RotorBench

A lightweight Python-based quadrotor flight simulator for benchmarking, rapid prototyping, and academic research.

Features

1. Lightweight quadrotor simulator
   A pure Python implementation designed for rapid prototyping and low-level control algorithm development.
2. Physics engine
   • Accurate SE(3) rigid-body dynamics update with RK4 integration and Cayley transform.
   • First-order delayed ESC (Electronic Speed Controller) model.
3. GPU-accelerated batched rollouts
   Built on PyTorch for massively parallel simulation.
4. Rich controller suite out-of-the-box
   • Geometric Tracking Controller
   • Linear Quadratic Regulator (LQR)
   • H∞ Robust Controller
   • Nonlinear Model Predictive Control (NMPC)
   • Reinforcement Learning (PPO via Stable-Baselines3)
5. Trajectory planning and visualization
   • YAML-configurable reference trajectories
   • Real-time animated 3D visualization of quadrotor motion

Installation

pip install -r requirements.txt

Run Simulations

Geometric Tracking Controller

python run.py --ctrl=GEOMETRIC_CTRL

Linear Quadratic Regulator (LQR)

python run.py --ctrl=LQR

H∞ Controller

python run.py --ctrl=HINFTY_CTRL

Nonlinear Model Predictive Control (NMPC)

python run.py --ctrl=NMPC

Reinforcement Learning (Experimental)

Download a pre-trained policy:

mkdir -p runs/ppo_quadrotor/best
wget -O runs/ppo_quadrotor/best/best_model.zip https://raw.githubusercontent.com/shengwen-tw/rotor-bench/blob/runs/ppo_quadrotor/best/best_model.zip

Start simulation with the RL controller:

python run.py --ctrl=RL

To train an RL policy:

python train_rl.py --traj HOVERING --iterations 1000 --n-envs 64 --total-steps 1000000000000
tensorboard --logdir runs/ppo_quadrotor

Benchmarking (Experimental)

Run multiple independent simulations in parallel for benchmarking. It currently serves as stress testing.

python benchmark.py [--ctrl CONTROLLER] [--workers NUMBER]

Common options:

• --ctrl {GEOMETRIC_CTRL,NMPC,LQR,HINFTY_CTRL,RL}: controller to benchmark
• --workers N: number of worker processes to launch
• --traj {HOVERING,CIRCLE,EIGHT}: reference trajectory
• --iterations N: simulation steps per run
• --dt DT: simulation timestep

Example:

python benchmark.py --ctrl NMPC --workers 16

The benchmark computes statistics such as wall time, completed steps, and position/velocity tracking error.

Project Structure

rotor-bench/
├── benchmark.py             # Parallel benchmarking script
├── run.py                   # Entry point of the simulation
├── train_rl.py              # RL training script
├── trajectory_planner.py    # Trajectory planner
├── models/
│   ├── dynamics.py          # Rigid-body dynamics
│   ├── esc.py               # ESC model
│   ├── quadrotor.py         # Gym-compatible quadrotor environment
│   ├── se3_math.py          # SE(3) math utilities
│   └── thrust_allocator.py  # Control allocation
├── control/
│   ├── care_sda.py          # CARE solver via SDA
│   ├── geometric_control.py # Geometric controller
│   ├── hinf_syn.py          # H∞ control synthesizer
│   ├── hinfty_control.py    # H∞ controller
│   ├── lqr_control.py       # LQR controller
│   ├── nmpc.py              # Nonlinear MPC
│   └── rl_control.py        # RL controller
├── configs/                 # Vehicle and trajectory configs
├── viz/                     # 3D visualization
├── assets/                  # Images and media
└── requirements.txt

Preview

Citation

If you find RotorBench useful for your research, please consider citing:

@misc{rotorbench2026,
  author = {Cheng, Sheng-Wen},
  title  = {RotorBench: A Lightweight Python Quadrotor Simulator for Control Benchmarking},
  year   = {2026},
  url    = {https://github.com/shengwen-tw/rotor-bench}
}

An extended version of this work is currently under preparation for submission to arXiv and a peer-reviewed journal.

For the H∞ control, please cite:

@inproceedings{cheng2022robust,
  title={Robust State-Feedback H∞ Control of Quadrotor},
  author={Cheng, Sheng-Wen and Hung, Hsin-Ai},
  booktitle={2022 International Automatic Control Conference (CACS)},
  pages={1--7},
  year={2022},
  organization={IEEE}
}

License

This project is licensed under the MIT License.
Feel free to use, modify, and distribute without restriction.