This paper presents a novel application of recursive fast orthogonal search (R-FOS) to produce a time-varying linear state-space system model based on a historical record of input and outputs. The model is integrated into a larger unmanned aerial system to control the altitude of a quadcopter using adaptive model predictive control (MPC). Similar to a dual Kalman filter (KF), R-FOS is also used in conjunction with a KF observer in the form of a dual fast orthogonal Kalman filter to estimate the system states. Four variations of R-FOS are implemented and compared in terms of mean squared error during a number of simulations. The R-FOS adaptive MPC strategy provides better altitude control when executed online than when an offline model is used. Comparable performance is achieved with fewer model terms when the FOS criteria for evaluation of additional terms is considered. Finally, the performance is significantly improved when a forgetting factor is incorporated to give greater weight to more recent observations. These findings demonstrate that by combining R-FOS with adaptive MPC, one can optimally control a system with unknown, time-varying dynamics.
This paper proposes a system for guidance of autonomous vehicles based on a generalized robust model-predictive control (R-MPC) technique. It offers two scenarios, in which R-MPC is used to provide optimized motion plans guaranteed to satisfy constraints and avoid obstacles in the presence of bounded uncertainty. The first scenario involves a military unmanned aerial vehicle flying over a target, while avoiding enemy defenses. The second scenario involves an assistive care robot safely navigating to waypoints throughout a cluttered home. In each case, obstacles are represented as a collection of linear inequality constraints that adapt to changes in the environment. Results show that when R-MPC is formulated using a series of positive-semidefinite relaxations on linear inequality constraints, safe optimum trajectories can be planned by solving a convex quadratic program. When formulated in terms of perturbations on feedback predictions, this solution is guaranteed to be robustly stable.
Inspired by the natural mobbing behavior of birds, this work presents a novel, quasi-distributed swarming strategy called the Dynamic Lemniscatic Arch. It resolves the problem of producing globally-stable, evenly-spaced lemniscate (or, figure-eight) trajectories while relying on local interactions only. Such trajectories are advantageous in applications where energy consumption and mechanical strain must be minimized. Previous work in lemniscate curves has typically relied on predetermined trajectories, rather than on the emergent structure of the swarm. Furthermore, we enrich the traditional 2-dimensional lemniscate plane curve structure by forming an arch in the third dimension. This arch provides more consistent coverage in surveillance type tasks and, with minor variations in parameters, can be used to produce mobbing behavior. The technique relies on time-varying quaternion rotations linked to the positions of dynamically induced virtual agents. We provide a mathematical proof of stability, which demonstrates the swarm converges to the desired geometry. Simulations show that the strategy performs well with multiple agents and in numerous different configurations.
The potential benefits of big data and machine learning techniques are yet to be fully realized in real-time, safety-critical applications like unmanned aerial vehicle control. This is because of challenges related to interpretation, error susceptibility, and resources requirements. Due to their robustness and reliability, traditional model-based design techniques still dominate this landscape. However, a growing body of research in adaptive control has demonstrated the potential benefits of merging these two distinct design philosophies. This article investigates the benefits of using a combination of machine learning techniques to automatically tune parameters within a strictly defined model predictive control architecture. Fast orthogonal search and finite action-set learning automata are used to tune model coefficients and objective function weights, respectively. The strategy is validated experimentally on an actual Quanser Qball2 quadcopter and through several simulations of a Parrot AR.drone. Results demonstrate that the proposed approach improves performance while reducing design effort.