Job shops and fabricators make many different parts in small batches, so classic robot teach-in or offline programming costs more than the work itself. AI-driven cells scan every part, detect seams or surfaces, generate the tool path and process parameters, and adapt in closed loop while welding, sanding or polishing. This addresses skilled-welder shortages and makes automation viable for batch sizes down to one.
Problem & value
- Today: a robot programmer teaches points by pendant or generates paths offline from CAD; parts must be fixtured precisely. Variation (tack-weld position, gaps, casting tolerance, worn surfaces) is handled by skilled people. Welding, grinding and sanding are also physically hard and health-burdening jobs.
- AI value: no per-part programming, tolerance to part and fixture variation, consistent quality, documented process data, and use of existing robots on high-mix work. Vendors report large reductions in programming time and higher utilisation; treat these as vendor claims and verify on your own parts [1][2].
- Why distinct: ROBO-003 learns from a human demonstration or instruction; here paths derive from scan/CAD plus learned contact and process models. ROBO-001 is pick-and-place without a process; OPT-009 tunes parameters without a robot.
Variants
- Arc welding (MIG/MAG/TIG) of steel and aluminium structures: seam detection, multi-pass planning, seam tracking.
- Surface finishing: sanding, grinding, deburring, polishing of composites, metal, wood (aerospace, marine, furniture).
- Spray painting / coating: path and film-thickness planning from the scan.
- Collaborative welding / cobot cells for micro-SMEs with simplified programming (see ROBO-016 for safety).
- Large structures: shipbuilding, rail and construction steel with mobile or gantry-mounted robots.
- Laser, spot and friction welding with similar scan-to-path logic but different sensors.
How it works
- part placed in cell (loose fixturing)
- 3D scan (point cloud / mesh), optional CAD registration
- AI perception: seam / joint / surface segmentation, defect and gap detection
- path planning: tool path, orientation, collision-free motion planning
- process model: parameters from material, thickness, joint, finish target
- simulation / digital-twin check (reachability, collisions)
- operator review (approve, edit, release)
- execution with closed-loop sensing (seam tracking, force control, arc signals)
- inspection and quality record (scan after weld/finish, NDT link: VIS-053)
- Perception: point-cloud segmentation and learned seam detectors localise joints even on tack-welded, dirty or reflective parts.
- Planning: geometric path generation plus motion planners (e.g. MoveIt-class); learned models propose parameters; reinforcement learning is used for contact-rich finishing policies.
- Closed loop: laser seam tracking, adaptive fill, force/torque control and post-process scan trigger rework.
- Human-in-the-loop: a qualified welder/programmer approves the first article and parameter sets; certified welding procedures stay under a responsible welding coordinator.
Data preparation
The AI needs three kinds of data: 3D scans of parts, CAD models and weld drawings or WPS, and process records (material, thickness, joint type, sensor streams). In an SMB these typically live in CAD/PDM folders, on file shares, in the ERP (order and material master data) and partly on paper (drawings, weld logs).
Typical gaps: outdated or missing CAD for repeat or reworked parts, scanned drawings without metadata, WPS documents scattered across folders, inconsistent material names, no record of which parameters produced good seams, and no labelled examples of defects.
Preparation steps: collect drawings and WPS in one structured location (PREP-009, PREP-010), digitise paper logs (PREP-006, PREP-008), harmonise material and joint-type names (PREP-019), export order and material data from the ERP (PREP-005) and start capturing quality results per part as labels (PREP-020). Calibrate sensors and fix the scan setup early, because scan quality drives path quality more than any model choice. Use open tools such as Open3D or CloudCompare to check point-cloud quality.
Rough share of effort: about 20 to 30 percent of the project. With a bought cell much of this shrinks to clean master data and one-off parameter libraries; for own builds it is higher. See GUIDE-003 for the readiness check.
Implementation paths
Which path for whom
- SMB: Path A, because mature welding and finishing cells with built-in scanning and path generation exist at SMB prices and the process (seams, surfaces) is standard.
- Enterprise: Path B, because own cells, CAD/PLM data and weld documentation must be integrated with vision and path-planning services.
- Developer: Path C, because open-source stacks (point-cloud libraries, ROS 2 MoveIt, simulation) run on HW1 and allow building own cell software.
Path A — Buy (recommended for most SMBs)
- Autonomous welding: Path Robotics (US; autonomous welding cells with AI vision), Lincoln Electric, Fronius, Yaskawa, FANUC, ABB, KUKA, Universal Robots welding kits; EU: Carl Cloos Schweißtechnik, Fronius (AT), Pemamek (FI), Valk Welding (NL), Inrotech (DK, seam detection), Novarc (CA) for pipe.
- Surface finishing: GrayMatter Robotics (US; Scan&Sand, deployed with Janicki Industries) [2], 3M robotic finishing partnerships; EU: regional integrators of FANUC/KUKA/ABB finishing cells.
- Scanning sensors: Zivid, Photoneo, Keyence, Micro-Epsilon, Pickit (EU suppliers marked by integrators).
- Choose A when: parts are weldable/finishable within a vendor's envelope; prefer a turnkey cell with safety certification and CE marking.
Path B — Build on cloud APIs
- Little of the control loop can run in the cloud (latency, safety). Cloud is useful for CAD/WPS parsing with LLMs/VLMs, fleet analytics, path-quality dashboards and offline simulation farms.
- Robot interfaces: vendor SDKs, ROS 2, OPC UA robotics companion specs; cloud simulation via NVIDIA Isaac Sim / Omniverse in EU regions.
- Choose B when: you integrate a standard robot and want custom orchestration (MES/ERP job data to cell).
Path C — Self-host (open source)
- Stack: ROS 2 with MoveIt 2, Open3D / PCL for point clouds, open segmentation models (e.g. SAM-family, PointNet++/Point Transformer variants), NVIDIA Isaac Lab / Isaac ROS (open components) and MuJoCo for simulation [3][4].
- Hardware: HW1 (Jetson Orin class or industrial PC) at the cell for inference; HW2 workstation for training/simulation.
- Effort: needs robotics, welding-process and safety expertise; certification of the integrated machine stays with the integrator.
Path D — Custom training
Worth it for a specific high-volume family of parts or exotic materials. Needs scans plus labelled seams and process outcomes (weld macro sections, NDT results, finish measurements); sim-to-real (ROBO-010) can generate synthetic seams and parts. Process-outcome data is the scarce asset.
State of the art (2026)
- Scan-to-path without teach-in is production-proven in welding and finishing cells, but within bounded part families, materials and joint types [1][2].
- Process-aware AI: models link scan data and in-process signals to weld quality; learned force-controlled sanding policies now handle curved composites.
- Foundation-model trend: vision-language-action work (ROBO-004) and LLM-assisted job setup from drawings are emerging, not yet qualified for certified welding.
- Limits and failure modes: reflective or dark surfaces degrade scans; thick multi-pass, out-of-position and highly distorting welds still need experts; heat distortion changes geometry mid-process; learned parameters do not replace a qualified welding procedure; surface finish quality is hard to measure objectively; calibration drift (tool centre point, scanner) causes systematic errors.
Shared resources & combinations
- 3D scanning + segmentation shared with ROBO-001 and ROBO-014; simulation/digital twin with ROBO-010; process parameter models with OPT-009; post-process inspection with VIS-053 and other VIS inspection cases.
- Safety functions for fenceless or collaborative cells: ROBO-016.
- Programming by demonstration for non-scannable tasks: ROBO-003.
- Edge GPU (HW1) and ROS 2 infrastructure are reusable across robot cells.
Effort, cost & KPIs
- Path A: weeks to a few months for a cell; cost dominated by robot, safety, fixtures and scanner, plus software subscription.
- Path B/C: 6-18 months of engineering for a robust production system; high integration and validation cost.
- KPIs: programming time per new part, first-pass yield, rework/scrap rate, arc-on time or cycle time, robot utilisation, consumables use (wire, abrasives), operator hours saved, weld defect rate per NDT.
People & change
Involve management (budget, risk), the production or welding lead as process owner and the welders and finishers early, ideally in the discovery workshop (GUIDE-001), because they know the seams and tolerances the AI must meet. IT and IT security join for network segmentation of the cell; legal/compliance checks machinery and CE aspects.
Works council: depends. The cell itself does not monitor people, but logs of operators, cell cameras and per-shift output data can be suited to monitor behaviour or performance, which triggers co-determination under §87(1) no. 6 BetrVG (objective suitability suffices, per BAG). In any case the works council must be informed in good time about AI-planned work processes (§90(1) no. 3); an expert may be engaged (§80(3)). Public employers involve the Personalrat. Agree up front that process data is not used for individual performance evaluation.
Data protection: scans of parts are not personal data, but cell cameras may capture staff; involve the DPO and run a DPIA only if people are systematically recorded.
AI literacy: support staff skills through short hands-on training for operators on how the system decides, when to intervene and how to report errors (GUIDE-006). Typical concern is job loss; address it by positioning the cell as relief from repetitive, hazardous work and by upskilling welders to cell supervisors and programmers (GUIDE-008). Start with a pilot group of two to three volunteers on one part family (GUIDE-009) and keep the human in control: operators approve first parts, can stop the cell and override parameters.
Risks, regulation & compliance
- EU AI Act: typically minimal risk for the process AI itself. However, AI that is a safety component of machinery falls under Annex I via the Machinery Regulation (EU) 2023/1230 (applies from 20 Jan 2027) and may be high-risk; this affects safety functions, not path planning (see ROBO-016) [5][6]. AI literacy (Art. 4) applies to staff. No Art. 5 prohibition is relevant.
- Machinery safety: CE conformity of the whole cell (Machinery Directive 2006/42/EC until Jan 2027, then Regulation 2023/1230), ISO 10218-1/-2 and ISO/TS 15066 for collaboration, risk assessment of arc, fume, UV and noise hazards.
- Welding quality: EN ISO 3834 and EN 1090 (structural steel), ISO 15614 welding procedure qualification, EN ISO 9606 welder qualification; auto-generated parameters must sit within a qualified WPS.
- GDPR: generally no personal data; operator logs or camera views of workers can create it, so apply data minimisation and works-council agreements.
- Workplace: fume extraction, ergonomics, works-council involvement when job profiles change.
- Mitigations: operator approval of first article, parameter limits from WPS, post-process inspection, logs of AI decisions.
Getting started
- Pick 5-10 representative part families and quantify programming time, rework and batch sizes.
- Check which welding/finishing standards and qualified procedures apply to them.
- Request a feasibility trial from 2-3 Path A vendors using your real parts, including hard cases.
- Plan safety (risk assessment, CE) and fixturing; involve the welding coordinator and works council.
- Track first-pass yield and programming time for the first 3 months before scaling.
Sources
- Path Robotics — autonomous welding — https://www.path-robotics.com
- GrayMatter Robotics — Scan&Sand / robotic surface finishing — https://www.graymatter-robotics.com
- MoveIt 2 motion planning — https://moveit.ai
- NVIDIA Isaac Sim / Isaac Lab documentation — https://developer.nvidia.com/isaac/sim
- Regulation (EU) 2023/1230 (Machinery Regulation) — https://eur-lex.europa.eu/eli/reg/2023/1230/oj
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Open3D — https://www.open3d.org