en

Industrial AI for welding robots

Robot welding automation is about automatically converting engineering intent into executable robot programs with speed, consistency, and traceability.
Starting from messy CAD data and incomplete specifications, it should generate deterministic, production-ready robot code that meets industrial requirements for quality, throughput, and repeatability.

Beyond physical AI

AI models can be trained to deal with incomplete or ambiguous specifications and unstructured environments , but fall short in industrial automation contexts requiring tight quality tolerances and strict validation and approval processes.

That is why most AI approaches fail when exposed to real welding production.

ArcNC developed an Industrial AI framework designed to run production cells: leveraging the power of AI, while adhering to the highest industry standards for quality, throughput and certifiability.

ArcNC uses a three-layer industrial AI architecture. Each layer has a clearly bounded role. None of them alone is sufficient, but together, they form a production-grade welding AI system.

  • 1. Function Models

    execute narrow, well-defined welding tasks

  • 2. Application Agents

    reason at application level and handle intent

  • 3. Industrial Platform

    enforces determinism, validation, and control


1. Function Models

Task-Specific Welding Intelligence

Function models are custom-trained AI models that execute one specific task.

Examples include:

  • Extracting weld seams from CAD geometry
  • Classifying joints and weld types
  • Selecting optimal robot configurations or sense locations
  • Determining optimal torch orientation and approach angles
  • Converting 2D drawings or images into structured 3D data

These models are deterministic, domain-specific, and verifiable.

They do not make global decisions. They provide precise, reliable outputs for a single well-defined function.

This is the execution layer of welding intelligence.

2. Application Agents

Context-Aware Welding Experts

Application agents operate at a higher level.

They are general-purpose models that:

  • Interpret natural language instructions
  • Read drawings, notes, and unstructured inputs
  • Interact with ArcNC like a human programmer
  • Decide which function models to use and when

Over time, they build an understanding of each customer:

  • Preferred weld parameters
  • Standard joint treatments
  • Company-specific conventions

If similar parts have been welded before, the agent reuses that knowledge.
Specifications do not need to be restated.

This largely automates the specification phase as well, not just the motion planning and process parameter assignment.

3. ArcNC Platform

Industrializing Welding AI

The ArcNC platform is the control layer that makes AI usable in production:

  • Enforces constraints on AI decisions
  • Guarantees deterministic execution paths
  • Validates results before output
  • Logs every decision and transformation
  • Supports inspection, replay, and auditing

AI is not allowed to act freely, as every decision is bounded, checked, and recorded.

This is what converts AI capability into industrial reliability.

Welding AI in production

Neither function models nor application agents are sufficient on their own.
Uncontrolled AI systems would introduce non-repeatable behavior, opaque decision paths, and lack the necessary traceability.

These properties are incompatible with industrial automation. Welding production requires bounded behavior, predictable outcomes, and full accountability.

In demanding production environment, ArcNC's Welding AI results in:

  • Reduced dependence on senior experts

  • Consistent weld execution across parts

  • Captured and reusable welding knowledge

  • Scalable automation for high-mix environments

and ultimately: fully autonomous robot welding.