This project aims to investigate vulnerabilities in AI-driven autonomous drone systems and develop a reusable framework for adversarial testing, verification, and validation.
We will study how adversarial attacks can affect learning-based components of a drone autonomy stack, develop methods to systematically discover and characterize failures, and investigate defensive methods to improve robustness. Our current focus is
A major goal of the project is to build a , rather than an attack that works only on one specific neural network. The framework should make it possible to plug in different autonomy algorithms and systematically evaluate their robustness across nominal, off-nominal, and adversarial scenarios.
Students may work on topics including:
- Learning-based drone obstacle-avoidance algorithms and implementation of them on simulation and actual drones
- Automated scenario and test-case generation
- Defensive methods against discovered vulnerabilities
- Strong programming, communication, and collaboration skills
- Proactive and fast self-learner
- Interest in robotics, autonomous systems, machine learning, or cyber security
- Familiarity with learning-based obstacle avoidance algorithms.
- Familiarity with cyber security and vulnerabilities
- Familiarity with software: ROS2, Python, C++, Linux, Bash, Git, docker, Isaac Sim
- Integrate and evaluate modern learning-based perception and obstacle-avoidance algorithms for drones
- Learn how adversarial attacks are formulated and evaluated against autonomous systems
- Design reproducible simulation experiments and automated test scenarios
- Contribute to the design of a reusable verification and validation framework
- Work with graduate students, researchers, and faculty across robotics and cyber security
- Potentially contribute to research publications, demonstrations, and open-source software
Not a hard limit. We care more about talent and potential.