Novel probabilistic and distributed algorithms for guidance, control, and nonlinear estimation of large-scale multi-agent systems

Multi-agent systems are widely used for constructing a desired formation shape, exploring an area, surveillance, coverage, and other cooperative tasks. This dissertation introduces novel algorithms in the three main areas of shape formation, distributed estimation, and attitude control of large-scale multi-agent systems. In the first part of this dissertation, we address the problem of shape formation for thousands to millions of agents. Here, we present two novel algorithms for guiding a large-scale swarm of robotic systems into a desired formation shape in a distributed and scalable manner. These probabilistic swarm guidance algorithms adopt an Eulerian framework, where the physical space is partitioned into bins and the swarm's density distribution over each bin is controlled using tunable Markov chains. In the first algorithm - Probabilistic Swarm Guidance using Inhomogeneous Markov Chains (PSG-IMC) - each agent determines its bin transition probabilities using a time-inhomogeneous Markov chain that is constructed in real-time using feedback from the current swarm distribution. This PSG-IMC algorithm minimizes the expected cost of the transitions required to achieve and maintain the desired formation shape, even when agents are added to or removed from the swarm. The algorithm scales well with a large number of agents and complex formation shapes, and can also be adapted for area exploration applications. In the second algorithm - Probabilistic Swarm Guidance using Optimal Transport (PSG-OT) - each agent determines its bin transition probabilities by solving an optimal transport problem, which is recast as a linear program. In the presence of perfect feedback of the current swarm distribution, this algorithm minimizes the given cost function, guarantees faster convergence, reduces the number of transitions for achieving the desired formation, and is robust to disturbances or damages to the formation. We demonstrate the effectiveness of these two proposed swarm guidance algorithms using results from numerical simulations and closed-loop hardware experiments on multiple quadrotors. In the second part of this dissertation, we present two novel discrete-time algorithms for distributed estimation, which track a single target using a network of heterogeneous sensing agents. The Distributed Bayesian Filtering (DBF) algorithm, the sensing agents combine their normalized likelihood functions using the logarithmic opinion pool and the discrete-time dynamic average consensus algorithm. Each agent's estimated likelihood function converges to an error ball centered on the joint likelihood function of the centralized multi-sensor Bayesian filtering algorithm. Using a new proof technique, the convergence, stability, and robustness properties of the DBF algorithm are rigorously characterized. The explicit bounds on the time step of the robust DBF algorithm are shown to depend on the time-scale of the target dynamics. Furthermore, the DBF algorithm for linear-Gaussian models can be cast into a modified form of the Kalman information filter. In the Bayesian Consensus Filtering (BCF) algorithm, the agents combine their estimated posterior pdfs multiple times within each time step using the logarithmic opinion pool scheme. Thus, each agent's consensual pdf minimizes the sum of Kullback-Leibler divergences with the local posterior pdfs. The performance and robust properties of these algorithms are validated using numerical simulations. In the third part of this dissertation, we present an attitude control strategy and a new nonlinear tracking controller for a spacecraft carrying a large object, such as an asteroid or a boulder. If the captured object is larger or comparable in size to the spacecraft and has significant modeling uncertainties, conventional nonlinear control laws that use exact feed-forward cancellation are not suitable because they exhibit a large resultant disturbance torque. The proposed nonlinear tracking control law guarantees global exponential convergence of tracking errors with finite-gain Lp stability in the presence of modeling uncertainties and disturbances, and reduces the resultant disturbance torque. Further, this control law permits the use of any attitude representation and its integral control formulation eliminates any constant disturbance. Under small uncertainties, the best strategy for stabilizing the combined system is to track a fuel-optimal reference trajectory using this nonlinear control law, because it consumes the least amount of fuel. In the presence of large uncertainties, the most effective strategy is to track the derivative plus proportional-derivative based reference trajectory, because it reduces the resultant disturbance torque. The effectiveness of the proposed attitude control law is demonstrated by using results of numerical simulation based on an Asteroid Redirect Mission concept. The new algorithms proposed in this dissertation will facilitate the development of versatile autonomous multi-agent systems that are capable of performing a variety of complex tasks in a robust and scalable manner