⚡ electrical engineering

CAD-prior-driven teaching-free robotic welding trajectory generation for complex structural components

This paper proposes a CAD-prior-driven method for teach-free robotic welding of complex structural components that integrates active perception, coordinate decoupling, and 2D profile-based feature extraction to compensate for assembly deviations and generate high-precision welding trajectories, achieving a 98.6% point cloud coverage rate and sub-millimeter accuracy in physical trials.

Original authors: dahao sun, Kepeng Chen, Haisheng Liu, Zhongren Wang, Rongli Zhao, Wen Zhang, Mengfeii Wang

Published 2026-10-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: dahao sun, Kepeng Chen, Haisheng Liu, Zhongren Wang, Rongli Zhao, Wen Zhang, Mengfeii Wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of heavy manufacturing, building massive structures like ship hulls, bridge frames, or mining equipment often involves joining thick metal plates together. For decades, the most reliable way to ensure these plates stay fused was to have a human welder guide the torch by hand, or to teach a robot exactly where to move by physically moving its arm through the path once. This "teach-and-repeat" method works well when every piece is identical and fits together perfectly. However, real-world metal parts are rarely perfect. When large plates are assembled, tiny gaps appear, and the metal can warp slightly due to heat or pressure. If a robot is blindly following a pre-programmed path based on a perfect digital drawing, it might miss the actual seam entirely, leading to a weak joint or a wasted part. To solve this, engineers have turned to vision systems, giving robots "eyes" to see the weld as they go. Yet, simply having a camera is not enough; if the camera cannot see the seam because of shadows, glare, or the robot's own body blocking the view, the system fails. The challenge lies in teaching the robot not just to look, but to know exactly where to stand and how to angle its camera to get a clear, complete picture of a complex, twisting seam before it even begins to weld.

Researchers at Hubei University of Arts and Science have developed a new approach that allows a robot to figure out the best way to see and weld a complex metal structure without any human guidance. Instead of relying on a human to show the robot the path or on a camera that just scans whatever is in front of it, the system starts with a detailed 3D digital model of the part, known as a CAD model. The researchers taught the robot to read this model like a map, identifying exactly where the metal plates touch and where the weld should theoretically go. Using this digital map as a guide, the robot plans its own movement. It calculates a smooth path for its camera to follow, ensuring the lens stays at the perfect distance and angle to avoid shadows and reflections, while also making sure the robot's arm doesn't get stuck or blocked by the part itself. This process, which the team calls "active perception," allows the robot to gather a complete and clear 3D picture of the seam, even in tight corners where traditional methods often miss data.

Once the robot has captured this clear picture, it faces another problem: the real metal part is never exactly the same as the digital map. There are small gaps and bends that the computer model doesn't know about. To fix this, the team created a method to find the true center of the weld line by looking at the data in a clever way. Instead of trying to find the seam in the messy, three-dimensional cloud of points the camera captures, the system slices the data into flat, two-dimensional cross-sections. Imagine looking at a loaf of bread; instead of trying to trace a curve through the whole loaf, you look at the flat face of a single slice. By doing this, the robot can easily spot where the two metal plates meet, filtering out the noise and errors that usually confuse computer vision. It then builds a smooth, continuous line representing the actual, physical location of the weld, correcting the original digital map to match reality.

With this corrected map in hand, the robot generates its final welding path. It translates the smooth line of the actual weld into a set of instructions for its own joints, ensuring the welding torch stays perfectly aligned with the seam as it moves. The researchers tested this system on a complex metal structure featuring both straight lines and twisting, three-dimensional curves. They found that their method captured 98.6% of the weld area, a significant improvement over traditional fixed-angle scanning, which only managed to see about 72% of the same area due to blind spots. The measurements of the weld's length were incredibly precise, with an average error of just 0.19 millimeters. When the robot actually performed the welding, it successfully created continuous, strong seams along both the straight sections and the difficult curved intersections, without stopping or making errors.

The study demonstrates that by combining a digital blueprint with smart, self-correcting vision, robots can handle the messy reality of industrial manufacturing without needing a human to teach them every move. The system proved that it could adapt to the inevitable imperfections of real-world assembly, finding the true path of the weld and executing it with high accuracy. This approach moves beyond simple automation, offering a way for machines to understand and navigate complex physical environments on their own, ensuring that the heavy structures holding up our world are built with a level of precision that was previously difficult to achieve without constant human supervision.

Technical Summary: CAD-Prior-Driven Teaching-Free Robotic Welding Trajectory Generation

Problem Statement
The paper addresses the challenges inherent in the teaching-free robotic welding of complex structural components, such as multi-curvature assemblies and irregular pipe frames. Traditional methods relying on manual teaching or offline programming struggle with frequent workpiece changes and random assembly deviations. While 3D vision offers a path to autonomy, existing systems suffer from insufficient perception stability. Specific issues include:

Visual Occlusion and Specular Reflection: Harsh industrial conditions lead to missing data or high-frequency noise in point clouds.
Limited Adaptability: Many systems rely on fixed-view scanning or manual adjustments, failing to account for the sensor's field of view, observation quality, and robot kinematic reachability.
Feature Extraction Failures: Random assembly gaps and local deformations create complex spatial deviations between theoretical CAD models and measured data. Directly fitting noisy point clouds often results in trajectory drift, leading to weld defects like lack of fusion.
Underutilized CAD Priors: Current visual tracking studies often ignore the explicit structural constraints (geometry, topology, normals) embedded in CAD models during the active pose-planning stage.

Methodology
The authors propose a comprehensive framework for active perception and trajectory generation driven by CAD priors. The methodology proceeds through four main stages:

Automatic Theoretical Weld Identification:
- The system parses the 3D CAD model to extract theoretical weld seams automatically.
- Using topological adjacency and contact-geometry constraints, it identifies contact lines between solids.
- A multi-objective evaluation function filters candidate boundaries based on reference direction projection, arc length, and consistency with adjacent surface normals to select the optimal theoretical weld line.
Active Perception Trajectory Planning:
- Local Reference Frame: A preferred local observation direction is constructed using the weld tangent and the angle-bisector of adjacent surface normals.
- Candidate Pose Generation: Candidate camera poses are generated around sampling points, ensuring the optical axis targets the local weld region.
- Quality Evaluation: A comprehensive function evaluates candidates based on five metrics: field-of-view coverage, observation distance, incidence angle, local occlusion, and robot kinematic feasibility.
- Dynamic Programming: To ensure continuity, a layered graph is constructed where transition costs account for spatial displacement, pose variation, and joint-space changes. Dynamic programming solves for the optimal imaging sequence, which is then smoothed using B-spline interpolation.
Global Registration and Coordinate Unification:
- To compensate for clamping errors, the system acquires a global coarse point cloud.
- Using hand-eye calibration and forward kinematics, data is transformed into the robot base frame, then back-mapped to the theoretical model coordinate space.
- The Iterative Closest Point (ICP) algorithm registers the actual point cloud with the theoretical CAD model to determine the rigid-body pose deviation ( $\Delta T$ ).
Deviation-Field Reconstruction and Trajectory Generation:
- Dimensionality Reduction: For each theoretical sampling point, the local 3D point cloud is projected onto a 2D normal cross-section perpendicular to the weld tangent. This transforms the 3D feature search into a 2D contour-intersection problem, suppressing noise and outliers.
- Feature Extraction: Robust line fitting is applied to the projected point sets (left and right plates) to find the intersection, representing the actual weld feature point.
- Deviation Field Reconstruction: The discrete deviation vectors are decomposed into lateral, depth, and tangential components. A regularized energy functional is minimized to reconstruct a smooth, differentiable continuous deviation field, filtering out high-frequency measurement noise.
- Source Correction: The smooth deviation field is superimposed on the theoretical weld model to generate the actual weld curve.
- Trajectory Synthesis: A moving welding coordinate system is established based on the corrected curve and dihedral-angle bisectors. Inverse kinematics converts these poses into executable robot commands.

Key Contributions

CAD-Prior-Driven Active Planning: The paper introduces a method that utilizes CAD geometric and topological priors to constrain candidate camera poses and guide continuous imaging trajectory planning, rather than relying on fixed views or manual adjustment.
2D Projection for Noise Suppression: By projecting local 3D point clouds onto 2D normal cross-sections, the method converts complex 3D feature extraction into a 2D contour intersection problem, significantly reducing the impact of specular reflection and outliers.
Regularized Deviation Field Reconstruction: Instead of directly fitting noisy data, the authors propose a regularized energy functional to reconstruct a smooth deviation field. This ensures the generated trajectory is continuous and differentiable, preventing drift caused by local visual noise.
Integrated System Validation: The study validates the entire pipeline—from CAD parsing to physical welding—demonstrating the generation of executable trajectories for complex spatial intersections without manual teaching.

Experimental Results
Experiments were conducted on a structural component featuring straight fillet welds and spatial curved intersections using a Siasun SR10C robot and a Kinect V2 sensor.

Point Cloud Coverage: The active planning strategy achieved a point cloud coverage rate of 98.6% in complex intersection regions, a significant improvement over the 72.4% achieved by traditional fixed-view scanning.
Dimensional Accuracy: The mean absolute error in point cloud length was 0.19 mm, with a maximum unidirectional coordinate error of 0.28 mm.
Feature Extraction Accuracy: The mean absolute errors for extracted feature points were 0.16 mm (X), 0.14 mm (Y), and 0.17 mm (Z).
Physical Welding: Physical trials confirmed the successful formation of continuous weld seams across both straight and spatially curved regions. The robot executed the trajectory without over-limit or singularity alarms, demonstrating the method's engineering executability.

Significance
The paper claims that its primary significance lies in transforming the analyzable geometric relationships within a workpiece model into constraints for the perception and control process, thereby eliminating the need for manual teaching and fixed experience-based perspectives. By systematically integrating CAD priors into the active visual pose-planning stage, the method addresses the root causes of perception instability (occlusion, reflection, noise) and trajectory drift. The study demonstrates that a teaching-free approach is feasible for complex structural components, providing a robust, data-driven basis for autonomous trajectory generation that adapts to assembly deviations and complex geometries.

More like this