RoVerFly

RoVerFly is a MuJoCo and Stable-Baselines3 framework for reinforcement-learning control of quadrotor–payload systems. One policy handles taut/slack cable transitions, payload variation, actuator delay, and external disturbances without explicit mode switching.

Mintae Kim, Jiaze Cai, and Koushil Sreenath, RoVerFly: Robust and Versatile Implicit Hybrid Control of Quadrotor–Payload Systems. arXiv:2509.11149

Install

RoVerFly supports Python 3.10 and 3.11.

git clone https://github.com/mintaeshkim/roverfly.git
cd roverfly
python3.11 -m venv .venv
source .venv/bin/activate
pip install -e .

Optional plotting, export, and development dependencies are available as extras:

pip install -e '.[plot,export,dev]'

Train

roverfly train --env payload --num-envs 32 --device cpu --id exp_1

Available training environments are falcon, mini, payload, and random. Each experiment is self-contained under runs/<id>/, including TensorBoard logs, evaluation results, checkpoints, and the final model. The three-vehicle multi environment remains available through the Python API for simulation experiments, but is not offered by the training CLI because its cooperative reward is not yet implemented.

Evaluate and export

roverfly evaluate runs/exp_1/checkpoints/best_model.zip --env payload --episodes 20
roverfly export runs/exp_1/checkpoints/best_model.zip --format onnx

MNN export additionally requires MNNConvert on PATH. ONNX export requires the export extra.

Layout

roverfly/
├── assets/        MuJoCo models and textures
├── control/       action filters
├── envs/          Gymnasium environments
├── export/        deployment formats
├── integrations/  ROS integration
├── math/          geometry and rotations
├── simulation/    MuJoCo runtime and rendering
├── training/      PPO configuration and runner
└── trajectories/  reference trajectories

Development

pytest
ruff check .

MuJoCo smoke tests are marked integration and can be run explicitly with pytest -m integration on a machine with a working OpenGL runtime.

Citation

@article{kim2025roverfly,
  title={RoVerFly: Robust and Versatile Implicit Hybrid Control of Quadrotor-Payload Systems},
  author={Kim, Mintae and Cai, Jiaze and Sreenath, Koushil},
  journal={arXiv preprint arXiv:2509.11149},
  year={2025}
}