As robotic systems become increasingly prevalent in our lives (e.g., by harvesting food, assisting with disaster response and defense, and transporting persons and our goods around the world) there is a growing need to ensure that they can cooperate to achieve their intended goals. Robotic cooperation in the real-world is inherently a hard problem. There are limitations on both computation and communication abilities, risk of component failure, hard deadlines for decision making to occur, and other actors operating around them that may interfere with individual robots. The algorithms necessary to coordinate robotic teams in fielded applications must perform reliably in spite of these complications and the uncertainty they cause.
This dissertation addresses these problems by developing novel methods of distributed non-myopic robotic coordination. These methods distribute the planning and decision making authority across the team, which diversifies the computational burden of the coordination problem across the team. An additional benefit of distributing the planning authority across the team is the resulting increase in system robustness to communication failures and the removal of the single point of failure of centralized systems. However, distributing planning authority can lead to conflicts between team members when they simultaneously plan their individual actions. To address these potential conflicts, we develop coordination methods that allow individual agents to search for their individual actions while accounting for the likely actions of their team members. These methods focus on the idea of temporal auctions and discounting rewards to resolve conflicts between team members.
The problem for coordinating a team of exploring robots is largely the problem of selecting a set of poses from which to observe the unexplored portions of the environment. Methods of coordinated exploration are inherently myopic as the ability to predict future actions is limited by knowledge of the environment. To address this limitation, we present a technique for inferring the unobserved portions of the environment and allowing the robots to plan their actions beyond what has been directly observed. Then, individual agents select poses that maximize the expected information gain while accounting for information their team members are likely to collect. Agents resolve conflicting claims using an auction of expected travel costs and discounting the conflicted reward from the losing robot. The developed method was evaluated in a series of simulated trials where it was able to reduce the time required to explore environments by 13.15%.
To enable coordination of distributed robotic teams in generalized robot scenarios we also provide a novel algorithm that allows for non-myopic coordination in real-time of heterogeneous robotic teams. Our algorithm allows each agent to plan non-myopically over their own potential actions while accounting for the impact of those actions on the team's cumulative reward. To determine the impact of the team reward individual robots, discount the expected rewards from their planned actions to account for the probability and time a team member may complete those actions. This approach allows a distributed team of heterogeneous robots to non-myopically coordinate in real-time to complete a wide range of tasks. The developed method was evaluated in a series of simulated and fielded hardware trials where we found that it was able to increase the cumulative team reward by a maximum of 47.2% in the trials conducted.