Nano-Drone

This benchmark is based on real-world flight data from a Crazyflie 2.1 Brushless nano-quadrotor (~50 g), recorded in a motion-capture arena across four aggressive trajectories. The available dataset can be downloaded here, you can also find a detailed description of the benchmark setup and some baseline results through this link.

A real-flight nano-drone dataset (~75k samples) with synchronized motor inputs and full-state outputs are provided. The goal is to model the Nano-Drone dynamics from the motor angular velocities to the drone position, velocity, orientation and angular velocity. Furthermore, reference identification models, are available in the accompanianing repository and paper including: physics-based models, purely data-driven neural models, hybrid models, and recurrent (LSTM-based) models. This results in a standardized multi-step prediction benchmark, evaluating open-loop error propagation up to 0.5 s (50 steps at 100 Hz). 

Previously published results on the Nano-Drone benchmark can be found below. You can submit your own results through this form. Note that the reported results are curated, only complete submissions with meaningful contributions will be included. Candidate entries should make use of the Python dataloader functionalities and figure of merit calculation functions provided through this link.