Code for the paper "ARMOR: Attack-Resilient Reinforcement Learning Control for UAVs"

Key Features

• Attack-resilient model-free RL controller for robust UAV operation under adversarial sensor manipulation (e.g., GPS spoofing).
• Learns robust latent state representations of the UAV's physical state for reliable control under sensor corruption.
• Two-stage training framework: a teacher encoder trained with privileged attack information generates attack-aware latent states for RL policy training; a student encoder learns to mimic the teacher using only observable (attacked) sensor data at deployment.
• Compatible with UAV autopilot pipelines for adaptive, model-free control under adversarial conditions.

Installation

Tested on Ubuntu 22.04

git clone https://github.com/dashpritam/safe-rl.git
cd safe-rl/

conda create -n drones python=3.10
conda activate drones

pip3 install --upgrade pip
pip3 install -e . # if needed, `sudo apt install build-essential` to install `gcc` and build `pybullet`

Running Scripts

1. Teacher

Train (or evaluate with flag) a reinforcement learning (RL) teacher and encoder model using:

python train_teacher.py --vehicle drone --rl_file configs/rl.yaml \
  --environment_file configs/env.yaml --teacher_file configs/teacher.yaml \
  --attack_file configs/attack.yaml --output_folder teacher_exp

Expected Output Directory: results/train_teacher/<timestamp>

Arguments:

• --vehicle: Vehicle type (drone or car)
• --rl_file: RL model config YAML
• --environment_file: Environment config YAML
• --teacher_file: Teacher encoder config YAML
• --attack_file: Attack config YAML (training & eval fallback)
• --output_folder: Output directory (default: timestamped)
• --gui: Enable GUI (default: False)
• --eval_attack_files: Comma-separated attack YAMLs for post-training eval
• --num_eval_episodes: Episodes per attack (default: 10)
• --eval_only: Skip training, only evaluate (requires --rl_model)
• --rl_model: Path to saved PPO model (.zip) for eval

---

2. Student

Train (or evaluate with flag) a student encoder model using:

python train_student.py --vehicle drone --rl_file configs/rl.yaml \
  --environment_file configs/env.yaml \
  --rl_model results/train_teacher/01.15.2025_14.30.00/models/best_model.zip \
  --student_file configs/student.yaml --teacher_file configs/teacher.yaml \
  --attack_file configs/attack.yaml --output_folder student_exp

Expected Output Directory: results/train_student/<timestamp>

Arguments:

• --vehicle: Vehicle type (drone or car)
• --rl_file: RL model config YAML
• --environment_file: Environment config YAML
• --rl_model: Trained teacher PPO model (.zip)
• --student_file: Student encoder config YAML
• --teacher_file: Teacher encoder config YAML
• --attack_file: Attack config YAML
• --output_folder: Output directory (default: timestamped)
• --gui: Enable GUI (default: False)
• --eval_attack_files: Comma-separated attack YAMLs for final eval
• --num_eval_episodes: Episodes per attack (default: 10)
• --student_encoder: Pre-trained student .pth for warm-start or eval
• --skip_training: Skip training, only evaluate (requires --student_encoder)

---

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{dash2026armor,
      title={ARMOR: Attack Resilient Reinforcement Learning Control for UAVs}, 
      author={Dash, Pritam and Chan, Ethan and Lawrence, Nathan and Pattabiraman, Karthik},
      booktitle = {IEEE International Conference on Robotics & Automation (ICRA)},
      year={2026}
}