Probabilistic Methods for Robotic Gas Source Localization.

Abstract

The field of robotics is a complex one. Many things can go wrong when an autonomous machine interacts with the real world, from sensors producing unreliable measurements to actuators behaving in unpredictable manners. A robot operating with such unreliable tools must account for the uncertainty that they cause in its estimations of the state of the world.

Probabilistic robotics is an approach that embraces this uncertainty by modeling all the knowledge and estimation processes of the robot with the formal tools of probability theory. This creates a rigorous, robust way of handling uncertainty even as the complexity of the problems increases. Such an approach is beneficial in most cases, but becomes crucial when the information that the robot obtains can only be related to the state it tries to estimate through unreliable models.

One such problem is gas source localization (GSL), where a robot equipped with a gas sensor and an anemometer must solve a fluid dynamics problem to figure out which point in the environment a gas is being emitted from. Due to the complexity of the phenomenon of gas dispersion, GSL techniques must make assumptions and simplifications to be able to reason about it, which increases uncertainty.

In this thesis, we tackle the subject of source localization with an autonomous mobile robot from the perspective of probabilistic robotics. We explain the difficulties the problem presents, and propose solutions based on probabilistic modeling and Bayesian reasoning.