Recognising Tens of Thousands of Plate Specs in 6 Seconds
Cut steel plate parts arrive in heaps in the heavy industry stockyard, up to 12 metres long, weighing several tonnes, in tens of thousands of specifications. This project uses AI plus 3D vision and a digital twin to recognise presence and type, then guides a robot to sort and stack. Recognition accuracy is 99.99 percent or better, a single recognise-and-sort cycle completes within 6 seconds, and sorting capacity rises several times over, subject to site acceptance results.
Tens of thousands of specs break the fixture approach
Sorting a fixed set of parts can rely on fixtures, templates or taught positions. At tens of thousands of specifications all three fail: fixtures cannot be built for every part, templates do not cover the set, and taught positions are meaningless. The system has to recognise what it sees and grasp what it recognises. Flexibility stops being a bonus and becomes the entry condition.
Point cloud for geometry, AI for which part
A high-precision 3D camera captures the point cloud and with it the shape and position; the AI model judges presence and type from the cloud. The 99.99 percent figure tests not single-image classification but generalisation across similar parts, occluded parts and parts lying at arbitrary attitudes. Heaped input is the real test set, and it is the one a supplier should be asked to demonstrate.
Grasp safety lives in the pose
Recognition is followed by the part's position and orientation in space, and the grasp pose is sent to the robot. For a part weighing tonnes, grasp point selection and robot payload limits have to be considered in the planning stage together: grasping accurately is only the beginning, grasping stably is what decides whether the system runs for years.
The digital twin is for trying things first
Planning and validating sorting for tens of thousands of specifications on site would stop the yard. The digital twin moves stockyard layout, robot motion and sorting cycle into a virtual environment, where the plan is tried before it is applied. What it saves is not algorithm time but on-site commissioning and downtime. Ask for the model's fidelity, since a twin that does not match the yard saves nothing.
Six seconds against heaped input
Heaped material means occlusion, arbitrary attitude and look-alike parts mixed together. The spatial expressiveness of the point cloud and the robustness of the model carry recognition, and a recognise-and-sort cycle inside 6 seconds is what keeps pace with continuous sorting. A slow recognition step makes the robot wait, and the capacity gain disappears.
Limits and what to verify
The scheme suits plate part sorting in construction machinery, mining machinery, steel structures and similar heavy industries. Selection should assess maximum part size and weight, container form, robot payload and count; at part sizes of several metres, 3D capture range and robot reach need their own calculation. The AI model needs samples, so a new specification is not a parameter change.
Questions before rebuilding a stockyard
How can tens of thousands of specs be recognised?
Through the generalisation of a 3D point cloud plus AI model, not a rule per specification. Sample coverage decides stability on look-alikes and occluded parts, and new specs need added training.
Can one camera capture a 12 m part?
Capture range is assessed against part size and camera layout, with segmented capture and stitching where needed. It has to be calculated together with robot reach, not on its own.
How is a stable grasp assured?
Pose computation gives the grasp point and orientation, chosen together with payload limits. Parts at tonne scale also need gripper and coordinated handling assessed.
Is the digital twin just a gimmick?
Not here. Validating tens of thousands of specifications on site would be unaffordable in downtime; task allocation and paths are proven virtually first.