Robots operating in the same space may not have the capability to coordinate explicitly with one another beyond local sense-and-avoid procedures. Consider the UAV package delivery domain where multiple heterogeneous UAVs, each equipped with in-house cost-based planning algorithms, must share the same cluttered urban airspace. Severe airspace congestion and potentially dangerous conflict incidents can result if there is no coordination between the UAVs. Furthermore, explicit inter-UAV cooperation in such a large and dynamic scenario can be computationally intensive if not impossible without prior arrangement. We address this problem by introducing a decentralized and distributed system of high-level learning agents that can manage traffic flow through assigned sectors of the airspace without needing to know the internal planners of each robot. The sector agents learn to dynamically adapt the cost of travel through their respective sectors to mitigate and alleviate congestion throughout the entire airspace. The policies learned by our algorithm have demonstrated a reduction in the total number of conflict incidents experienced in the airspace while maintaining throughput performance.

In this talk, I will briefly present the problems of maintenance in multi-robot systems. Maintenance problems arise when a group of robots, able to interact through an ad-hoc communication network or a relative sensing equipment, are tasked to perform a mission while maintaining some collective property of their interconnection. I will first quickly introduce the notion of graph and frameworks, the related matrices, the connectivity and rigidity properties. I will then show how these properties can be represented by a simple algebraic concepts such as the algebraic connectivity and the rigidity eigenvalue. Secondly I will introduce the maintenance problems and present gradient-based algorithm to solve them. I will then show the methodologies that allow those algorithms to be implemented on distributed multi-robot architectures. Thirdly I will show how additional requirements can be easily embedded in maintenance problems by the definition of suitable adjacency relations. All the methodologies will be presented with a critical approach, showing pros and cons of each choice. Finally I will conclude the talk providing some practical examples illustrating the presented methodologies.

The challenge in multi-robot coordination design is the mapping from implementation details (described by Measures of Performance) to specifications while reasoning about how to achieve the operational goal (described by Measures of Effectiveness). In this talk, this challenge is formulated at a meta-level above the actual implementation level as a non-cooperative game (called ‘Reasoning’) with one player (called ‘Details’) and the other player (called ‘Specifications’). These two players have to come as fast as possible to a design decision with guaranteed resulting effectiveness for the actual system implementation. This means that by solving this game a design solution is found which is applicable to an efficient independent Verification and Validation evaluation process. Since hidden information makes this evaluation process complicated and independence between ‘Details’ and ‘Specifications’ leads to a minimization of relevant hidden information, the key idea for a methodology to solve the game is to add as a constraint to the game formulation the necessity of following an ‘Independence Plan’ for the meta-level players. An iterative optimization algorithm is outlined which starts with an initial ‘Independence Plan’ developed by inspecting the structure of controller, estimator, utility and final reward equations for chances to implement opportunities for meta-level independence. This initial ‘Independence Plan’ is then mapped upon a physically viable coordination scheme. The following optimization steps then search for more efficient solutions along the directions of the parametric description of controller, estimator and utility while rejecting optimization steps without an associated new ‘Independence Plan.’

We present information-theoretic control policies that enable teams of robots to efficiently build 3D maps. Although these policies are intractable in general, we develop a series of approximations that make them suitable for real time use. We apply these policies to autonomously construct 3D maps with ground and aerial robots. By using Cauchy-Schwarz Quadratic Mutual Information, we show substantial computational improvements over similar information-theoretic measures. To map environments faster, we adopt a hierarchical planning approach which incorporates trajectory optimization so that robots can quickly determine feasible and locally optimal trajectories. Finally, we also present a high-level planning algorithm that enables heterogeneous robots to cooperatively construct maps.