Robocentric Model-Based Visual Servoing for Quadrotor Flights

Abstract

This article proposes a robocentric formulation for quadrotor visual servoing. This formulation presents the task-specific state dynamics of the quadrotor in its body reference frame. Compared with other visual servoing methods, our method allows tightly and integrated state estimation and control on the same robocentric model, and allows a faster system response in aggressive quadrotor flights. On the theory level, we prove the controllability and observability of the proposed robocentric model. Then, we design an on-manifold Kalman filter for the estimation and an on-manifold iterative model predictive control for motion planning and control. We verify our proposed formulation and controller in two crucial quadrotor flight tasks: 1) hovering; and 2) dynamic obstacle avoidance. Experiment results show that the quadrotor is able to resist large external disturbances and recover its position and orientation from two reference visual features. Moreover, the quadrotor is able to avoid dynamic obstacles reliably at a relative speed up to 7.4 m/s, demonstrating the effectiveness of our visual servoing methods in agile quadrotor flights.