Physics-informed-Machine-Learning-with-Heuristic-Feedback-Control-Layer-for-AV-Control

Physics-informed Machine Learning with Heuristic Feedback Control Layer for Autonomous Vehicle Control.

Table of Contents

• Project Structure
• Controllers
• Key Features
• Dependencies
• Usage
• Results
• Author

---

Project Structure

.
├── DPCModel/                 # DPC controller files
├── HFAMPCModel/              # HFAMPC controller files
├── HFRPCModel/               # HFRPC controller files
├── RPCModel/                 # RPC controller files
├── simulation/               # Simulation scripts
├── DataSetGeneration/        # Scripts to generate training and testing data
├── results/                  # Output results and performance summaries
├── mainLaneChangingSixController.py  # Main script for testing all controllers
├── mainNumericalSimulationSingleExample.py  # Script for a single example simulation
└── PlotLoss/                 # Training loss visualization files

---

Controllers

This project implements the following controllers:

1. MPC (Model Predictive Control)
2. AMPC (Approximate MPC)
3. HFAMPC (Heuristic Feedback AMPC)
4. DPC (Differentiable Predictive Control)
5. RPC (Recurrent Predictive Control)
6. HFRPC (Heuristic Feedback RPC)

---

Key Features

• Physics-Informed Learning: Combines machine learning with model-based control for enhanced generalization.
• Heuristic Feedback Layer: Improves steady-state error and generalization.
• Performance Comparison: Numerical and graphical analysis of computational efficiency, trajectory accuracy, and generalization capabilities.

---

Dependencies

To run the project, the following dependencies are required:

• Python 3
• numpy
• pandas
• matplotlib
• torch
• casadi
• CUDA

---

Usage

1. Generate Dataset: Use DataSetGeneration/ scripts to generate training and testing data. You can also download the dataset using the following link:
   Google Drive Dataset

2. Train Controllers: Run training scripts in the respective model directories (e.g., RPC_train_GPU.py).

3. Run Simulations:

   • Single example simulation: mainNumericalSimulationSingleExample.py
   • Multiple controller comparison: mainLaneChangingSixController.py

4. Visualize Results:

   • Training loss: Training Loss over Epochs.png.

---

Results

• Performance Analysis:
  • Performance comparisons are shown in Lane-Change Instance Within Training Set Boundaries.png and Closed-loop Numerical Simulation Results.png.
• Training Loss:
  • Visualized in Training Loss over Epochs.png.

---

Author

Xianning Li
New York University
Email: [email protected]