Doctoral Research
Trustworthy Physical AI for Robotic Welding
Toward trustworthy, LLM-mediated robotic welding: vision-based weld-bead control and secure human–robot interaction on a ROS 2 middleware.
Overview
Welding underpins Canada's heavy-manufacturing base — shipbuilding, pressure vessels, pipelines — yet the sector faces a chronic shortage of skilled welders alongside rising demands for quality and consistency. A new wave of research proposes foundation-model welding, where large language models turn a CAD model and a plain-language request into an executable weld. That is transformative for high-mix, low-volume fabrication, but it routes the output of an opaque model straight onto actuators that move a robot and strike an arc, opening a safety and trust gap that no current welding system addresses.
My doctoral research develops and experimentally validates a trustworthy, vision-in-the-loop, LLM-mediated welding system on an industrial ROS 2 / Fanuc platform. Perception closes the loop on bead geometry, a language interface makes the cell operator-directable, and a layered security architecture makes the language-to-actuator path verifiably safe. The work is carried out across SFU's School of Mechatronic Systems Engineering, the SFU Trustworthy AI Lab, and an instrumented industrial welding cell at BCIT.
Research Thrusts
Three interlocking thrusts on a shared ROS 2 / Fanuc platform — each starting from a working baseline rather than from scratch.
Vision-based perception & closed-loop control
A passive eye-in-hand camera estimates weld-bead width and height in real time, closing the loop on bead geometry through a ROS 2 → Fanuc interface. A feed-forward model predicts the travel and wire-feed speeds needed to hit a target bead shape before deposition begins, and the pipeline is characterized for deployment on low-cost embedded hardware.
LLM-mediated human–robot interaction
A multimodal interface maps a natural-language request and a CAD model to a weld-seam trajectory and process parameters. It is prototyped on a hosted model and matured into a locally fine-tuned, welding-specific model that runs on-premise — preserving data privacy and staying operational without a network connection.
Trustworthiness & security of the controller
A layered, latency-aware defense — a deterministic constraint guard at the controller boundary, a domain-specific constitutional classifier, and human-in-the-loop arming — so that no operator, model, or attacker-issued command can drive the cell outside its safe envelope. Evaluation is adversarial, with a route toward certified guarantees.
Beyond the Bench: Cybersecurity, AI Ethics & Governance
My welding work is one instance of a broader question I care about — how powerful but opaque AI can be deployed safely, securely, and accountably in systems that act in the physical world.
Cybersecurity
Moving from a hosted API to locally fine-tuned open weights shifts the trust boundary inward: the weights and training data become a concrete attack surface. I study backdoor and poisoning threats to AI-driven cyber-physical systems, adversarial red-teaming of the language-to-actuator path, and the deterministic guards that contain them.
AI Ethics & Safety
As powerful but opaque models begin to act in the physical world, I care about deploying them safely and responsibly — balancing capability against over-refusal, keeping a human meaningfully in the loop, and pursuing certified, verifiable guarantees rather than best-effort assurances.
AI Governance
I am interested in how AI can be governed as it enters critical manufacturing — provenance and audit trails for every command, safe-operating envelopes tied to engineering standards, accountability for automated decisions, and the policy questions raised by autonomous systems on the factory floor.
Supervision & Collaboration
The research spans robotics and trustworthy-AI groups, pairing an industrial welding platform with certified-robustness expertise.
Senior Supervisor
Prof. Mehrdad Moallem
School of Mechatronic Systems Engineering, Simon Fraser University
Co-Supervisor
Dr. Linyi Li — SFU Trustworthy AI Lab
Certified robustness, trustworthy AI, and LLM safety & security
Visit the labExperimental Platform
BCIT Centre for Welding Technologies & Metallurgy Research
Industrial Fanuc CRX-10iA cell with a Lincoln Power Wave R450 supply
Interested in trustworthy AI, robotics, or the security of cyber-physical systems? I am always glad to talk about collaborations and research directions.