Multi-Robot Connectivity-Formation, Active Sensing and Control in Cooperative and Adversarial Settings

Abstract

Progress in robotics technologies have led to an increasing interest in the usage of mobile robots for a variety of tasks. Expanding their operational domains from controlled environments and indoor spaces, robots have started being used in outdoor settings. However, there are still challenges we need to overcome before reliably deploying mobile robots at scale. One of these challenges is the problem of connectivity. When operating as a group, the robots usually need to communicate with each other if they are jointly working on a task together. Yet, in practical applications, robots often lose communication with each other due to being out of the connectivity radius or sensor malfunction. Resilience to, and recovery from such failures in the communication network is crucial for safe deployment. In addition to communication faults, the robot's measurements to carry out a task, or information about its motion dynamics may be imperfect. Active planning provides a promising solution to overcome uncertainties in sensor measurements or motion dynamics. The core of this dissertation is divided into three parts which focus on the problems of multi-robot connectivity-formation, active sensing with imperfect sensors, and control of robotic systems with uncertain dynamics. The first part is dedicated to the problem of network formation for a large group of robots who may be initially disconnected. We tackle the following question: Where should the robots move in order to establish a connected network as quickly as possible? We study offline and online versions of this problem, which also correspond to centralized and decentralized settings, respectively. In the offline version, the robots have two modes of communication: a long-range but low-bandwidth mode for exchanging small amounts of data such as position coordinates, and a short-range but high-bandwidth channel for sharing large-sized data such as images. The robots in this case are tasked with forming a high-bandwidth network by exchanging coordinates over the low-bandwidth channel. In the online case, the robots have only one mode of communication, therefore starting from an initially disconnected configuration, they need to search for each other to form a connected network. For both these versions of the problem, we present algorithms with theoretical guarantees and verify our results in simulations. We also provide proof-of-concept implementations of our algorithms on various multi-robot platforms so as to demonstrate their practical feasibility in real-world settings. In the second part of the thesis, we study the problems of active localization and tracking of targets. In this case, the robot is tasked with either localizing a set of static targets or tracking a single target. In the active localization setting, the robot is equipped with a noisy sensor obtaining relative measurements, such as bearing or range observations. One challenging aspect of obtaining these noisy measurements is due to the unbounded uncertainty of the information a single measurement provides. Thus, the robot needs to aggregate multiple noisy measurements over a time horizon and optimize its trajectory. In the active tracking version, there are two challenges: the target's motion can be adversarial and the robot's visibility is constrained. The visibility constraints of the robot increase the complexity of the search problem since the target can escape from multiple different paths that are not directly visible to the robot. Our contribution to both of these problems is to introduce novel representations for the robot to represent its belief state for the target positions. Specifically, we present learning-based methods to merge robot's measurements and use these compressed belief states for efficiently computing the robot's actions. In the last part of the thesis, we study the challenging problem of controlling a system (e.g., the robot) without knowing its dynamics model. This is a difficult problem especially when the state space is high dimensional and the system has complex dynamics. Existing approaches either generate control laws using computationally expensive optimization, or they are sample inefficient during training. Our approach is to learn the dynamics from data collected by interacting with the system and to train a controller using loss functions adapted from the continuous-time analog of the Bellman equation that integrates the learned dynamics. The resulting controller can be used in real-time to navigate high-dimensional systems with uncertain dynamics and rearrange a group of robots without any collisions. Overall, this dissertation provides methods for robot systems to accomplish complex tasks such as active localization and tracking of multiple targets. The practical feasibility of our approaches are validated in a set of simulated and real-world experiments. We believe that our studies will contribute to the deployment of robots at large scales in scenarios where they need to plan under sensing or dynamics uncertainties.