Design and optimization of a decentralized multi-robot exploration behavior taking into account energy constraints

The robot revolution is just around the corner. Robots have already started appearing in public spaces and are becoming available to everyone. However, most robots today still need a human in the loop supervising, recharging batteries, or making decisions, especially in the face of uncertainty. There are still some pieces of the robot autonomy puzzle that are still either missing or that need to be refined to enable robots to work completely on their own. One such piece is the robot exploration problem: how can robots explore very large spaces without human intervention? This piece is crucial for tasks, such as space exploration or disaster relief. This thesis work focuses on designing and optimizing a distributed exploration strategy called Decentralized Frontier-based Exploration (DFBS). This strategy aims to allow every robot to make its own decisions based on their perception of the world by using a shared map created by all the robots in the group. This work builds on top of the Frontier-based exploration strategy, that defines a frontier as the borderline between explored and unexplored space, and extends it by using of a fully decentralized approach that tackles fault tolerance, and also considers robots with limited energy reserves and their replenishment. Field and service robotics has been one of the main foci of the Center of Excellence in Generic Intelligent Machines, part of the Automation and Systems Technology department of Aalto University. In this laboratory, the MarsuBot robot fleet was created as a test bed for multi-robot exploration. This work aims to contribute to the efforts of the laboratory by creating a simulation environment called MarSim, in which the MarsuBot fleet can be simulated, and the DFBS strategy can be implemented and optimized. Experiments were performed using MarSim, simulating scenarios of varying difficulty to evaluate the performance of the proposed exploration strategy in comparison with a purely stochastic and reactive strategy. Over six million simulations were performed in Triton, Aalto University's supercomputer, to search for the best parameter combination that minimized the energy spent by the whole fleet while exploring the different scenarios. These results were compared to the ones obtained by an off-the-shelf Bayesian Optimization tool running on a single computer. The results offer a complete analysis of the performance of the Decentralized Frontier-based Exploration in comparison to a basic reactive behavior, along with the optimization of its parameters using Bayesian Optimization and verification of the results by comparing them to the experimental results obtained using Grid Search. This work also analyzes the performance of Bayesian Optimization as a tool to optimize the starting parameters of a robot in a very small amount of experiments to evaluate the possibility of using it to optimize the real MarsuBot fleet.