Multi-Robot Odor-Source Searching Based On Reinforcement Learning

In real life and industrial production processes, accidents of toxic and harmful odor/gas leakage are of common occurrence, which have caused great harm to human beings. So it has important significance to localize the odor/gas leakage sources. In the early 1990s, inspired by the olfaction abilities of creatures, researchers started the research work in the active olfaction field. The so-called active olfaction means finding, tracking plumes and finally declaring the odor source actively by using mobile robots. Studies of robot active olfaction have great applicable values in many areas like judging toxic/harmful gas leakage locations, checking contrabands and searching for survivors after disasters.This thesis mainly studies on the multi-robot based odor source searching in turbulence dominated airflow environments. The main contributions are listed as follows:Firstly, the research background and the significance of odor source localization (OSL) are briefly summarized. Besides, the application environments, the research status, the application prospect and the chief difficulties related to the OSL are analyzed.Secondly, in order to establish the simulation platform for the existing indoor experimental site, a two-dimension indoor dynamic flow field model used for analyzing and verifying the odor source searching algorithms is built by using the FLUENT software of computational fluid dynamics, and a two-dimension indoor plume model is established by combining the two-dimension indoor dynamic flow field and the filament based plume model. The comparison results show that the built indoor dynamic flow field has good consistency with the measured data, and it can meet the requirement of simulation researches on dynamic plume tracing.Thirdly, aiming at the odor-source searching in turbulence dominated airflow environments, a strategies sharing based multi-robot reinforcement learning method is adopted for training the robots to finish the odor-source-searching tasks rapidly and reliably. A lot of simulation experiments in both the outdoor large-scale and indoor plume platforms show that the success rate and efficiency of the multi-robot odor source searching could be improved by using the strategies sharing and group reinforcement learning.