Paper | GitHub | arXiv | Published in L4DC 2025 (Oral) + Submitted to IEEE TAC.

This repository is the official implementation for the papers titled "Neural Operator Predictors for Delay-Compensated Nonlinear Stabilization" and "Neural Operators for Predictor Feedback Control of Nonlinear Delay Systems". Please refer to the example below for a quick notebook reproducing the paper results. We have included pretrained datasets and models as well as the code to train one's own models and dataset if interested.

Note, since the original L4DC publication, these result have been improve and updated to the latest packages as of May 2025. Thus, the direct numerical values may differ slightly from those published in the paper, but the overall result remains the same.

• To get started, please setup your virtual envrionment (Virtual Env tutorial) and install the corresponding packages and versions given in requirements.txt.
• Additionally, we have published both the dataset and models on huggingface. Please clone the repositories below following the instructions on huggingface and place the resulting files in the dataset and models folders of this repository (See below for structure).
  • Hugging face: Dataset
  • Hugging face: Models

The repository directory should look like this:

NeuralOpertorPredictorFeedback/
├── examples/
├── src/
├── config/
├── datasets/ # Place the cloned datasets here
│   ├── ManipulatorDatasets/    
│   ├── UnicycleConstDelayDataset/   
│   ├── UnicycleTimeVaryingDelayDataset/
├── models/ # Place the cloned models here
│   ├── ManipulatorModels/    
│   ├── UnicycleConstDelayModels/   
│   ├── UnicycleTimeVaryingDelayModels/
├── media/ 
•   •   •
•   •   •

See examples folder for two notebooks that have implementations of neural operator predictors for both constant and time-varying delays for the unicycle system. These models are quick to train and thus can be run with any modern GPU system (and higher end laptops) although we do provide pretrained models for those who are interested. The Jupyter-notebooks should be self-explanatory as they have section headers guiding one through each portion of the code.

See examples folder for a third notebook that replicates the manipulator results for a 5DOF manipulator that contains the same parameters as a single arm of a Baxter manipulator without the last two joints. To generate a dataset for the manipulator, it takes significant computation time as the example's forward dynamics are expensive. Furthermore, for the models, they can be trained either through the notebook or using the train_model.py file given that one modifies the configuration. Without significant computing resources, it is not recommended that the user develops their own models for the manipulator.

If you have any issues with any of the notebooks or models in this repo, feel free to create an issue in this github repo or email [email protected] and I am more than happy to assist!

If you found this work useful or interesting for your own research, we would appreciate if you could cite our work:

@misc{bhan2025neuraloperatorspredictorfeedback,
      title={Neural Operators for Predictor Feedback Control of Nonlinear Delay Systems}, 
      author={Luke Bhan and Peijia Qin and Miroslav Krstic and Yuanyuan Shi},
      year={2025},
      eprint={2411.18964},
      archivePrefix={arXiv},
      primaryClass={eess.SY},
      url={https://arxiv.org/abs/2411.18964}, 
}

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.