Postal Flats and Letter-Tray Handling: Thin-Item Data for Robots

What is distinctive about mail

Parcels, totes and cartons are boxes with some give. Letters and flats are the reverse, a close cousin of deformable object handling: thin, light, and prone to bending, sticking together and sliding off a stack. A letter tray holds a vertical row of envelopes pressed against each other; a flats tray holds larger, floppier pieces. Moving them means gripping a slim edge, fanning or jogging a stack, and putting it into a tray without folding the corners. These behaviors set this page apart from parcel sortation hand tasks, which deals with parcels, and from item singulation and induction, where single items are separated for a sorter.

Not every site works the same way. Large processing plants often rely on machines for much of the sorting, with people loading and unloading trays, clearing jams and handling exceptions. Corporate and institutional mailrooms sort by recipient by hand. The right supplier type depends on whether you want tray handling, hand sorting, or exception work.

Task anatomy

1. Tray handling. Trays are lifted from carts or containers, stacked, strapped or sleeved, and moved.
2. Feeding and jogging. A worker loads a stack of mail into a feed point, squaring edges so the pieces align.
3. Hand sorting. In a mailroom, a worker reads an address or department and drops the piece into a case or slot.
4. Exception handling. Jammed, torn, oversized or misdirected pieces are pulled for manual treatment.
5. Dispatch. Filled trays are labeled and loaded onto carts or racks.

Reading destination information is a perception task covered more by barcode scanning workflows and, if relevant, by sources outside this page.

A cautious use of ARMBench

The one source cited here is a warehouse manipulation benchmark. ARMBench was collected in an Amazon warehouse pick-and-place workcell and offers images, videos and metadata for more than 235,000 pick-and-place activities on more than 190,000 unique objects, with tasks including segmentation in clutter, object identification and defect detection (Mitash et al., 2023). That is a retail-item workcell, not a mail operation, and it does not feature the thin flats and letters this page is about. Use it as a reference for how a manipulation benchmark packages images, videos and metadata, and how it might define tasks such as defect detection; do not infer anything about mail from its numbers or results.

What data can support

• Tray and stack perception. Overhead and side views of letter trays, flats trays, carts and containers support detection of trays, fill levels and stack alignment.
• Step structure. Segmented sequences of tray moves, feeding and exception handling support task understanding.
• Exception recognition. Labeled examples of jams, folded pieces and torn items help evaluate whether a system flags problems.
• Process records. Throughput logs and tray counts can indicate where steps begin and end; they are weak labels for video.

Limits: what this data cannot support

None of these assets contains robot joint states, commanded actions or contact forces. Training a humanoid to peel a single flat from a stack or square a tray needs robot or teleoperation data, and probably tactile sensing for thin items; footage, SOPs or recognition data alone do not provide it. See human demonstrations vs teleoperated robot demonstrations and tactile sensing data.

Task-specific limits:

• Edges and thickness are hard to see. Thin pieces on a belt or in a tray present little visual detail, and grip is by feel.
• Machines hide the hand work. Where machines sort, video mostly shows loading and clearing, not the item-level manipulation.
• Content is private. Mail contains personal and commercial information. Footage with legible addresses or contents is generally sensitive, so what remains after masking may be unusable.
• Material variability. Paper stock, windows, plastic wrap and magazines behave differently; coverage may be narrow.
• Embodiment. A humanoid hand's thin-edge grasp differs from a human thumb-and-finger pinch; see the human-to-humanoid embodiment gap.

Illustrative package: tray handling and flats feeding

# Illustrative only - synthetic values, not an available dataset
package: mail_tray_v0
source: processing-plant and mailroom station video, 2 sites (site_M1, site_M2)
clips:
  - clip_id: ml_0078
    task: feed_flats_stack          # tray_move | feed_letters | hand_sort | clear_jam
    item_class: large_flat | letter | magazine
    stack_state: aligned | skewed | stuck_together
    outcome: fed | jam | pulled_for_manual
mail_content: synthetic or de-identified test pieces, no legible addresses
views: [station_overhead, side_close]
annotations: [step_segments, stack_state, exception_flag]
excluded: addresses, recipient names, worker identity, forces, robot actions

A pilot proposing "an arbitrary 200 clips per task" would use arbitrary inputs, not minimums. Requirements depend on architecture, task, embodiment and intended use.

Quality and acceptance criteria

• Privacy of content verified. Spot-check frames for legible addresses and personal information.
• Close views present. At least some clips must show the contact between fingers and the edge of a piece.
• Item-class balance. Letters, flats and magazines should each be represented.
• Exception sampling. Include jam and misfeed events with causes.
• Hold-out. Keep out sites or item classes for evaluation; see training vs evaluation datasets.

Who may hold it

Mail processors, commercial presort and mail-service providers, and large corporate or institutional mailrooms may hold station video and throughput records. Because real mail is private, commissioned recording with test mail may be a more practical route than licensing archived footage; compare in existing footage vs commissioned recording. Tray handling also resembles tote and container handling.

Permissions and provenance

Mail is private to senders and recipients, and a mailroom's footage may show addresses, return labels or documents. Ask whether any frame shows legible mail content, how that was handled, and who can authorize licensing. Workers are also on camera. Mail may also be subject to specific legal protections and to confidentiality terms with mailers, so ask what restrictions apply to recording, retaining or licensing images of mail, and whether test mail can replace real items. Masking does not by itself settle permission questions. Evaluating a robot's tray handling is separate from monitoring individual workers' sorting rates, and licenses should say so. See worker consent and identifiable people. This is not legal advice.

Questions to ask suppliers

1. Which operations are recorded: tray handling, feeding, hand sorting, exceptions?
2. How much of the footage shows hand-level contact with mail versus machines?
3. Is any mail content legible in the frames, and how was it removed?
4. Can I get close-up views of grasping a stack or piece?
5. Which item classes and tray types appear?
6. Are jams, misfeeds and damaged pieces included with reasons?
7. Could recording be done with test mail instead of real mail?
8. What notice did workers receive, and who can license?
9. Which legal or contractual restrictions on handling and recording mail apply at the site, and who has reviewed them?

Discuss your requirements

For mail-handling work, describe the operations, item classes and permitted uses to SourceX. SourceX looks for US businesses that may hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match; SourceX follows up about potential sources. Start from the humanoid training data guide or the warehouse hub, and see supply chain and logistics records.

Discuss your humanoid training-data requirements with SourceX

Sources

• ARMBench: An Object-centric Benchmark Dataset for Robotic Manipulation — warehouse pick-and-place benchmark used here only as a capture, labeling and format reference; it is not a mail dataset.