FANUC and Google build an AI agent that welds straight from blueprints

Teaching an industrial robot how to weld has always been a painful, time-consuming bottleneck. Technicians usually spend hours manually programming waypoints and guessing voltage settings for every single new part.

FANUC and Google just teamed up to change that. On September 11, the robotics giant introduced the AI Welding Agent, a Physical AI platform that can read an engineering drawing and start welding with zero setup and zero teaching.

Show the tablet a drawing, walk away

The entire workflow basically comes down to taking a picture.

Using the built-in camera on a standard FANUC CRX tablet teach pendant, an operator snaps a photo of an arc-welding blueprint. That is all the visual input it needs.

Under the hood, Google's Gemini Enterprise analyzes the drawing to understand the materials, joint designs, and final target shape. It then generates the required welding parameters, such as current and voltage, alongside the complete robot motion paths.

Workers can either run the auto-generated code right away or fine-tune specific parameters before pulling the trigger.

Designed for factory floors, not lab demos

What makes this launch notable is how low-friction the implementation looks.

Factories do not need dedicated computer vision rigs or external sensors; the tablet's existing camera is enough. It is also agnostic to welding hardware, meaning it works with any welding power source compatible with a FANUC robot.

On the business side, manufacturers do not have to sign a separate enterprise deal with Google Cloud. The service can be activated directly through existing FANUC contracts.

To address industrial IP concerns, FANUC noted that all blueprints and production files run through Gemini Enterprise safeguards and will not be used to train external AI models.

FANUC demonstrated the setup at the International Welding Show in Tokyo on September 16, with commercial shipments slated for late December.

Why it matters

Skilled arc welders are in desperately short supply across shipbuilding, construction, and automotive manufacturing. High-mix, low-volume shops often avoid robotics simply because programming each batch takes longer than welding it by hand.

If FANUC's system delivers on its "zero-teaching" promise, it dramatically lowers the barrier to automating custom metalwork.

We think FANUC took a very smart approach here. Instead of forcing clients to overhaul their shop floors with proprietary sensors and complicated cloud vendor setups, they piggybacked on existing tablet hardware and established distributor relationships.

That said, blueprints represent ideal geometry, while real metal parts have fit-up gaps, burrs, and thermal distortion. Whether Gemini's drawing-level reasoning can adapt to real-world fabrication imperfections without real-time seam tracking remains the big question to watch once shipments begin in December.