• autonomous search
• mobile robot
• olfaction
• entrotaxis
• infotaxis
• algorithmic
• autonomous vehicles
• active sensing

The objective of this thesis work is twofold. Firstly, a formalization of the Active Sensing problem is provided in the most general sense, with a focus on the history and various algorithms presented over the years for autonomous vehicles. Subsequently, the thesis work specifically focuses on the cognitive search by a vehicle to find an accidental gas dispersion in an environment. Methods such as Chemotaxis, Infotaxis, and Entrotaxis are presented. The work concludes by presenting a specific algorithm, RE-Entrotaxis, through a MATLAB implementation for the control of a unicycle and the search for a gas leak.
The need for new tools to address new requirements has driven the field of Robotics Research and Development towards a new branch called Active Sensing. It aims to actively acquire information from the environment so that the vehicle or robots in general can make autonomous decisions and act accordingly. Among these, algorithms such as SLAM (Simultaneous Localization and Mapping), object search within an environment, or, in a more practical and real-life scenario, the search for a gas dispersion.
Indeed, the latter topic has been of greater interest in the literature. Currently, dogs are used for search and rescue operations, but they are expensive, difficult to train, and can become inefficient after a few hours of work. For this reason, based on animal behavior, several algorithms have been developed that allow for fairly accurate estimation of the position of a gas dispersion source.
Despite the versatility of mobile agents, the success of the mission depends on the chosen algorithm and its effectiveness in real simulations. One of the first algorithms developed for this purpose is Chemotaxis, which involves following the local gradient of a chemical concentration. However, this, and other subsequent algorithms, have problems regarding local maxima/minima that may occur. To address this, algorithms that also consider the influence of wind, called cognitive strategies, have been developed.
The difference between these strategies and the previous ones lies in the cost function that is intended to be maximized so that the autonomous vehicle can find the source, and in the formulation of the problem as a Partially Observable Markov Decision Process (POMDP). Thus, adopting this philosophy, some algorithms have been developed that have achieved great success in the literature. One of these is Infotaxis. With this type of algorithm, the agent chooses the direction that locally maximizes the expected rate of information acquisition. Specifically, the vehicle autonomously chooses, among the neighboring sites on a lattice and stationary, the move that maximizes the expected reduction of entropy of the posterior probability field. Therefore, the search process is designed to acquire information about the source's location. Subsequently, Entrotaxis was introduced, which, unlike Infotaxis, maximizes entropy rather than its expected value.
Thanks to the advent of certain "reinforcement" algorithms, which tend to improve their robustness, search speed, and accuracy in more complex search environments, to address the main problems of previous algorithms, where pure exploitation or pure exploration are not practicable, the RE-Entrotaxis algorithm was developed. This algorithm proves to be computationally efficient, thanks to its ease of use and computational simplicity.
Assuming knowledge of the average lifespan of the air mass, atmospheric diffusivity, sensor radius, sampling time, wind direction and speed, the RE-Entrotaxis algorithm was simulated using MATLAB software, demonstrating its search effectiveness even in turbulent environments. It is shown that this approach can also work in three-dimensional environments at the cost of higher search time and slightly higher computational calculations.