Learning-based Decentralized Control of Groups of Multiple Robots

Over the past years many advancements have been made in devising decentralized control and estimation schemes for groups of multiple robots. From the initial consensus-like laws for solving simple rendez-vous or formation control problems [1], we have witnessed a flourishing or more complex algorithms able to embed and enforce much more complex "behaviors" for the group, such as inter-agent and obstacle collision avoidance, possibility of changing the sensing/communication topology at runtime, maintenance of "global" group properties such as connectivity or rigidity, implementation of navigation or exploration strategies in unknown environments, and interfacement with a human operator in charge of giving high-level commands to the group (or part of the group itself) [2,3,4,5]. While several theoretical frameworks have been proposed for embedding an arbitrary number of "complex" behaviors  in a single (but "arbitrarily complex") decentalized control law, this complexification has some costs, one of which being the increasing computational and communication complexity of the control law as a function of the "complexity" of the sought behaviors and of the group size.

One possible (and promising) alternative to the implementation of analytical (but computationally complex) decentalized laws is the use of modern learning techniques in the following illustrative pipeline: first, an analytical (or model-based) decentralized is designed for implementing the desired group behaviors. Then, a suitable neural network is trained on the model-based law in order to reproduce its effects. If the training is successful, the neural network can then be used in place of the model-based law with a key advantage: the computation complexity of the network is in general very law and scalable, contrarily to the model-based law, and thus much more amenable to be used on computationally-constrained robots such as small UAVs with limited onboard hardware [6,7,8].

We have recently started investigating these ideas with promising results in reproducing the behavior of a previously proposed decentralized connectivity maintenance law with inter-agent and obstacle avoidance requirements. The trained network could successfully reproduce the model-based controller in many simulated tests. The goal of this Thesis is start from this initial work and expand it in several directions, for instance: how to train the network in a decentralized way, how to include sensory data (images from onboard camera) in the learning, how to include the presence of humans either in the team or commanding the team. The activities will be tested experimentally on the quadrotor UAVs available in the Rainbow team.