Coordination and Control of Robotic Multi-agent Systems in Confined Environments

In nature, ants are well known for collectively carrying loads, birds fly in large flocks, and fish gather in schools. Such collective behavior is not only relevant in nature, but also for an increasing number of robotic systems. Multi-agent Systems (MAS) can provide benefits for a wide range of application areas, as groups of robots can collectively perform tasks that are way beyond the capabilities of individual agents. In this thesis, we introduce methods to coordinate and control robotic agents in both antagonistic and collaborative MAS in confined environments. We exemplarily work on two different scenarios, which are autonomous driving respectively racing, and formation control for a flock of quadcopters. For each of these scenarios, we use a three-pronged approach consisting of theoretical analysis and reasoning, implementation and simulation, and hardware experiments. For the autonomous racing task, we use machine learning to drive an agent in a timed racing scenario. We show that in simulation, the model-based deep Reinforcement Learning (RL) algorithm outperforms a number of other model-free RL algorithms, and we empirically demonstrate that this control method is able to successfully transfer the learned policy from simulation to a real-world test environment. In our collaborative MAS consisting of a group of drones, we control quadcopters to form and maintain a flock formation. We introduce Spatial Predictive Control (SPC) as a fully distributed control method that is based only on the position of the individual drone itself and on those of neighboring drones. Hardware experiments demonstrate SPC’s robustness against a potential sim-to-real transfer gap and its capability to perform properly in the presence of significant sensor noise and the extra latency of positional and control signals. For scenarios, where even positional observations are not possible, we present Distributed Distance-based Control (DDC), which is fully distributed and solely based on scalar distance measurements and local position estimation. To the best of our knowledge, we are the first to demonstrate such a controller on aerial MAS and perform experiments with real hardware drones.