Goal: Build an AI-assisted framework that generates multiple diverse implementations of software for an autonomous system while ensuring that each implementation still accomplishes the required mission.
Why this matters: Deploying identical software across many autonomous vehicles creates a software monoculture: knowledge gained from one implementation may transfer to every other unit. This project explores whether large language models and automated program transformations can create meaningfully different software variants while retaining required behavior.
The team will investigate techniques for generating software variants, automatically execute them in a controlled environment or simulator, and measure both correctness and diversity. The initial work can use manageable software components rather than an entire drone stack; a drone/autonomous-system simulator can be added as a case study if feasible.
The team should investigate the following research questions:
RQ1. How successfully can AI generate diverse implementations that still satisfy the same requirements?
RQ2. How should implementation and behavioral diversity be measured?
RQ3. What generation strategies provide the best tradeoff between diversity and correctness?
Expected outcome: A prototype variant-generation and evaluation framework, an empirical study of generated variants, and preliminary results that could support future research on resilient autonomous systems.
working knowledge of LLMs/code generation; automated testing; program analysis; software diversity; potentially autonomous-system simulation. Feel free to enroll in CS 499/569 offered in Fall 2026 to learn more about these concepts. More guidance and resources will be provided as the project progresses.
Manish Motwani: motwanim@oregonstate.edu
