| Filename | Latest commit message | Latest commit date |
|---|---|---|
| LICENSE | ||
| README.md | ||
Physics-guided residual learning for UAV trajectory prediction
This repository contains the code, final figures, and result tables for a journal submission on multi-step UAV trajectory prediction. The project evaluates a hybrid physics-guided residual-correction model against three baseline variants: a standalone physics-based rollout, an LSTM model, and an MLP fusion model.
The main idea is to use the physics-based prediction as a structured trajectory prior and then learn a data-driven residual correction. This allows the final model to retain useful physical guidance while compensating for drift, phase-transition effects, and unmodeled disturbances in real UAV telemetry.
What This Repository Contains
This is a lightweight sharing package for journal review. It includes the essential scripts, final paper figures, and reported metric files. Large raw datasets, trained model checkpoints, and large intermediate prediction arrays are intentionally excluded.
.
|-- src/
| |-- lstm_unified.py
| |-- Physics_model.py
| |-- MLP.py
| `-- PIML.py
|-- figures/
|-- results
| |-- metrics/
| |-- run_info_piml_v12.json
| `-- piml_summary.csv
|-- data/
| `-- README.md
|-- docs/
| |-- REPRODUCIBILITY.md
| `-- latex_tables.md
|-- requirements.txt
|-- CITATION.cff
`-- .gitignore
Model Variants
Physics_model.py: analytical physics-based trajectory rollout.lstm_unified.py: LSTM-only data-driven sequence prediction model.MLP.py: learnable MLP fusion baseline combining physics and LSTM predictions.PIML.py: proposed physics-informed residual-correction model.
Key Outputs
The figures/ folder contains the final paper-ready visualizations, including:
- horizon-error comparison over the forecast horizon
- phase-wise real-vs-predicted trajectory plots
- accuracy and physical-feasibility trade-off plot
- residual-correction magnitude plot
- physical-feasibility effect plot
- rolling forecast timeline diagram
The results/metrics/ folder contains the CSV and JSON files used to report:
- overall RMSE, MAE, ADE, and FDE
- phase-wise prediction performance
- physical-feasibility metrics
- residual correction magnitudes
- accuracy-feasibility trade-off data
Installation
Create a Python environment and install the dependencies:
pip install -r requirements.txt
Data Availability
The raw UAV telemetry data are not included in this lightweight repository package. To rerun the full training pipeline, place the dataset folders under data/ as described in data/README.md.
The included figures and metric files are sufficient to inspect the reported journal results without rerunning the full pipeline.
Reproducibility
See docs/REPRODUCIBILITY.md for the expected execution order and notes on the final reported result source.
Suggested GitHub Description
Physics-guided residual learning framework for multi-step UAV trajectory prediction, including LSTM, physics-based, MLP fusion, and PIML/PGRL model variants with final journal figures and evaluation metrics.