Researchers tested whether a digital twin and cobot could let an operator remotely program and supervise small-batch metal AM finishing without being in the cell.
A research team from the University of North Carolina at Charlotte and the University of Oulu in Finland has developed a cyber-physical system that allows an operator to program, supervise and intermittently teleoperate a robot to finish metal additive-manufactured parts without being physically present at the manufacturing cell.
The system combines a digital twin, extended reality (XR), a collaborative robot and a centrifugal finishing system. The researchers say their laboratory prototype demonstrates that the approach can support remote programming and supervision of finishing operations for small-batch and one-off production.
“Production itself is highly automated,” the authors write, “but the finishing step that follows remains largely manual and time-consuming.”
The research, published Sept. 8, was conducted by Jose Outeiro and Jia Holt of UNC Charlotte’s Digital Engineering for Advanced Manufacturing Laboratory, along with Tero Kaarlela and Khalil Chakal of the University of Oulu in Finland.
Moving the operator away from the cell
The researchers built the system around a Universal Robots UR30 six-axis collaborative robot and a Mass Finishing RF-50 centrifugal disk finisher. The robot is equipped with a three-finger pneumatic gripper to move an additively manufactured part in and out of the finishing station.
The physical finishing environment presents a challenge for remote supervision. The researchers measured a sound pressure level of approximately 85 dB(A) during finishing. The process also generates vibration, dust and limited visibility.
Rather than attempting to eliminate the human operator, the researchers designed the system to move that person away from the physical process.
The digital twin was developed in Unity and delivered as a WebGL application. It operates on a desktop or through WebXR-compatible immersive devices. The twin communicates with the physical robot using Message Queuing Telemetry Transport (MQTT) protocol.
Within the virtual environment, an operator can manipulate the robot, control its joints and linear movements, operate the gripper, record waypoints and simulate and execute programmed movements. The virtual environment also allows the operator to inspect the robot, part and finishing bowl from viewpoints that may not be feasible in the physical work area.
The system doesn’t simply send every operator command directly to the robot. The digital twin checks proposed robot poses against joint and collision constraints before they are transmitted. The robot controller provides another layer of protection. It independently checks incoming commands against configured joint limits and Cartesian safety planes.
Th cell has a light curtain to stop remote commands when a person enters the safeguarded space and was configured according to the applicable ISO 10218 requirements, with ISO/TS 15066 requirements applying during collaborative operation.
The twin carries process information, not just robot data
One of the researchers’ central questions was whether information about the robot itself would be enough for an operator to supervise the finishing process remotely.
The answer was no.
The robot’s joint and Cartesian states are exchanged automatically between the physical system and the digital twin. But the finishing process itself is not yet fully instrumented.
The researchers currently measure the surface topography of the part offline at fixed intervals using 3D coherent scanning interferometry. The resulting information includes a height map, aerial surface roughness and surface uniformity.
The operator can use that data to determine whether the part needs to be repositioned or rotated in the finishing media, or whether the target surface roughness has been reached.
The researchers emphasize that the current digital twin therefore represents the robot’s kinematics more completely than it represents the physical finishing process, with no force, vibration, or acoustic signals gathered from the bowl or the part.
The surface-topography data is the only process information transferred to the twin. Fully instrumenting the process is identified as a primary area for future development.
The researchers validated the system with a pick-and-place and finishing task. The operator remotely programmed the robot to pick up an AM part, place it in the centrifugal bowl, reposition it during finishing and remove the finished part.
The digital twin only permitted robot positions to be programmed when the selected end positions were validated as collision-free. The system successfully synchronized the digital and physical robots, programmed the robot remotely and executed the resulting program. The maximum reported steady-state joint synchronization error was 0.12 degrees.
But the larger limitation was communication latency, with a minimum round-trip latency of 431 milliseconds and an average of 563 milliseconds. This was caused by the polling interval in the robot controller rather than the MQTT itself.
The resulting system cannot support continuous motion streaming or direct manual control. Instead, the operator works in what the researchers called a “move-and-wait regime.” A pose is established, verified and executed rather than continuously streamed as an operator guides the robot.
The authors say this is sufficient for the type of supervisory programming demonstrated in the study, but it prevents an operator from making rapid reactive corrections while the robot is moving, making the measured performance sufficient for offline programming and intermittent teleoperation, but not for continuous teleoperation.
They also identify a relatively straightforward technical target for improving the system: replacing the controller’s polling loop with an event-driven interface.
Human part of the control loop
The approach was not an attempt to make the finishing process completely autonomous. Fully autonomous finishing may not be economically feasible for small production batches. Human operators can adapt finishing strategies based on part geometry and process behavior, while the robot performs the physical work.
The digital twin therefore becomes the interface through which the operator makes those decisions remotely. The authors describe the result as human decision-making combined with robotic execution. The robot’s position alone does not tell the operator whether the part is finished; surface information determines what should happen next.
The study’s second research question therefore produced a similarly direct answer: robot kinematic state is not sufficient for remote supervision of the finishing process. Process-level information is also required.
From lab to manufacturing
The system was demonstrated using industrial hardware, but the implementation remains a laboratory prototype and is “not yet available for production use.” The next steps will focus on making the system more capable and reducing the amount of manual intervention required.
To accomplish this, they plan to add force, acoustic and vibration sensing to the finishing process and introduce in-situ metrology so that surface topography can be measured without removing the part from the gripper. They also plan to replace the controller polling loop to reduce latency.
For now, the work demonstrates a narrower proposition than fully autonomous digital manufacturing: an operator can remain responsible for process decisions while using a digital twin and XR interface to program and supervise a finishing robot from outside the physical cell.
The results indicate the approach can work for intermittent remote control. They also identify exactly where the current system falls short: continuous teleoperation, comprehensive process sensing and validation with non-expert users.
