GymFC is flight control tuning framework with a focus in attitude control. GymFC was first introduced in the manuscript "Reinforcement learning for UAV attitude control" in which a simulator was used to synthesize neuro-flight attitude controllers that exceeded the performance of a traditional PID controller. Since the projects initial release it has matured to become a modular framework for tuning flight control systems, not only for synthesizing neuro-flight controllers but also tuning traditional controllers as well. GymFC is the primary method for developing controllers to be used in the worlds first neural network supported flight control firmware Neuroflight. Details of the project and its architecture are best described in Wil Koch's thesis "Flight Controller Synthesis Via Deep Reinforcement Learning".

Please use the following BibTex entries to cite our work,

@article{koch2019reinforcement,
  title={Reinforcement learning for UAV attitude control},
  author={Koch, William and Mancuso, Renato and West, Richard and Bestavros, Azer},
  journal={ACM Transactions on Cyber-Physical Systems},
  volume={3},
  number={2},
  pages={22},
  year={2019},
  publisher={ACM}
}

@article{koch2019flight,
  title={Flight Controller Synthesis Via Deep Reinforcement Learning},
  author={Koch, William},
  journal={arXiv preprint arXiv:1909.06493},
  year={2019}
}

Table of contents

• Features
• News
• Installation
• Getting Started
• Tutorials
• User Modules
• Custom Modules
• Examples
• Development Team
• Contributions

Features

• Support for IMU, ESC and battery sensors
• Aircraft agnostic - support for any type of aircraft just configure number of actuators and sensors.
• Digital twin independence - digital twin is developed external to GymFC allowing separate versioning.
• Google protobuf aircraft digital twin API for publishing control signals and subscribing to sensor data.
• Flexible agent interface allowing controller development for any type of flight control systems.
• Support for Gazebo 8, 9, and 11. Gazebo plugins are built dynamically depending on your installed version.

News

• May 2020 - NF1 quadcopter model and reward functions used in this thesis work is published in the examples directory.
• August 2019 - GymFC synthesizes neuro-controller with new level of performance.
• August 2019 - Thesis is defended, Flight Controller Synthesis via Deep Reinforcement Learning.
• July 2019 - GymFC v0.2.0 is released.
• December 2018 - Our GymFC manuscript is accepted to the journal ACM Transactions on Cyber-Physical Systems.
• November 2018 - Flight controller synthesized with GymFC achieves stable flight in Neuroflight.
• September 2018 - GymFC v0.1.0 is released.
• April 2018 - Pre-print of our paper is published to arXiv.

Installation

Quick start

To install GymFC and its dependencies on Ubuntu 18.04 execute,

sudo MAKE_FLAGS=-j4 ./install_dependencies.sh
pip3 install .

Dependencies

GymFC runs on Ubuntu 18.04 and uses Gazebo v10.1.0 with Dart v6.7.0 for the backend simulator. To use Dart with Gazebo, they must be installed from source. For why Gazebo must be used with Dart see this video. The easiest way to install the dependencies is with the provided install_dependencies.sh script. By default it will run make with a single job. You can override the make flags with the MAKE_FLAGS environment variable. Building Gazebo from source is very resource intensive. If you have sufficient memory increase the number of jobs to run in parallel. For example to run four jobs in parallel execute,

sudo MAKE_FLAGS=-j4 ./install_dependencies.sh

Note, this script may take more than an hour to execute. If your build fails check dmesg but the most common reason will be out-of-memory failures.

GymFC

(Optional) It is suggested to set up a virtual environment to install GymFC into. From the project root run, python3 -m venv env. This will create an environment named env which will be ignored by git. To enable the virtual environment, source env/bin/activate and to deactivate, deactivate.

Install GymFC,

pip3 install .

This will install the Python dependencies and also build the Gazebo plugins and messages.

Developing with GymFC

If you plan to modify the GymFC code you will need to install in edit/development mode.

pip3 install -e .

You will also have to manually install the Gazebo plugins by executing,

gymfc/envs/assets/gazebo/plugins/build_plugin.sh

If you deviate from this installation instructions (e.g., installing Gazebo in a different location other than specific in install_dependencies.sh), you may need to change the location of the Gazebo setup.sh defined by the variable SetupFile in gymfc/gymfc.ini.

Verifying install

GymFC requires an aircraft model (digital twin) to run. The NF1 racing quadcopter model is available in examples/gymfc_nf/twins/nf1 if you need a model for testing. To test everything is installed correctly run,

python3 tests/test_start_sim.py --verbose examples/gymfc_nf/twins/nf1/model.sdf

If everything is OK you should see the NF1 quadcopter model in Gazebo.

You will see the following error message because you have not built the motor and IMU plugins yet.

[Err] [Plugin.hh:187] Failed to load plugin libgazebo_motor_model.so: libgazebo_motor_model.so: cannot open shared object file: No such file or directory
[Err] [Plugin.hh:187] Failed to load plugin libgazebo_imu_plugin.so: libgazebo_imu_plugin.so: cannot open shared object file: No such file or directory

Also the following error message is normal,

[Err] [DARTJoint.cc:195] DARTJoint: SetAnchor is not implemented

To use the NF1 model for further testing read examples/README.md.

Install by Docker

This repository includes an experimental docker build in docker/demo that demos the usage of GymFC. It has been tested on MacOS 10.14.3 and Ubuntu 18.04, however the Gazebo client has not been verified to work for Ubuntu. This docker image can help ensure you are running a supported environment for GymFC.

Install dependencies

For Mac, install Docker for Mac and XQuartz on your system. For Ubuntu, install Docker for Ubuntu.

Build and test

Build the docker image

docker build  . -t gymfc:demo

This will take a while as it compiles mesa drivers, gazebo and dart. It is recommended to give Docker a large part of the host's resources. All incoming connections will forward to xquartz:

xhost +

Example usage, run the image and test test_step_sim.py using the Solo digital twin,

docker run -ti -e DISPLAY=<hostip>:0 \
-v <path-to-gymfc-digitaltwin-solo>/models/solo/model.sdf:/gymfc/demo/models/solo/model.sdf \
gymfc:demo \
"python3 /gymfc/tests/test_step_sim.py --gymfc-config /gymfc/gymfc.ini --verbose /gymfc/demo/models/solo/model.sdf  1 1 1 1"

Replace <hostip> by the external ip of your system to allow gymfc to connect to your XQuartz server and <path-to-gymfc-digitaltwin-solo> to where you cloned the Solo repo. Take special note that the test_step_sim.py parameters are using the containers path, not the host's path.

Getting Started

The simplest environment can be created with,

from gymfc.envs.fc_env import FlightControlEnv
class MyEnv(FlightControlEnv):
    def __init__(self, aircraft_config, config=None, verbose=False):
        super().__init__(aircraft_config, config_filepath=config, verbose=verbose)

By inheriting FlightControlEnv you now have access to the step_sim and reset functions. If you want to create an OpenAI gym you also need to inherit this class e.g.,

from gymfc.envs.fc_env import FlightControlEnv
import gym
class MyOpenAIEnv(FlightControlEnv, gym.Env):  

For simplicity the GymFC environment takes as input a single aircraft_config which is the file location of your aircraft model model.sdf. The SDF declares all the visualizations, geometries and plugins for the aircraft.

Directory Layout

GymFC expects your model to have the following Gazebo style directory structure:

model_name/
  model.config
  model.sdf
  plugins/
    build/

where the plugin directory contains the source for your plugins and the build directory wi