Integrated AI Vision for Torch Cutting and Robotic Plate Handling in Harsh Environments

Torch cutting systems traditionally present unique challenges for automation due to the variability of raw materials, high heat, and lack of consistent part presentation. LM3 Technologies recently deployed a multi-layered inspection and handling solution combining 3D vision, AI-based object detection, and advanced robot guidance to achieve complete automation of the torch cutting process — from bin picking to final cut verification.

Application Overview

The client environment involves the processing of heavy steel plates of varying shapes and sizes, stacked in disordered configurations. Each plate requires:

1. Robotic picking and placement onto a torch cutting table

2. Datum hole localization for positional alignment

3. Adaptive cut path generation based on plate geometry

4. Post-cut inspection and dimensional validation
   

The automation goal was to enable fully unattended part handling and cutting, while maintaining high process repeatability and traceable inspection outcomes.

System Architecture

LM3’s solution integrates the following core technologies:

1. Robotic Pallet Picking with SICK PLB520

A SICK PLB520 3D vision sensor is mounted above the pallet to acquire structured light point cloud data of loosely stacked plates. The PLB520 identifies pickable targets and transmits pick coordinates to a 6-axis robot. This enables reliable autonomous part retrieval without requiring fixed orientation or spacing.

2. Datum Hole Detection via PAQi + QC Hero

Once the plate is placed on the cutting table, the PAQi vision controller utilizes AI object detectors trained in LM3’s QC Hero platform to locate critical datum holes. These detections are used to establish the part’s coordinate frame, enabling positional correction and rotation compensation prior to cutting.

The object detection models are robust to variation in part geometry, size, surface finish, and orientation — allowing for generalization across a wide range of SKUs.

3. Profile Scanning and Cut Path Adjustment with SICK Ruler

A SICK Ruler 3D profile sensor captures the edge geometry of the plate. This information is used to dynamically adjust the robot’s torch cutting path in real time, ensuring alignment with the actual part location and preventing overcuts or undercuts.

4. Torch Cut Validation via AI Inspection

After cutting, additional vision cameras perform inspection to detect common torch-related defects, including:

• Incomplete cuts or skipped sections

• Misaligned contours or offset profiles

• Excess material / dross along the cut edge

• Dimensional nonconformance relative to the programmed profile

  
  

The AI models used for defect detection were trained and validated using QC Hero, and deployed directly onto the PAQi unit for on-premise execution.

Data and Integration

The system is built around LM3’s PAQi controller, which provides:

• On-premise inference of AI models with no cloud dependence

• SQL-based data logging for image traceability and pass/fail records

• Integrated PLC communication for part routing (pass/fail binning)

• Live visualization of detections and robot commands for operator review

  
  

The vision system is synchronized with the robot controller and torch cutter via digital IO and Ethernet/IP communication, ensuring deterministic timing and traceable control flow.

Outcome and Value

By integrating robotic part handling, flexible part recognition, adaptive cut path guidance, and AI-based cut inspection, the solution delivered:

• Fully autonomous operation across part sizes, shapes, and orientations

• Real-time defect feedback and image-based traceability

• Reduced labor involvement and manual inspection

• Minimized part waste due to adaptive cut alignment

• Modular system architecture scalable to additional lines

Conclusion

This project demonstrates how LM3 Technologies combines industrial-grade hardware, AI model development, and robust system integration to deploy end-to-end vision-based automation in challenging environments. By leveraging our camera-agnostic platform (PAQi), rapid AI training pipeline (QC Hero), and deep experience in robotics and factory networking, we provide complete vision systems that outperform traditional solutions in flexibility, speed, and reliability.