Abstract
Many robotic applications, such as search and rescue, disaster relief, and inspection operations, are often set in unstructured environments that typically have communication constraints. In such environments, a multi-robot system must either be deployed to remain constantly connected, sacrificing operational speed and efficiency, or allow disconnections considering when and how to regroup. In this thesis, we insist that the latter approach is desired to increase operational efficiency and create more robust and predictable reasoning within the multi-robot system during disconnection. However, planning in unstructured and unpredictable environments when communication is unavailable requires computing an intractable sequence of possibilities. To address these challenges, we propose a novel epistemic planning approach to propagate beliefs about the state of the system during communication loss to ensure cooperative operations. If changes occur at runtime, robots must understand the social aspect of a scenario to interact with the agents around them alongside achieving mission objectives and ensuring logical belief and planning updates. Planning actions socially requires a robot to infer the intentions and beliefs of other agents, empathizing to predict what other agents want and know about each other. The capacity to reason about the perspective of another agent is the foundation of "theory of mind" which enables the "I know that you know that I know" paradigm without the need for direct and constant communication among actors. Using this architecture, we can increase the operational effectiveness of multi-robot systems during disconnected operations, allowing robots to reason about the capability of other robots and plan according to local observations and simulated beliefs. The contribution of our approach includes: i) a dynamic rendezvous location and decision-making algorithm using risk estimations and multi-objective weighted sum optimization for faster information relay, ii) an epistemic planning formulation, formalizing beliefs and knowledge for consensus-based coverage while disconnected, iii) a generalized epistemic task assignment and gossiping protocol for complex multi-robot tasks with considerations for connectivity constraints and failures during operations and iv) an efficient runtime plan adaptation framework that leverages active inference to reason about the goals of others and signal to others their own knowledge and intentions in communication denied environments. We apply our contributions to both homogeneous and heterogeneous robotic systems, taking into account the various capabilities and limitations of each robot, such as failures, disturbances, dynamics, and sensing capabilities. Our contributions are validated through comprehensive simulations and experiments. Additionally, we include a discussion on incorporating reinforcement learning into our multi-robot epistemic planning framework, aiming for a more adaptable system that can handle runtime uncertainties and form dynamic optimal sub-teams, thereby making our approach scalable for large multi-robot systems.
