Parameter-free Regression-based Autonomous Control of Off-the-shelf Quadrotor UAVs

Autonomous flight in unmanned aerial vehicles (UAVs) generally requires platform-specific knowledge of the dynamical parameters and control architecture. Recently, UAVs have become more accessible with off-the-shelf options that are well-tuned and stable for user teleoperation but due to unknown model parameters, they are typically not ready for autonomous operations. In this paper, we develop a method to enable autonomous flight on vehicles that are designed for teleoperation with minimal knowledge of the dynamical and controller parameters. The proposed method uses a basic knowledge of the control and dynamic architecture along with human teleoperated trajectories as demonstrations to train a thin-plate spline (TPS) regression model, which is then used to manipulate the pre-trained commands to generate new autonomous input commands for autonomous navigation over new trajectories. A statistical approach is also presented together with a satisfiability modulo theories (SMT) solver to assess the learned prediction error and correct to minimize errors in the input generation. A robust control-based strategy is also proposed to adjust autonomous input commands during run-time for closed loop trajectory tracking. Finally, we validate the proposed approach with trajectory-following experiments on a quadrotor UAV.