View Proposal

---

Proposer

Oliver Lemon

Title

Safety of Embodied AI

Goal

To test the safety of embodied AI models (different LLMs and VLMs, and VLAs) and attempt to make them safer!

Description

You will explore, extend, and modify our existing simulation environments for testing the safety of embodied AI models: See for example https://openreview.net/forum?id=J75kryV3RF You will explore adversarial red-teaming to try to jailbreak embodied AI models though various attack methods. See for example the ASIMOV papers from Google Deepmind. Some attacks may be via speech, for example: https://arxiv.org/pdf/2608.28518 You will explore mitigation strategies such as prompting, monitor agents, reinforcement learning, and formal verification. You may extend our existing software , which is used at the UK Robotics Summer School : https://github.com/olemon1/ukras-genair-lab

Resources

See code at https://github.com/olemon1/ukras-genair-lab Ollama LLMs and VLMs

Background

AI, evaluation, NLP, LLMs, robotics

Url

External Link

Difficulty Level

Moderate

Ethical Approval

InterfaceOnly

Number Of Students

4

Supervisor

Oliver Lemon

Keywords

generative ai, safety, llm, vlm, vla, human-robot interaction, evaluation

Degrees

Bachelor of Science in Computer Science
Master of Engineering in Software Engineering
Master of Science in Artificial Intelligence
Master of Science in Artificial Intelligence with SMI
Master of Science in Data Science
Master of Science in Human Robot Interaction
Master of Science in Robotics
Bachelor of Engineering in Robotics
MSc Applied Cyber Security