Box Erecting and Carton Folding Demonstrations

The deployment problem

Carton forming looks simple and is hard for robots. A cardboard blank starts as a flat, springy sheet. Handled correctly, it passes through intermediate shapes whose geometry is hard to model: bent flaps, a half-open sleeve, bottom flaps interlocking before tape. The end state is rigid, but the task is deformable until the last crease.

For a humanoid program that wants to support packing, erecting is the upstream step. Without formed cartons there is nothing to pack into. The task is also an example of deformable-to-rigid transition: perception must recognize shapes that appear in no CAD library, and manipulation must apply force at creases without crushing the board. This page is limited to the folding task; the packing step itself is covered in packing demonstrations, and breakdown of cartons at the other end is in cardboard breakdown and baling.

Task anatomy

What this data can support, and what it cannot

Can support. Recognition of blank, partial and finished carton states; segmentation of flaps and creases under occlusion by hands; step and substep segmentation; counting and cycle timing; detection of malformed cartons. Published deformable-object work indicates the kind of challenge involved: ZeroWaste targets highly deformable, often translucent objects in extreme clutter on a recycling sorting line (Bashkirova et al., 2021). Cardboard in a waste stream and cardboard in a packing station differ, but both stress segmentation of non-rigid, crumpled or occluded objects.

Cannot support. Video does not show how hard a worker presses a crease, how the board resists, or how fingers and thumbs sequence contact. Hands often hide the flap being folded, so the critical state is the one least visible. Human hands, with their dexterity and tactile feedback, are not a humanoid's hands; see the embodiment gap and when video is not enough. The data contains no robot joint states or commanded actions. Published human-video work in robotics generally reports human data as helping robot data, not replacing it, as summarized in what human demonstration video can and cannot do.

Existing footage and commissioned recording

Existing assets are pack-station and case-erecting cameras at fulfillment sites, and line footage at co-packers. They show real production pace and real material variation, such as blanks from different suppliers or humidity-affected board, but viewpoint is chosen for the line, not for hand detail. Commissioned recording can set up close, calibrated views, including wrist or head cameras, and ask workers to narrate or repeat difficult folds. It costs more and needs consent and site access. The trade-off is described in existing footage vs commissioned recording. Discuss with SourceX whether a commissioned recording is feasible if existing footage does not fit; see the sourcing route guide.

Illustrative collection spec

# Illustrative only - synthetic values a buyer would set
task: erect_carton_from_flat_blank
carton_types: [regular_slotted, auto_bottom, tray]   # list by style, not by customer
sizes: 4 size classes, stated in mm
views:
  - overhead: fixed, whole workspace
  - front_close: hand and flap detail
  - egocentric: optional, head or chest mount
frame_rate: constant, stated; shutter stated
labels:
  step: [pick_blank, open, fold_minor, fold_major, secure, check, place]
  state_per_second: [flat, sleeve, flaps_partial, formed, malformed]
  outcome: ok | refold | discard
metadata: [site_id, station_id, board_supplier_class, tape_type, operator_id_pseudonym]
excluded: [fold_force, joint_state]
holdout: by site and carton style

Quality and acceptance criteria

• Hand and flap visibility. In a sample of clips, count the frames where the folding flap is visible and where it is hidden by hands.
• Carton diversity. Styles, sizes, board colors and print on the board.
• Material variation. Different board suppliers, stiffness and humidity affect springback; ask for metadata.
• Failure coverage. Re-folds, tears and malformed cartons should be present with labels.
• Pace. Real production pace may blur frames; ask for shutter details.
• Held-out set. Hold out carton styles and sites for a generalization claim.

Supplier types

3PLs and fulfillment centers that erect cartons for e-commerce, co-packers that form cases for packaged goods, and distributors with in-house kitting. Cardboard itself is not usually client-confidential, but printed boxes, product names and workers may be. See 3PL providers as warehouse data suppliers and value-added services tasks.

Limitations and additional preparation

Footage needs step and state labels, usually by annotators, and may need close views that existing cameras lack. If a station uses a machine to erect most cartons, the manual remainder is a skewed sample. Gloves and sleeves obscure fingers. Redaction of faces is straightforward; redaction of printed branding can remove information about material. A buyer who wants a folding policy needs robot or teleoperated data from the target hand design, since the contact strategy of a human hand will not transfer directly.

What to look for in a sample

Ask for 20 random erecting clips and watch them with a pen. Note how many show the whole sequence from blank to formed carton, how often the worker's body hides the hands, and whether the same carton style repeats all day. Mark any clip where the worker recovers from a problem, such as a flap that springs back or a bottom that will not lock; recoveries are rare in clean footage and valuable for a robot team.

Check the lighting too. Brown cardboard under warm light on a tan table loses edges, and edges are what a flap detector needs. If a station tapes the bottom with a machine, the manual sequence ends earlier than the full task, so confirm which steps are performed by people. Finally, confirm that frame rate is constant: variable frame rate complicates any timing label and any hand tracking. These checks cost an hour and can save a purchase.

Questions to ask suppliers

1. Which carton styles and sizes are erected by hand, and which by machine?
2. What cameras cover the station, and at what distance and angle?
3. Do clips show the flaps being folded, or mostly the worker's torso?
4. Are re-folds, tears and discarded blanks recorded?
5. Is the board material or supplier recorded?
6. What frame rate and shutter settings are used?
7. Are workers identifiable, and what notice applied?
8. Is printed branding on cartons owned by clients who must approve licensing?
9. Would a commissioned close-view recording be possible at the site?

Discuss your requirements

If you can describe carton styles, views and labels, describe the requirement to SourceX. SourceX reviews it, follows up about potential sources among US businesses, checks the data and licensing permissions, and coordinates the license and delivery; any new recording would be scoped case by case. A request does not guarantee a match. Start at the humanoid training data guide or the warehouse and fulfillment hub. Related catalog context: first-person video of skilled manual work.

Describe the workplace demonstration data your robotics project needs

Sources

• ZeroWaste Dataset: Towards Deformable Object Segmentation in Cluttered Scenes — deformable object segmentation in clutter.