Neural Internal Model Control

This repository contains the implementation of Neural Internal Model Control (Neural-IMC) for quadrotor control, as presented in our IEEE Robotics and Automation Letters paper Neural Internal Model Control: Learning a Robust Control Policy Via Predictive Error Feedback. For the quadruped robot experiments, please refer to our dedicated repository.

Installation

conda create -n torchctrl python=3.9
conda activate torchctrl
pip install -e .

Trajectory examples used for experiments

Trajectory Type	Description	3D Trajectory	Per-axis Trajectory	Implementation
Circle	Circular trajectory			circle.py
Poly	Chained polynomial trajectories			chained_polynomial.py
Star	5-pointed star trajectory			pointed_star.py
Zigzag	Zigzag trajectory			zigzag.py

Usage

Tuning the controller parameters

conda activate torchctrl && cd scripts
# By default, it will use wandb to log data, please make sure you have set WANDB_API_KEY in your environment variables.
python run_quadrotor.py
# To run without wandb logging, use the following command:
python run_quadrotor.py wandb.mode=disabled

Refer to scripts/train/shell_scripts for more experiment scripts.

Quadruped experiments

Check this repo for more details.

Citation

If you find our codebase or our proposed technique (predictive error feedback) useful, please cite our paper:

@ARTICLE{gao2025neuralimc,
  author={Gao, Feng and Yu, Chao and Wang, Yu and Wu, Yi},
  journal={IEEE Robotics and Automation Letters}, 
  title={Neural Internal Model Control: Learning a Robust Control Policy Via Predictive Error Feedback}, 
  year={2025},
  volume={10},
  number={7},
  pages={6848-6855},
  keywords={Predictive models;Robots;Quadrotors;Adaptation models;Legged locomotion;Mathematical models;Training;Robustness;Robust control;Adaptive control;Robust control;reinforcement learning (RL);sensorimotor learning},
  doi={10.1109/LRA.2025.3573169}}