Neural Predictor for Flight Control with Payload

This repo contains the implementation of the paper "Neural Predictor for Flight Control with Payload" by Ao Jin, Chenhao Li, Qinyi Wang, Ya Liu, Panfeng, Huang and Fan Zhang*.

The Neural Predictor is a learning-based scheme to capture force/torque caused by payload and residual dynamics of tethered-UAV system. Specifically, inspired by DDKO theory, the formulation of lifted linear system (LLS) is derived and the LLS is learned from data to capture the force/torque caused by payload and residual dynamics of tethered-UAV system. The learned dynamics combined with nominal dynamics, which produces a hybrid model of tethered-UAV system. This hybrid model is incorporated into a model predictive control framework, known as NP-MPC. We demonstrate that our proposed framework not only provides much better prediction of external force and torque against state-of-the-art learning-based estimator, but also improve closed-loop performance in both simulations and real-world flight significantly.

Our paper: Neural Predictor for Flight Control with Payload

1. Getting Started

Before running the code, install the dependency packages in a virtual python env by executing the following command:

pip install -r requirements.txt

2. Running the Code

This repo includes the code of two parts: Numerical Evaluation and Physical Experiments.

2.1 Numerical Evaluation

We provided pretrained models for evaluation. Run bash scripts/evaluation.sh to reproduce the results that presented in the paper. In addition, the RMSE results on 13 unseen trajectories are shown in dump/evaluation/rmse_result.csv, which corresponds the results of Table I in the paper. If you want to train your own LLS, please follow the instructions below.

• Processing BEM data

  The dataset for training and testing in this work is adopted from BEM dataset. For sake of convenience, we provided BEM dataset in the data folder. Navigate to the data/BEM folder, run python process_data.py , then some figures and files will appear in the data/BEM folder

• Training LLS

  Run bash scripts/train.sh and this will take a couple minutes (The training time on i9-12900H CPU was around 8 min, and training time on a RTX 3060 laptop GPU was around 5 min). After training, the trained LLS will be located in the dump/evaluation folder.

• Evaluating Neural Predictor

  Run bash scripts/evaluation.sh. The validation results on 13 unseen trajectories will be located in dump/evaluation/test. In each trajectory folder, there are two figures that show the prediction results of Neural Predictor.

• Plotting

  Navigate to the plot folder. The BEM_Comparasion subfolder corresponds the result of Fig. 2 presented in paper. The Sample_Efficiency subfolder corresponds the result of Fig. 3 presented in paper.

2.2 Physical Experiments

We evaluate the Neural Predictor in the real-world experiments. The setup for real-world flight experiments is illustrated in Section VI.B of the paper.

Code: Coming soon

3. Citation

If you find this repo useful in your research, please cite our work:

@misc{jin2024neuralpredictorflightcontrol,
  title={Neural Predictor for Flight Control With Payload}, 
  author={Jin, Ao and Li, Chenhao and Wang, Qinyi and Liu, Ya and Huang, Panfeng and Zhang, Fan},
  journal={IEEE Robotics and Automation Letters}, 
  year={2025},
  volume={},
  number={},
  pages={1-8},
  doi={10.1109/LRA.2025.3573624}}