Physical AI

Vision inspection was the starting point. The lab is now expanding into Physical AI — perception-driven autonomy for manufacturing, where models do not only classify images but inform movement, handling and decisions on the floor.

The original experiment established the practical baseline: controlled capture, an annotated defect dataset, and detection models evaluated against real cycle-time constraints. Its findings — that capture conditions dominate model choice — carry directly into the next phase.

Perception feeding action: fixed and mobile cameras, ground platforms moving material, aerial inspection of large assets, and manipulation guided by live perception rather than fixed programs.

Manual visual inspection is inconsistent across shifts and operators, and the conditions that make it hard — lighting, fixturing, cycle time — are exactly the conditions a vision system has to survive. The same constraints apply, harder, once perception drives physical action.

Controlled capture conditions matter more than model architecture for practical industrial defect detection. Extending that further: reliable embodied behaviour depends more on constraining the physical environment and the perception loop than on the size of the model driving it.

• Capture setup with controlled lighting and fixturing.
• Image dataset with defect annotation.
• Detection and classification model comparison.
• Inline evaluation against cycle-time constraints.
• Perception layer shared between inspection and motion decisions.
• Ground and aerial platforms as data collectors and actors.
• Safety and supervision boundaries around any perception-driven action.

How much detection performance depends on capture conditions versus model selection — and how far that same perception stack can be trusted when its output triggers physical movement.

• Capture consistency dominated early results.
• Rare defect classes need deliberate collection; they do not appear in normal production sampling.
• Latency budgets tighten sharply once perception output drives motion rather than a report.