Short answer
Deflashing and trimming data is video of workers removing excess material from molded parts with knives, shears, files or abrasives, plus part-removal and sorting steps. It can support step segmentation and outcome labeling. It cannot show cutting force or tool commands and cannot train robot control alone. Cutting tools, gloves and part variety make capture and generalization harder.
The buyer's problem
After a part leaves a mold, it is rarely finished. Thin fins of material, called flash, may sit along the parting line; gates and sprues may need to be cut off; ejector marks or burrs may need attention. In some shops this work is done by hand, at a bench beside the press, by people with a blade or a small tool. It is repetitive, a little hazardous and hard to automate across many part shapes, which is why it appears on robotics lists.
A buyer wants to know whether footage of this work helps. It can help with some questions and not others. The decisions that matter, where to cut, how deep, how hard and when to stop, are made through hand and tool feel, and footage conveys little of it. This page sets out what a molder may hold, what a purpose-built recording would add and what stays out of reach.
Task anatomy
Post-mold work spans handling and trimming:
- Part removal: taking parts from the mold, ejector or conveyor; sometimes done by an in-press robot, sometimes by hand. Machine-attached versions of this are covered in machine tending.
- Gate and sprue separation: cutting or snapping parts off runners.
- Flash removal: trimming fins with a knife, scraper, shears, file or a rotary tool, or rubbing them off by hand.
- Inspection: checking edges, surfaces and dimensions visually or by feel.
- Sorting and placement: putting good parts in trays or bins, scrap in a reject container.
- Housekeeping: clearing trimmings and waste, which can affect the next cycle.
Failure modes include cutting into the part, leaving flash behind, scratching a cosmetic surface, dropping a hot or slippery part, and mixing good and rejected parts. The material matters: a rigid plastic, a soft elastomer and a rubber compound each trim differently, and the same hand motion produces different results.
What the data can support
Video from the press-side bench can support:
- Step segmentation: separating removal, trimming, inspection and sorting.
- Part state recognition: labeling parts as flashed, trimmed or rejected, and where the flash sits.
- Sequence and station design: showing how the work is laid out and what a worker moves between.
- Outcome labeling: tying a part image to an accept, rework or scrap decision.
For labeling, image datasets for industrial inspection give a sense of the structures involved. GDXray collects 19,407 X-ray images in five groups: castings, welds, baggage, natural objects and settings, with annotations as bounding boxes or labels (Mery, 2015). It is X-ray imagery for nondestructive testing and does not cover molded-part trimming, so use it only as a reference for how defect locations can be annotated.
The adjacent finishing operations, where abrasion and force dominate, are discussed in sanding, polishing and finishing, and handling steps that follow printing are covered in additive manufacturing post-processing.
What data helps each part
| Part of the task | Data that helps | Typical source |
|---|---|---|
| Removal and sorting | Overhead video with part and bin labels | Existing footage or commissioned recording |
| Trimming technique | Close view of blade, part and hands | Commissioned recording |
| Flash location | Annotated images of parts before and after | Quality records, commissioned imaging |
| Tool choice | Tool type per part and operation | Work instructions, annotations |
| Accept or scrap | Inspector decisions per part | Quality records |
| Part geometry | CAD or scans for each tooled part | Engineering files |
Molders usually have strong records on scrap and cycle counts, and weaker video. Their parts often come from customer tooling, so geometry may be restricted.
Illustrative recording package
# Illustrative only - synthetic values, not an available dataset
package: post_mold_trimming_v0
site: site_052 (synthetic)
parts: [housing_type_A, clip_type_B, seal_type_C] # synthetic part families
streams:
bench_overhead: 1080p, 30 fps
cutting_zone: close view, 60 fps, bright diffuse light, avoid glare on glossy parts
part_stills: before and after trimming, consistent lighting
labels:
steps: remove | separate_gate | trim_flash | inspect | sort
tool: knife | shears | file | deburr_tool
flash_regions: bounding boxes on the before still
outcome: accept | rework | scrap, with reason
metadata: material_family, mold_id (pseudonymous), operator_id operator_F11 (pseudonymous), glove_type
sync: shared clock or visible sync event
excluded: cutting force, tool commands, robot actions
rights: training and evaluation; customer part geometry stated separately
Limits: what this data cannot support
Trimming video does not carry the force applied to a blade, the angle at which the tool meets the part or the feedback a worker feels as flash separates. It does not show a robot hand holding a part and a cutter. Those need robot or teleoperation data, or instrumented handheld tools, in the robot's action space; footage, work instructions, quality records and image labels alone do not train control. Hands wearing cut-resistant gloves also change grip and visibility; see the human-to-humanoid embodiment gap.
More limits:
- Occlusion. The blade and the flash are tiny and hidden by fingers.
- Part variety. A dataset for three part families does not cover a fourth.
- Material state. Part temperature and moisture may vary through the shift and change how material trims.
- Safety bias. Sites may ask workers to slow down or take extra care when filmed, producing unrepresentative technique.
Permissions and provenance
Part geometry, tooling numbers and customer names may appear on parts, trays and labels. Workers are identifiable by face and hands, and bench cameras installed for other purposes may not have been intended for training use. Redaction or removing identifiers does not settle permissions, and blurring the part removes the evidence you need. Ask who owns the footage and the parts' designs and what workers were told. This is not legal advice. Cutting tools raise site safety rules for camera placement; see machine safety constraints for recording and manufacturing data confidentiality.
Questions to ask suppliers
- Which part families, materials and trimming tools are in the recordings?
- Is there a close view of the cutting zone, and how is it lit?
- Are trimming defects, over-cuts and rework preserved, or only accepted parts?
- How are accept, rework and scrap decisions recorded and by whom?
- Is part removal done by hand, by an in-press robot or by both?
- Does the footage show gloves, and what type?
- Which part geometries or customer tooling must be excluded?
- Can the molder host a new recording with your chosen viewpoints?
Discuss your requirements
If post-mold handling and trimming are on your task list, describe the part families, viewpoints and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources among US businesses, checks the data and the supplier's licensing permissions, and coordinates license and delivery. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match. Start from the humanoid training data guide or the manufacturing hub; for neighboring tasks see rework and repair operations and fixturing and part orientation.
Describe the workplace demonstration data your robotics project needs
Sources
- GDXray: The Database of X-ray Images for Nondestructive Testing — X-ray defect images and annotation format; labeling reference only.