Parcel Sortation Hand Tasks

Task anatomy

Manual sortation looks simple from a distance and is varied up close. A worker takes a parcel from a conveyor, chute or cage, finds the label, reads or scans it, maps the destination code to a location, carries or throws the parcel to that location and places it, often while the next parcel is already arriving. The steps repeat many times per shift, which makes the task attractive for automation and poor for careful demonstration.

The object stream is the difficulty. Parcels range from rigid cartons to soft polybags, envelopes and tubes. Labels sit at any orientation, may be creased or under tape, and sometimes face down. Failure modes include a misread label, a parcel that slides off a pile, a bag that will not fit the opening, jams at a chute and damaged or leaking packages that need to be set aside. Each failure is a case a robot must handle and a dataset must contain.

This page covers sourcing sortation data. For related warehouse tasks, see item singulation and induction, the tote and container handling page and the warehouse cluster. The overall map is the humanoid data guide.

Which data helps which part

Published warehouse datasets show the structure object-centric data can take. ARMBench was collected in an Amazon warehouse pick-and-place workcell and provides images, videos and metadata for more than 235K pick-and-place activities on more than 190K unique objects, with tasks including object segmentation in clutter, object identification and defect detection (Mitash et al., 2023). It records a robotic workcell, not manual sortation, so it is a reference for what annotated pick data can contain, not a substitute for sortation video.

Scan events can tie a label to a unit. The GS1 Serial Shipping Container Code is an 18-digit key that uniquely identifies a logistic unit such as a case, pallet or parcel (GS1); where a carrier uses it, linking each clip to a unit key lets you join video to scan history without storing the label image.

What this data can support

• Perception. Parcel segmentation, pose, deformable-package handling and label localization.
• Step understanding. Segmenting induct, scan, decide and place from video, with language descriptions.
• Evaluation. Comparing a system's sort decisions and error cases against human operators on held-out stations or shifts.
• Flow statistics. Rates of exception types and dwell times from logs, for scoping a pilot.

Published work on learning from human video reports that it helps, and does not replace, robot data; see what human demonstration video can and cannot do.

What this data cannot support

• Robot actions. Manual video has no joint states, commanded motions or contact forces. A humanoid needs robot-specific observations and actions from teleoperation, a robot or an instrumented device.
• Embodiment match. People throw, flip and slide parcels with two hands, trunk motion and speed that a robot may not reproduce. See the embodiment gap.
• Label content. Shipping labels contain names, addresses and phone numbers. Blurring them removes information a label-reading model needs, which forces a trade between privacy handling and training value.
• Rare cases. Damaged, leaking or hazardous items are scarce and often handled away from cameras.
• Pace. Production-speed video has motion blur and occlusion; hands and labels may be hidden at the critical moment.

Illustrative package: sortation clips with scan linkage

# Illustrative only - synthetic values, not an available dataset
package: parcel_sort_manual_v0
site: delivery_station_012
video:
  stations: [induct_A, sort_wall_B]
  camera: {mount: overhead_fixed, fps: 30, resolution: 1920x1080}
  egocentric_subset: {mount: chest, workers: 6, consent_ref: consent_set_3}
clips: {segmentation: per_parcel, ids: clip_000001..clip_000480}
labels:
  step_boundaries: [pickup, scan, decide, place]
  parcel_type: [carton, polybag, envelope, tube, other]
  outcome: [placed_ok, rescanned, dropped, set_aside]
linkage: {unit_key_hash: salted_hash, scan_events: events_000.csv}
redaction: {label_text: blurred, faces: blurred, notes: "label-reading value reduced"}
excluded: [robot_actions, forces, label_contents]
rights: {licensor: operator_X, client_data_cleared: partial}

Sourcing and permissions

Carriers, delivery stations, sort centers and third-party logistics operators are the holders. Existing station footage is cheapest but was recorded for security or process control, so viewpoint and labels need checking. Commissioned recording can set camera positions, include egocentric views and add scripted exceptions, at the price of time, consent and site access. See existing footage versus commissioned recording and 3PL providers as data suppliers.

Rights are high-risk here: the parcels belong to shippers and recipients, labels show personal data, and workers are identifiable. Ask whether clients have consented to recordings, whether contracts restrict reuse, and how worker consent was handled; see worker consent and identifiable people. Blurring labels or faces does not by itself settle whether the data may be licensed. This is not legal advice.

Quality and acceptance checks

Buyers set the thresholds. Checks worth writing into an acceptance test include the following.

• Label visibility. Sample clips and count how often a label is legible at the moment of the scan or read. If blurring has removed it, record that.
• Hand and parcel visibility. Check the share of frames in which the worker's hands and the parcel are both visible at pickup and at placement.
• Parcel diversity. Tabulate parcel types, sizes and weights, and compare them with the stream the robot will see. A dataset dominated by rigid cartons says little about polybags.
• Error retention. Check that rescans, drops and misdirected parcels remain. Related material is in pick errors and mispicks.
• Time alignment. Verify that scan timestamps and video timestamps agree within a tolerance you set, and that the offset is documented. See barcode scanning workflows.
• Held-out design. Split by station, shift or site so that evaluation parcels, workers and layouts are not seen in training.

For item-level images in clutter rather than parcel streams, bin-picking image datasets cover a closely related need.

Questions to ask suppliers

1. Which sort operations are covered (induction, wall, bag, cage, chute), and how many stations and shifts?
2. What camera positions and resolutions were used, and are labels legible in frame?
3. What parcel mix is represented, including polybags, envelopes and irregular items?
4. Are exceptions and failures kept, or were they edited out?
5. Can each clip be linked to a unit key or scan record without exposing label contents?
6. What was redacted, and does redaction obscure labels, hands or destination chutes needed for training?
7. What notice and consent did workers receive, and can the supplier evidence it?
8. Do carrier or shipper contracts restrict recording or reuse of footage?
9. Can sort plans and scan logs be supplied alongside the video?

Discuss your requirements

If your project involves parcel handling, describe the sort operation, modalities and permitted uses to SourceX. SourceX looks for US businesses that 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, and nothing is contracted until a supplier agrees. The catalog entry for supply chain and logistics data describes a related record type.

Describe the workplace demonstration data your robotics project needs

Sources

• ARMBench: An Object-centric Benchmark Dataset for Robotic Manipulation — Published warehouse pick-and-place benchmark with images, videos and metadata, showing how object-centric warehouse data is structured.
• Serial Shipping Container Code (SSCC) — GS1 key that identifies a logistic unit such as a parcel, used to link a label to a unit.