Resumen de Distributed consensus in multi-robot systems with visual perception
The idea of a team of robots executing cooperative tasks in an autonomous manner is everyday closer to become a reality. Multi-robot systems can perform complex tasks with more robustness or in less time than one robot working alone. On the other hand, the coordination of a team of robots introduces new challenges that the designers of these systems must face.
A globally consistent perception of the environment is a key component for the proper cooperation of the team of robots, which requires for the robots to communicate their observations to all the other members. When two or more robots have common observations the team needs to reach a consensus combining all of them. This must be done considering the limitations that each robot has, taking into account that not all the robots can communicate with each other.
To this end, we consider as the main objective of this work the development of distributed algorithms that make a team of robots reach an agreement about the information they perceive. We focus our work in solving this problem when the robots perceive the world using vision sensors. These sensors are very useful in many essential robotic tasks like autonomous navigation and mapping, due to the big amount of information a single image contains. However, in a distributed setup the use of these kind of sensors brings up many complications that need to be addressed.
In this Thesis we present a deep study of distributed consensus algorithms and how they can be used by a team of robots equipped with monocular cameras, solving the most important issues that appear because of the use of these sensors. In the first part of the Thesis we address the problem of finding global correspondences between the observations of the different robots. In this way, the robots know which observations must be combined in the computation of the consensus. We also deal with the problem of robustness and distributed outlier detection, giving a solution to discard erroneous measurements. To counteract the increase in the size of the messages caused by the previous steps, we use the properties of Chebyshev polynomials, reducing the number of iterations required to achieve the consensus.
In the second part of the Thesis, we focus on the problems of mapping the environment and controlling the motion of the team of robots. We apply well known computer vision algorithms to reach the consensus in these two scenarios. We show that using structure from motion, requirements such as the direct observation of the other robots during the control loop or the knowledge of a common frame are avoided. In addition, the lack of calibration information is not a major issue using our algorithms. The evaluation of the solutions is done using a large urban data-set and real non-holonomic robots.
All the algorithms presented in this Thesis are well designed to be executed in a distributed fashion by a team of robots with limited communication capabilities. We theoretically prove the main properties of all the proposed algorithms and test their quality using simulated and real data. Specifically, the main contributions of the Thesis are:
*A set of distributed algorithms that make possible for a team of robots equipped with cameras to reach a consensus about the information they perceive. In particular, we propose three distributed algorithms that solve the problem of finding global correspondences between the robots, detect possible outliers and reduce the total number of communication rounds required by the network to achieve the consensus.
*The combination of distributed consensus and structure from motion techniques in multi-robot perception and control tasks. We design an algorithm to cooperatively build a topological map of the environment considering planes as features and using homography constraints to relate the observations of different robots. We also propose a distributed control law using the epipolar constraint to make the team of robots to reach an agreement in their orientations without the necessity of directly observing each other.
