AcroRL: Learning Aggressive Quadrotor Inversion using Bidirectional Thrust

AcroRL is a simulation framework for learning aggressive quadrotor inversion maneuvers using bidirectional thrust and geometric control.

---

Key Components

1. Reference Modulation Policy $(\boldsymbol{\pi})$
   • Learned reference modulation policy for aggressive inversion maneuvers leveraging bidirectional thrust.
2. Hopf Fibration-Based Geometric Control (HFCA)
   • Open-source implementation of a Hopf-fibration-based geometric controller with custom inversion logic.
3. Thrust Model $(T(\Omega))$
   • Steady-state and stochastic transient thrust model for asymmetric propellers.
4. Optimal Control Allocation (OCA)
   • Projected gradient descent (PGD)-based optimal control allocation under actuator constraints.

---

Installation

Install mamba here.

Installation tested on Ubuntu 20.04 and 22.04.

If using a CUDA version other than 12, update the jax[cuda=XX] dependency in environment.yaml to match your local CUDA installation.

git clone git@github.com:rislab/acrorl.git
cd acrorl

mamba env create -f environment.yaml
mamba activate acrorl

pip install --use-pep517 -e .

---

Usage

For all scripts, run python3 scripts/{script} --help for a full list of options.

Training

python3 scripts/train.py --inversion_type nti --save_plots --save_data

Evaluate Reference Controller with Different Trajectories

Parallel nominal-to-inverted step command transition example usage:

python3 scripts/eval_control.py --inversion_method step_oca --inversion_type nti --save_plots --num_drones 20 --randomize_reset --show_hud

Evaluate Reference Controller with Learned Reference Modulation

Single nominal-to-inverted transition example usage:

python3 scripts/eval_learned_control.py --policy_name nti_final --inversion_type nti --save_plots

Export Learned Reference Modulation Policy to ONNX

Example export usage:

python3 export.py --policy-path /policies/my_policy --output ../policies/my_policy.onnx --verify

---

Citation

@misc{acrorl,
      title={AcroRL: Learning Aggressive Quadrotor Inversion using Bidirectional Thrust}, 
      author={Gabriel Rodriguez and Henri Sayag and Abhishek Rathod and John Stecklein and Siddharth Saha and Christopher Barngrover and Wennie Tabib},
      year={2026},
      eprint={2605.24301},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2605.24301}, 
}

If you use this repository in academic work, please also cite the original developers of the flightning simulator:

@inproceedings{flightning,
    title={Learning Quadrotor Control From Visual Features Using Differentiable Simulation},
    author={Johannes Heeg and Yunlong Song and Davide Scaramuzza},
    year={2025},
    booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
    url={https://arxiv.org/abs/2410.15979},
}

Forked from rpg_flightning.

Adapted animation module from pyplot3d.