ESAB ROBBI AI, Announced Oct. 22: A Self-Learning Welding Cobot for the Weld Cell

Low-volume, high-mix fabrication has always sat in the awkward corner of the automation roadmap: joints that fit up differently from one part to the next, scarce specialized welding skills, and processes such as MIG and tungsten inert gas (TIG) welding that fight fixed tool paths. In an Oct. 22 announcement, ESAB aimed its new ROBBI AI collaborative robot platform squarely at that corner, pairing artificial intelligence with camera and laser sensing so that welding motion and process parameters can change while the arc is on instead of after the pass is finished.

Read from a production floor, the meaningful change is not “one more arm in a cell.” It is the relocation of decision-making into the process itself. The system builds its own weld path rather than depending on predefined paths and fixed weld sequences, and it decides how to proceed as welding proceeds. What the plant still has to supply up front is procedure discipline: before the process begins, operators enter welding procedure parameters, including consumables, material type and weld type.

What the platform changes about programming effort

“Physical AI brings intelligence directly into the welding process,” said Dan Colvin, vice president of Advanced Technology Solutions at ESAB, framing the design intent. “Instead of programming every movement and anticipating every possible variation, ROBBI AI assesses the joint, determines how to proceed and can make adjustments while welding. It is also designed to learn from every weld, helping it recognize conditions that can lead to defects and respond by adjusting welding parameters or stopping the weld when needed.”

For managers evaluating retrofit projects, that reframes the cost curve. Offline teaching and variation-proofing normally consume the hours that make short-run work unattractive to automate. Here the declared behavior is joint assessment at the start of the weld and correction during it, within parameters set for the application.

Adapting in flight as fit-up shifts

The robot uses its sensing to evaluate the workpiece, identify the expected joint and decide how to proceed inside the limits established for the application.

“As welding progresses, ROBBI AI can adapt its motion and welding parameters to changing conditions,” Colvin said. “For example, if a joint gap opens or closes, the system can adjust its travel motion, oscillation or welding parameters as appropriate within the established procedure. Rather than relying solely on a predetermined weld sequence, it can make welding decisions in real time.”

Cells that are not ready to hand that latitude to the software are not excluded: for users who prefer traditional operation, the platform also supports conventional cobot use, and full-AI mode can be disabled.

Data capture, review loops and procedure governance

Learning before and after commissioning

Before installation, ROBBI AI completes a simulation-based learning period. Once it sits in a customer’s facility, it keeps collecting weld data and associated video that operators can review to investigate defects and find improvement opportunities.

When a condition that could compromise a weld is detected, the system can halt welding. Operators then review the recorded information and give feedback that helps the AI recognize similar conditions in future work, ESAB said. The practical implication for supervision is that the review step becomes part of the standard work of the cell, not an afterthought reserved for root-cause analysis.

Where the compute lives

Artificial-intelligence processing occurs locally on a computer inside the customer’s facility, avoiding reliance on continuous communication with a cloud-based system. ESAB said this keeps learning data onsite and supports real-time decision-making — two constraints that matter to plants with limited network reliability or strict data-handling rules.

Keeping AI decisions inside the qualified procedure

The company also relies on InduSuite software to help keep AI-driven welding decisions within established welding procedure specifications and qualification requirements.

“We use our existing InduSuite software to help keep AI-driven welding decisions within predefined welding procedure specifications and qualification requirements,” Colvin said. “ESAB’s approach is designed to reduce the risk of AI-driven adjustments that could result in welding outside those specifications or with an unapproved procedure.”

Hardware choices and the stack around the arm

Element	Available options	Effect on cell layout
Robot arm	Universal Robots UR10e arm or a UR20e arm	Sizing follows part mass and reach needs
Extended reach	seventh-axis configuration	Extends reach and supports a larger work envelope
Mobility	Wheeled cart	The cell can be moved around a shop
Welding power sources	EWM Titan XQ, EWM Tetrix XQ and ESAB Aristo Edge 500R	Compatible with listed sources
AI processing	Local computer inside the customer’s facility	No continuous cloud communication required
Procedure control	InduSuite	Constrains AI decisions to qualified procedures

Application case: TIG and pipe work for infrastructure

ESAB identifies TIG welding and pipe fabrication for data centers and other infrastructure projects as potential applications for ROBBI AI, especially where specialized welding skills are required and conventional automation can be difficult to implement.

The technical reason that target is narrow is well documented. TIG welding can be difficult to automate because it requires precise control of the torch, filler wire and molten weld pool. In research published in the International Journal of Pressure Vessels and Piping, investigators studying robotic orbital TIG welding of 304L stainless-steel pipes found that maintaining a stable weld pool during the root pass was particularly challenging as the torch moved around the pipe and the welding position changed.

A manual welder can continually adjust torch trajectory and filler-metal feed to control the pool; reproducing that control through automation requires carefully developed welding parameters. The study also found that the process remained sensitive to geometric variations in pipe joints. That sensitivity to geometry is precisely the variable fit-up problem that has kept this class of work manual.

“With ROBBI AI, we are extending the role of AI beyond digital analysis and into the physical welding process, where intelligent decisions can directly influence how a weld is made,” Colvin said.

Common misconceptions

• “Self-learning means the procedure specification disappears.” It does not: operators still supply consumables, material type and weld type, and InduSuite is used to keep AI-driven decisions within predefined welding procedure specifications and qualification requirements.
• “The cell must stay connected to a cloud service to keep learning.” The architecture described places AI processing on a local computer inside the customer’s facility, so learning data stays onsite and does not depend on continuous cloud communication.
• “Full-AI mode is all or nothing.” The platform supports conventional cobot use, and full-AI mode can be disabled for users who prefer traditional operation.
• “Any weld process can be dropped in and automated.” MIG and TIG are named as targets precisely because their variables have made them difficult to automate, and TIG remains constrained by torch, filler-wire and weld-pool control plus joint geometry sensitivity.
• “Detection and halt replaces quality supervision.” When a compromising condition is detected and welding is halted, operators still review the recorded weld data and video and provide the feedback that trains recognition of similar conditions.

Action recommendations for production and engineering teams

1. Inventory the work that is genuinely hard to automate today: variable joint fit-up, low-volume, high-mix production, pipe fabrication and TIG tasks, along with the MIG and TIG variables that defeated earlier automation attempts.
2. Get the procedure inputs under control first, since consumables, material type and weld type remain the operator-supplied parameters that bound the system’s decisions.
3. Plan for the simulation-based learning period in the commissioning schedule, and confirm the local compute location inside the facility before go-live.
4. Build a defect-review routine around the stored weld data and video, and treat operator feedback after a weld halt as a formal step so the system learns the condition once instead of repeatedly.
5. Verify power-source compatibility against EWM Titan XQ, EWM Tetrix XQ and ESAB Aristo Edge 500R before committing a cell layout.
6. Size the mechanical envelope deliberately: choose a Universal Robots UR10e arm or a UR20e arm, add a seventh-axis configuration where reach and work envelope require it, and consider a wheeled cart where the same asset must serve multiple areas of the shop.
7. Start with conventional cobot operation where shop-floor trust is thin, and enable AI-driven decisions only after procedure and review workflows are stable.
8. Document how AI-driven adjustments stay within qualified procedures, using InduSuite as the stated control point, so the quality function can sign off on the process.

FAQ

What did ESAB announce on Oct. 22?

A collaborative robot platform, ROBBI AI, that applies artificial intelligence together with camera and laser sensing to change welding motion and process parameters as conditions shift.

Which welding work does it target?

Metal fabrication with variable joint fit-up, low-volume, high-mix production, and tasks that have been hard to automate — including pipe fabrication and TIG welding, plus the variables that have made MIG and TIG difficult to automate.

Does the operator still program a weld path?

No. Rather than depending on predefined paths and fixed weld sequences, the system creates its own weld path and decides what to do as welding proceeds, after operators supply welding procedure parameters.

Can it stop itself?

Yes. When a condition that could compromise a weld is detected, the system can halt welding, after which operators review the recorded information and give feedback that helps the AI recognize similar conditions later.

Where is the AI processed?

Locally, on a computer inside the customer’s facility, which avoids reliance on continuous cloud communication and keeps learning data onsite.

What hardware can be configured?

A Universal Robots UR10e or UR20e arm, a seventh-axis configuration for extended reach and a larger work envelope, or a wheeled cart for mobility, with EWM Titan XQ, EWM Tetrix XQ and ESAB Aristo Edge 500R listed as compatible welding power sources.

For plants that have postponed automation because the work was too variable and the volumes too small to justify teach time, the announcement marks a different bet: put the sensing, the learning and the procedure guardrails inside the arc, and let the cell absorb the variation. The work left for the manufacturer is unchanged in kind — clean procedure inputs, disciplined review of the recorded data, and a clear decision about how much latitude the software should hold in each cell.