Skip to content

Training scripts, training data, and experimental data for Neural Fly

Notifications You must be signed in to change notification settings


Folders and files

Last commit message
Last commit date

Latest commit



4 Commits

Repository files navigation

Neural-Fly enables rapid learning for agile flight in strong winds

Michael O'Connell1, Guanya Shi1, Xichen Shi, Kamyar Azizzadenesheli, Anima Anandkumar, Yisong Yue, and Soon-Jo Chung2

This data and code is provided as part of the Science Robotics research article "Neural-Fly enables rapid learning for agile flight in strong winds", published on May 4th, 2022 here.

Training and validation script

Please run training-and-validation.ipynb, which demonstrates the Domain Adversarially Invariant Meta Learning (DAIML) algorithm. DAIML is the offline learning process for Neural-Fly. This script trains a wind-invariant representation of the aerodynamic effects on a quadrotor. After training the model, some simple statistics and plots are generated which show the model performance fitting to the training and testing data.

Filenaming scheme

Filenames are structured as


For specific details please refer to the article.

Field Description
  • custom: drone built using consumer off-the-shelf components with PX4 flight controller
  • intel: Intel-Aero RTF Drone, used for data collection for Neural-Fly-Transfer controller
  • random3: randomized trajectory created by randomly sampling 2 waypoints and generating a smooth spline from the current position through both waypoints; continuous derivatives through snap
  • random2: similar to random3 except only one random waypoint is generated
  • figure8: a lemniscate trajectory given by
      (x(t),y(t),z(t)) = (1.25 * sin(t), 0, 0.75 * sin(2 * t)
  • <METHOD>
  • 'baseline': nonlinear baseline control
  • 'indi': incremental nonlinear dynamics inversion control
  • 'L1': L1 adaptive control
  • 'NF-C': Neural-Fly-Constant, our adaptive controller without any learning
  • 'NF-T': Neural-Fly-Transfer, our learning based adaptive control with the ML model trained on data from the Intel-Aero drone
  • 'NF': Neural-Fly, our learning based adaptive control with the ML model trained on data collected with the custom drone ,
  • <CONDITION> wind condition for experiments, where the number corresponds to the fan array duty cycle and converts to
    • nowind = 0 m/s
    • 10wind = 1.3 m/s
    • 20wind = 2.5m/s
    • 30wind = 3.7 m/s
    • 35wind = 4.2 m/s
    • 40wind = 4.9 m/s
    • 50wind = 6.1 m/s
    • 70wind = 8.5 m/s
    • 70p20sint = 8.5+2.4sin(t) m/s
    • 100wind = 12.1 m/s

    Experiment data

    Additionally, the data from the experiment results present in the paper is provided. To load the data, run the following in python

    import utils
    Data = utils.load_data(folder='data/experiment')

    This will load all of the experiment data as a list of dictionaries. The ith experiment, field field, at the jth timestep, can be accessed with Data[i][field][j]. Most fields are ndarrays except the metadata fields, pulled from the filename. Available fields are given in the following table.

    field description
    't' time in seconds
    'p' position vector in meters
    'p_d' desired position vector in meters
    'v' velocity vector in m/s
    'v_d' desired velocity in m/s
    'q' attitude represented as a unit quaternion
    'R' rotation matrix (body frame to world frame)
    'w' (this is a misnomer - should be $\omega$) angular velocity in rad/s
    'T_sp' thrust setpoint sent to the flight controller
    'q_sp' attitude command sent to the flight controller
    'hover_throttle' throttle at hover computed as a best fit function of the battery voltage
    'fa' aerodynamic residual force computed using numerical differentiation of v and T_sp, q, and hover_throttle
    'pwm' motor speed commands from the flight controller
    'vehicle' <VEHICLE> field from filename
    'trajectory' <TRAJECTORY> field from filename
    'method' <METHOD> field from filename
    'condition' <CONDITION> field from filename


    Check out our overview video:

    Full overview video

    60 second overview video created by Caltech's Office of Strategic Communications:

    60 second overview video


    The data and code here are for personal and educational use only and provided without warranty; written permission from the authors is required for further use. Please cite our work as follows.

    @article{ doi:10.1126/scirobotics.abm6597, author = {Michael O’Connell and Guanya Shi and Xichen Shi and Kamyar Azizzadenesheli and Anima Anandkumar and Yisong Yue and Soon-Jo Chung }, title = {Neural-Fly enables rapid learning for agile flight in strong winds}, journal = {Science Robotics}, volume = {7}, number = {66}, pages = {eabm6597}, year = {2022}, doi = {10.1126/scirobotics.abm6597}, URL = {}}


    1. equal contribution and alphabetical order 2

    2. corresponding author -


    Training scripts, training data, and experimental data for Neural Fly






    No releases published


    No packages published