Accurate rotational dynamics modeling for the Crazyflie 2.1 nano-quadrotor

Develop an accurate predictive model of the rotational dynamics (attitude and angular velocity evolution) of the Crazyflie 2.1 Brushless nano-quadrotor that captures low-frequency actuation dynamics beyond the standard quadratic thrust/torque mapping between propeller angular speeds and body torques, so as to improve multi-step open-loop orientation and angular-rate prediction under the benchmark’s evaluation protocol.

Background

The paper introduces a benchmark and baseline models for nonlinear system identification on a Crazyflie 2.1 Brushless nano-quadrotor, evaluating multi-step predictions of position, velocity, orientation, and angular velocity. While translational dynamics are predicted with high fidelity, the authors find persistent difficulty in modeling rotational dynamics, particularly due to limitations of the standard quadratic motor model that maps propeller speeds to thrust and torques.

Empirical analyses show that the quadratic model fails to capture slow, low-frequency components of the actuation influencing torques, leading to systematic errors in rotational predictions. The authors adopt a simplified physical baseline (setting angular acceleration to zero) and demonstrate that residual and LSTM-based corrections remain insufficient over the tested 0.5 s windows, highlighting the need for more expressive temporal models or enhanced physics-based rotor/motor descriptions.