Skip to main content
Dogeartech

REVA University Uses YOLOv8 and RoboDK to Adapt Robotic Gluing Paths

A UR3, RealSense camera and YOLOv8 vision system adapted glue paths to part orientation without external markers or laser scanners.

REVA University Uses YOLOv8 and RoboDK to Adapt Robotic Gluing Paths

A camera guided a Universal Robots UR3 to adjust its adhesive path to a workpiece’s position and orientation, allowing the system to apply glue without external markers or costly laser scanners. Researchers at REVA University in Bengaluru combined YOLOv8 vision, Python and RoboDK to build the system, targeting a common manufacturing problem: maintaining a consistent adhesive bead when parts don’t arrive in exactly the same pose.

The team presented the work at a conference on digital signal processing. Its researchers were Prof. Bharath S.H, Akshay Ajit Bhawar and Joshua John, from the university’s School of Electronics and Communication Engineering. The project joined machine vision with robot programming to make a flexible gluing cell without relying on a more complex point cloud scanning setup.

Why automate adhesive application?

Adhesive joining is used across manufacturing, with materials including hot melt glue, cold glue and epoxy. Applying it by hand makes consistent bead size and placement difficult to maintain throughout a shift. Variation can produce defective glue patterns, scrap and added cost per part.

A robot can repeat a programmed dispensing motion, which helps keep application consistent across a production run. More accurate deposition can also reduce wasted adhesive and the number of parts rejected for application errors. Automating repetitive work may reduce workers’ exposure to solvent fumes and awkward postures, though the particular risks depend on the adhesive and process.

The REVA team used a collaborative robot, or cobot, for the application. Cobot use can allow a robot to work near people and may reduce the need for some safety fixtures. The actual safeguards still depend on the cell, task and risk assessment; the project description doesn’t provide details of its safety configuration.

YOLOv8 identifies part features

The vision system centered on an Intel RealSense camera and a YOLOv8 object detection model. The researchers trained the model using a custom dataset of component images that they labeled for features such as circular openings, cavities and reference holes. They used GPU acceleration during training and selected a model intended to run detection in real time.

YOLO, short for “You Only Look Once,” uses neural networks for computer vision tasks including object detection, classification and segmentation. In this setup, the camera image went through the trained model, which identified the workpiece and its full 360° orientation. That orientation mattered because the robot needed to locate the part’s features before applying the programmed glue curve.

Feature-based detection offers a different route from point cloud scanning. The researchers describe point cloud methods as potentially difficult to deploy because of instability, cost and dependence on complex laser cameras. Their system instead used a camera, trained image model and software transformation to adapt a known path to a detected part.

From image coordinates to robot motion

After YOLOv8 detected the part, the system represented its pose as a 4×4 transformation matrix. Through the RoboDK Python API, that transform was applied to the reference frame and glue curve in the robot program. The result was a path adjusted to the detected workpiece pose, stored as a curve-follow project in RoboDK.

RoboDK post-processors converted the path into joint commands for the UR3. The commands were then streamed to the robot in real time over Ethernet. This workflow linked camera detection to the robot’s motion without requiring the researchers to build the application around ROS, a framework commonly used in robotics research. The team’s choice reflects a practical design trade-off: fewer software layers can make a system easier to manage, although the project does not provide a direct performance comparison between RoboDK and ROS.

For dispensing, the setup used a pneumatic syringe with a 0.3 mm nozzle. The dispenser was designed to be operated by a foot pedal, so the researchers wired the pedal to a robot digital output and triggered it from the RoboDK program. Coordinating the glue signal with the robot path allowed the dispensing action to follow the programmed trajectory.

Results and next steps

The team reports that the system achieved part alignment, path adaptation and synchronized glue deposition in real time, with minimal error rates and deviation. No numerical error measurements or detailed test conditions are provided, so the result is best read as a demonstration of the approach rather than a quantified production benchmark.

The key engineering point is the connection between vision and dispensing: the robot can adapt a planned curve to the detected orientation of a part without relying on external markers. For applications where workpieces vary in placement, that could reduce the need for precise manual positioning while retaining a defined adhesive path. Production use would still depend on factors such as camera setup, model performance across the full range of parts and the dispensing material’s behavior.

The researchers plan to extend the system to multi-part assemblies, use deep learning with more complex geometries, and explore related processes including sealing, welding and coating. Those applications would require their own process validation, especially where tool alignment and trajectory control affect joint quality. The next test is whether the vision-guided approach can keep that alignment as part shapes and approach angles become more demanding.

Source: Alex Owen-Hill · robodk.com

Filed alongside