12 m Plate Parts: AI Plus 3D Sorts Tons Without Operators
Plate cut parts arrive at the heavy-industry stockyard in stacks, up to 12 metres long and weighing tons, with many types and similar shapes. This cell pairs AI with 3D vision and a digital twin to identify presence and type and drive multi-robot collaborative sorting and stacking. Recognition accuracy is at least 99.99 percent, a single recognition takes within 6 seconds, the work cycle is about 7 seconds, and figures are subject to site acceptance results.
A 12 metre part needs a stitched point cloud
One field of view cannot cover a 12 metre part. The cell acquires point clouds segment by segment with a scanning approach and stitches them into a complete part cloud, and the AI model recognises the assembled whole. Stitching is more than joining data: coordinate consistency at the seams determines how reliable the type verdict is.
What 99.99 percent under occlusion actually takes
Stacking leaves point clouds incomplete, and that is normal in a stockyard. The model is trained on many occluded samples and is robust to partial point clouds, so presence and type can still be judged when part of the surface is hidden. The 99.99 percent figure is an engineering result for stacked conditions, not a laboratory number, and is subject to site acceptance results.
Vision solves half; multi-robot simulation solves the rest
Recognition and pose computation tell you where to grasp and how. Whether the grasp will hold is another matter. Clamping a ton-scale part carries high failure risk, so grasp point selection, gripper capability and robot payload must be calculated together, and grasp feasibility checked in simulation rather than on the floor.
Several robots handling ton-scale parts need task allocation and paths that avoid collision and mutual waiting. The digital twin simulates stockyard layout, robot motion and sorting cycle, running the task sequence and paths in the model before execution. Recognition, scheduling and execution are decoupled: recognition outputs type and pose, scheduling plans tasks, execution moves the part.
Safety in an unmanned area is designed, not removed
Robots and heavy parts occupy a wide area, so the safety system divides robot working zones from personnel zones and separates the risk spatially. Unmanned does not mean removing people and leaving the rest; it means redefining where people may be.
The gap between 6 and 7 seconds
A single recognition completes within 6 seconds and the full work cycle is about 7 seconds. The difference is grasp, transfer and stacking motion. It also shows the bottleneck is not recognition: to go faster, change robot coordination and paths rather than adding recognition compute.
Where the retrofit stops
The cell suits plate cut part sorting in construction machinery, mining machinery and steel structures. Selection must assess maximum part size and weight, bin type, robot payload and count. The AI model needs sample accumulation, and new types enter through additional training rather than a raised threshold.
What to settle before an unmanned project
Can a camera see a 12 metre part in one shot?
One field cannot cover it. Scanning acquisition plus point-cloud stitching produces the complete cloud, and seam consistency must be validated at proposal stage.
Does recognition survive stacked occlusion?
The model is trained on occluded samples and can recognise from partial clouds. Actual accuracy is subject to site acceptance results, and sample coverage directly affects performance.
How are ton-scale parts sorted?
High-payload robots are configured by weight, with several collaborating on grasp and transfer. The digital twin plans allocation and paths, avoids interference and verifies grasp feasibility in simulation.
What happens when a new part type arrives?
It is added through sample training and template expansion without hardware changes, provided type definition and sample collection are in place first.