Multi-fidelity black-box optimization for time-optimal quadrotor maneuvers

Abstract

In order to generate time-optimal trajectories for highly maneuverable vehicles, such as quadrotor aircraft, a dynamics model that considers complex aerodynamic and electromechanical phenomena is required. Although generating time-optimal trajectories has been widely studied in the literature for several robotic vehicles, most existing research assumes a conservative dynamics model to simplify the optimization formulation. The resulting trajectories of these methods are often not time-optimal in reality. In this thesis, we propose a trajectory optimization algorithm that iteratively refines the dynamics model to achieve time-optimality in real-world environments. We consider the problem of minimum-time dynamically feasible quadrotor trajectory generation that connects a set of prescribed waypoints.

The boundary of the set of dynamically feasible trajectories is hard to model as it involves limitations of the entire system, including hardware and software, in agile high-speed flight. In this extreme condition, every part of a trajectory interacts with other parts, so the feasibility boundary has to be modeled for the entire trajectory. Moreover, the number of evaluations for modeling has to be minimized, because the experiments near the feasibility boundary are risky and expensive. In this work, we design a multi-fidelity Bayesian optimization framework to generate a time-optimal trajectory while approximating the feasibility boundary. We formulate the modeling of the feasibility boundary as a classification problem by utilizing the existing work that generates quadrotor maneuvers with smooth polynomials.

Also, we use a multi-fidelity optimization technique to combine the information from low fidelity sources, such as analytical approximation, numerical simulation, and keep the number of costly real-world flight experiments to a minimum. The algorithm is thoroughly evaluated in both simulation and real-world flight experiments at speeds up to 11 m/s in a 22 m x 11 m x 5.5 m room. The results demonstrate significant improvements in flight time up to 22 percent compared to the existing quadrotor trajectory optimization method.