Neural-ESO: Provably Robust Learning-Based Control for Quadrotors

This repository contains the official implementation of the Neural-Extended State Observer (Neural-ESO) for robust quadrotor flight control. By bridging traditional ESO with a learning-based dual-pathway architecture, this framework dynamically predicts and compensates for unmodeled dynamics and external disturbances in real-time.

The codebase includes offline PyTorch training scripts with spectral normalization (Lipschitz constraints) and online ROS/MAVSDK deployment scripts for hardware flight testing using PX4 and OptiTrack.

Features

• Data-Driven Disturbance Prediction: Utilizes a lightweight Multi-Layer Perceptron (MLP) to predict Z-axis disturbances ($z_3$) based on 5 system states (altitude, velocity, thrust, roll, pitch).
• Provable Stability: Incorporates Spectral Normalization during training (Lipschitz constant $K=1$) to bound the neural network's output and guarantee closed-loop stability.
• Linear ESO Core: Robust baseline control using Linear ESO, modified to accept neural feedforward commands alongside the standard Extended State Observer.
• Hardware-Ready Deployment: Fully asynchronous Python deployment script utilizing MAVSDK for PX4 offboard control and vrpn_client_node for high-fidelity OptiTrack motion capture.

Repository Structure

• TrainingBoxLandingDNNchangable.py - PyTorch training script for the Neural-ESO.
• Test_BoxLandingLearningWithNN.py - ROS/MAVSDK online deployment and flight script.
• requirements.txt - Python package dependencies.
• README.md - This documentation file.

Requirements & Installation

The system requires Python 3.8+, ROS Noetic, and a PX4 Autopilot environment (either SITL or physical hardware).

1. Download and extract the code: Download the provided .zip file and extract it to your local workspace.

   cd path/to/extracted/Neural-ESO-3051
   

2. Install the required Python dependencies:

   pip install -r requirements.txt

   (Note: Ensure your ROS Noetic environment is sourced properly so rospy and geometry_msgs are available).

Usage

Phase 1: Train the Neural-ESO Model

The neural network learns the disturbance mapping from baseline flight logs.

1. Place your baseline ESO flight log (.csv) in the directory.
2. Update the FILENAME_TO_LOAD variable in TrainingBoxLandingDNNchangable.py.
3. Run the training script:

   python TrainingBoxLandingDNNchangable.py

4. Outputs: The script will generate three files required for deployment:
   • disturbance_model_BoxLanding.pth (Model Weights)
   • scaler_X_BoxLanding.pkl (Input Scaler)
   • scaler_y_BoxLanding.pkl (Output Scaler)

Phase 2: Hardware Deployment

Ensure your OptiTrack system is publishing pose data to /vrpn_client_node/Drone1/pose and your PX4 autopilot is listening for MAVSDK UDP connections on udp://:14550.

1. Launch the flight script:

   python Test_BoxLandingLearningWithNN.py

2. Flight Sequence: The quadrotor will automatically arm, take off to a hover state to lock its initial position, and then execute the pre-programmed trajectory while running the Neural-ESO feedforward loop at a target rate of 50Hz ($dt = 0.02s$).
3. Logging & Visualization: Upon landing, the script will automatically save a new .csv flight log and generate matplotlib figures comparing the actual trajectory, ESO residual estimations, and NN predictions.

License

This project is licensed under the MIT License