an aroma or pheromone. However, inferring the source of a chemical plume from a concentration time series along a spatial path is under-determined and nontrivial because time-varying wind patterns, turbulence, diffusion, and deposition of the chemical onto surfaces all occur within a highly complex geometry (e.g. units in a chemical plant or buildings in a city).

To develop and benchmark policies for robots to search for chemical plumes, we are modeling the transport and dispersion of chemical plumes in realistic environments (cities, chemical plants, etc.) using the Quick Urban \& Industrial Complex (QUIC) Dispersion Modeling System. For a given geometry, we can vary the location, intensity, and chemical composition of the plume source and the meteorological conditions to simulate thousands of scenarios. Using this simulated data, then, we will benchmark insect-inspired algorithms to learn an optimal policy for a drone equipped with a chemical sensor and anemometer to search for the plume source. This policy will map the path of the drone so far, and the history of sensor measurements along the path, to an action---a direction to move. The action must balance exploitation and exploration.