A Learning-based Quadcopter Controller with Extreme Adaptation

This repo contains the code associated with the paper A Learning-based Quadcopter Controller with Extreme Adaptation

Paper and Video

If you use this code in an academic context, please cite the following publication:

Paper: A Learning-based Quadcopter Controller with Extreme Adaptation

Video: YouTube

@article{zhang2024learningbasedquadcoptercontrollerextreme,
   title={A Learning-based Quadcopter Controller with Extreme Adaptation},
   ISSN={1941-0468},
   url={http://dx.doi.org/10.1109/TRO.2025.3577037},
   DOI={10.1109/tro.2025.3577037},
   journal={IEEE Transactions on Robotics},
   publisher={Institute of Electrical and Electronics Engineers (IEEE)},
   author={Zhang, Dingqi and Loquercio, Antonio and Tang, Jerry and Wang, Ting-Hao and Malik, Jitendra and Mueller, Mark W.},
   year={2025},
   pages={1–17}
}

Setup

To set up the codebase, follow these steps:

1. Make sure you are in the xadapt_ctrl folder
2. Run the following commands:

# Initialize git submodules
git submodule update --init --recursive

# Create and activate conda environment
conda env create -f environment.yml
conda activate xadap

# Install py3dmath package
cd py3dmath
rm -rf py3dmath.egg-info
pip install -e .
cd ..

# Run the simulation
python simulate.py

If successful, you should see the simulation of trajectory tracking with our controller.

⭐ Learning the Adaptive Controller ⭐

The learning directory contains the core implementation of our adaptive controller training code (paper section II.B-D):

• ppo.py: Unified RL (PPO) + IL curriculum learning implementation for Phase 1 training (despite its name, it's more than just PPO 😉)
• dagger.py: DAGGER algorithm implementation for Phase 2 training
• train.py: Unified training script for the two-phase process
• hyperparam.yaml: Configuration file for training hyperparameters

To use this code, you'll need to:

1. Set up your own simulation environment
2. Integrate these training scripts with your infrastructure
3. Adjust hyperparameters as needed for your specific setup

Note: This code is provided as a reference implementation. The scripts are not directly runnable and need integration with your own training infrastructure.

Docker Setup

If your operating system doesn't support environment of this project, docker is a great alternative.

First of all, you have to build the project and create an image like so:

## Assuimg you are in the correct project directory
docker build -t xadaptctrl .

To use a shortcut, you may use the following command:

## Assuimg you are in the correct project directory
make docker_build

After the image is created, copy and paste the following command to the terminal to run the image:

## Assuimg you are in the correct project directory
xhost +
docker run -it --network host   --env="DISPLAY" --env="QT_X11_NO_MITSHM=1" --device=/dev/video0:/dev/video0 --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" --entrypoint /bin/bash xadaptctrl

NOTE: You only have to run xhost + once each time you log into the machine. These settings persist per login session. To use a shortcut, you may use following command:

make docker_run

Running the package at docker image

If you are in the docker image , this project is already sourced and the project can be run as the following command;

python3 simulate.py