Learning to Fly: Computational Controller Design for Hybrid UAVs with Reinforcement Learning

Jie Xu, Tao Du, Michael Foshey, Beichen Li, Bo Zhu, Adriana Schulz, Wojciech Matusik

ACM Transactions on Graphics, 38(4) 42:1-42:12 (SIGGRAPH), 2019

This repository implements the code for the paper Learning to Fly: Computational Controller Design for Hybrid UAVs with Reinforcement Learning (SIGGRAPH 2019).

Code Structure

• SimulationUI: C++ rigid body simulation to visualize the controller's performance. (C++)
• Training: Code to train the controller with our method with Reinforcement Learning. (Python)
• Firmware: Modified Ardupilot firmware code to implement our controller on real hardware. (C++)

Simulation UI

• The code is implemented to visualize the controller's performance in a realistic simulation environment and provide the user interaction to control the hybrid UAV through keyboard.

• The code has been tested on Ubuntu 16.04 and Ubuntu 18.04.

• Compiling the code relies on cmake >= 3.1.0. Please install it through the instruction in this link before compiling the code.

• Compile the code:

  • Enter SimulationUI folder.

  • Create build folder and enter it by:

    mkdir build; cd build;
    

  • Run cmake and make to compile the code.

    cmake -DCMAKE_BUILD_TYPE=Release ../
    make -j4
    

  • If cmake fails, please install the missing libraries listed in the error message.

• Run the code:

  • The default code will run the simulation for QuadPlane controlled by a pre-trained neural network controller.

  • Run the following command inside the build folder:

    ./projects/design_viewer/design_viewer
    

  • It will open a window as below where you can control the hybrid UAV through keyboard commands.

    

• Keyboard Command:

  • We provide a bunch of keyboard command to allow you to easily control the hybrid UAV.
  • p: start/pause simulation, you need to press p after you create the window.
  • w/s: increase/decrease velocity z (vertical velocity, negative value is going up).
  • a/d: increase/decrease target yaw angle.
  • up/down: increase/decrease velocity x (forward velocity).
  • left/right: increase/decrease velocity y (side velocity).
  • You can use the mouse scroll wheel to zoom in/out (scroll) and change the view (press and move) of the camera.
  • For more keyboard operation, please refer to the code.
  • To be noted, there is a dead zone for the velocity z, thus the controller is not triggered until the velocity z is smaller than -0.5.

• Tail-sitter: we also provide the config file and the pretrained controller for tail-sitter. You can modify the code in DesignViewer.cpp to load its config file and replace the controller parameter file projects/copter_simulation/Controller/NN/NNModelParameters.h with its controller file (which you can find in projects/copter_simulation/Controller/NN/TailSitter/).

Training

• The code has been tested on Ubuntu 16.04.

• The code relies on a modified version of OpenAI baselines in [email protected]:eanswer/openai_baselines.git. To run the code, please first clone the repository

  git clone [email protected]:eanswer/openai_baselines.git
  

  and then follow the README.md to install baselines.

• We provide two models in the data folder.

• To start training a controller for a model, for example QuadPlane, in one thread,

  python run.py --model QuadPlane
  

  or in multiple threads (e.g. 4),

  mpirun -np 4 python run.py --model QuadPlane
  

  You can also add a --visualize to visualize the training by Mujoco. (If you don't install Mujoco, this option does't work).

• The trained models are saved in results folder.

• To test a trained mode, for example model_afterIter_400.ckpt, you can run

  python run.py --model QuadPlane --controller model_afterIter_400 --play
  

• To save a trained model to a c++ format to be plugged into the SimulationUI code or ardupilot firmware code.

  python save.py --controller model_afterIter_400
  

Firmware Deployment

• The firmware code is based on the Ardupilot codebase and tested in Pixhawk 1.0.

• First plug the model description file generated by training code (should named AP_MotorsPolicyDefinition.h) into the firmware codebase:

  cp AP_MotorsPolicyDefinition.h ../../Firmware/ardupilot/libraries/AP_Motors/
  

• Connect the Pixhawk 1.0 to your computer by usb cabel.

• Enter the folder ArduCopter by cd Firmware/ardupilot/ArduCopter.

• Compile and upload the firmware code by

  make px4-v2-upload
  

• Please refer to Ardupilot official document for more detials.

Citation

If you find our paper or code is useful, please consider citing:

@article{xu2019learning,
  title={Learning to fly: computational controller design for hybrid UAVs with reinforcement learning},
  author={Xu, Jie and Du, Tao and Foshey, Michael and Li, Beichen and Zhu, Bo and Schulz, Adriana and Matusik, Wojciech},
  journal={ACM Transactions on Graphics (TOG)},
  volume={38},
  number={4},
  pages={1--12},
  year={2019},
  publisher={ACM New York, NY, USA}
}