This thesis presents a recollection of developments and results towards enabling reactive task allocation in dynamic environments for large-scale autonomous robotic systems. Achieving such coordination of robotic teams is challenging in multiple ways, including how to design the local autonomy and the overall team orchestration. Although modern advances, with high-performing methods for localizing based on the environment and high-performance on-board computation units, are enabling the deployment of large systems in complex environments, the deployment of multi-agent systems has proven to be a significant challenge in robotics. This thesis explores different aspects of the robotics research field, including reactive task allocation for coordinating teams of agents in settings where the specific tasks are unknown a priori and how the local autonomy can be synthesized based on the specific task, and agent capabilities, at hand.

The articles included in this thesis are primarily focused on presenting key contributions towards three main research directions. First, a reactive auction-inspired task allocation framework is presented and demonstrated in environments with limited a priori information of the specific tasks to be completed. This multi-agent architecture is shown to be effective in realistic laboratory environments using multiple ground agents. Second, the creation of the necessary local autonomy for completing the tasks as prescribed by the task allocation framework is explored. A method for synthesizing behavior trees, based on the individual robot capabilities, for specific tasks is presented and evaluated for enabling more resilient and flexible overall mission execution. Finally, the third research direction lies within deploying such systems in harsh large-scale environments. Towards this direction, large-scale experiments are presented, showcasing the possibility to deploy teams consisting of multiple aerial vehicles in underground mines to perform routine inspection missions.