Short answer
Singulation and induction data records items being separated from a pile and placed one at a time onto a conveyor or sorter. Parcel hubs and fulfillment sites hold induction-station video, jam and rejection records, and scan events. It can support item segmentation in clutter, pick-order and placement-pose modeling and fault detection, but it holds no robot actions and little contact information.
The deployment problem
Induction is where a flow of mixed items becomes a controlled stream. Someone, or something, takes an item from a bulk container or chute, orients it label-up if the scanner needs that, and places it on a belt with spacing the sorter accepts. A bad induction means a no-read, a jam or an item sent to the wrong destination.
For a humanoid or arm-based system the task has two hard parts. First, perception in clutter: items overlap, bags deform, and labels hide. Second, placement on a moving surface, with timing set by the belt. This page is about sourcing data for both. The downstream sort is covered in parcel sortation hand tasks, and chute-to-tote handling is covered in tote and container handling.
Task anatomy
| Step | What the operator or system does | Typical failure |
|---|---|---|
| Surveying the pile | Choose the item that is accessible and unobstructed | Picks buried item, shifts pile |
| Acquiring | Grasp or scoop one item | Two items lifted together |
| Orienting | Turn label or barcode toward the scanner | Label hidden, no-read |
| Gapping | Wait for belt space | Item placed too close |
| Placing | Release onto belt at belt speed | Item rolls, tumbles, stands on edge |
| Verifying | Scan or sorter reads the item | No-read, divert to reject |
| Clearing | Remove jams or oversize items | Jam not cleared before next item |
What this data can support, and what it cannot
Can support. Segmentation of items in a cluttered source container, detection of multiple-pick events, label visibility checks, belt-state recognition, and station-level step segmentation. Published warehouse datasets show what is collectible: ARMBench was collected in a pick-and-place workcell and includes images, videos and metadata for 235K+ pick-and-place activities on 190K+ unique objects, with segmentation, identification and defect-detection tasks (Mitash et al., 2023). That is a robotic workcell and not a manual induction station, but the same fields of interest recur: segmented items, outcome labels, object metadata.
Deformable items are a particular problem. ZeroWaste targets highly deformable, often translucent objects in extreme clutter on a recycling sorting line (Bashkirova et al., 2021); it is a different industry, but it shows how bag-like objects stress segmentation, and the same concern applies to polybags and soft mailers.
Cannot support. Manual induction video shows a human using two hands, wrist twists and tactile feedback that a gripper or humanoid hand will not replicate. It does not contain commanded actions, grip pressure or slip. Belt-synchronized timing is visible, but the control signal that matched hand speed to belt speed is not. The human-to-humanoid embodiment gap applies, and what human video can and cannot do sets the general limits. Existing robotic induction cells may log robot data, but those logs describe a specific cell design; see existing warehouse robot logs.
Capture and labels
Existing assets: overhead and oblique station cameras, scanner no-read logs, sorter rejection and jam records, throughput counters. New collection can add a defined rig: a top-down depth camera on the source container, a side camera on the belt, and a hand or wrist camera for operators. Useful labels, to be confirmed with a supplier:
- item masks in the source container, with occlusion flags;
- pick events: single, multiple, regrasp, drop;
- placement outcome: scan read, no-read, divert, jam;
- item class (envelope, polybag, box, tube, irregular), as coarse classes rather than customer product names;
- belt speed and gap at the moment of placement.
Illustrative event record
# Illustrative only - synthetic fields
station_id: induct_03
site_id: site_044
frame_rate: constant, stated
belt_speed_m_s: logged per minute
events:
- {event_id: e_000812, t: 2026-01-12T09:14:03.2Z, type: pick_start}
- {event_id: e_000813, t: 2026-01-12T09:14:04.1Z, type: place, gap_to_prev_mm: 410}
- {event_id: e_000814, t: 2026-01-12T09:14:04.9Z, type: scan_result, result: read}
item_class: polybag_soft
labels: [item_mask, multi_pick_flag, label_visible_flag, placement_outcome]
excluded: [grip_force, joint_state, operator_identity]
Supplier types
Parcel hubs, e-commerce fulfillment sites and 3PLs with sorters run induction stations. Camera coverage is often present for jam analysis. Items may belong to retailers or shippers, which affects rights; see client-owned inventory rights and 3PL providers as warehouse data suppliers. If existing footage lacks the viewpoint you need, discuss whether a commissioned recording is feasible, using the sourcing route guide to weigh options.
Limitations and additional preparation
- Check whether throughput pressure produced fast or partial hand motions that a robot should not imitate.
- Motion blur on the belt can reduce training value of frames; ask for exposure and shutter details.
- Records about jams or no-reads are outcome signals, not diagnoses. They need time alignment and join keys before they label a video interval.
- The station may have a fixed item mix. Test sets must be held out by item class, site and time, not by frame.
- Labels on packages can carry personal addresses; redaction is needed before sharing, and it may remove label-reading signal you want.
- Moving-belt placement needs robot-side data from the target hardware.
Choosing between existing footage and a designed recording
Existing induction video is good for the messy, high-volume reality of a hub: item variety, jams, no-reads, shift effects. It is weak where a robot team needs control: the camera sits where the supervisor needed it, frame rates vary, and the operator's hands are small in the frame. A designed recording can fix viewpoints and add depth, but it samples a smaller set of items under a protocol the site agrees to.
A practical split is to buy existing footage for perception breadth and for failure mining, then scope a limited designed set for the details your model needs, such as a top-down depth view of the source container. Keep the designed set out of any test partition you plan to reuse, and record exactly what protocol the operators followed so the set can be described honestly in a datasheet.
Questions to ask suppliers
- How many induction stations and layouts are represented?
- What item classes and sizes enter the station, and in what proportions?
- Are cameras fixed above the source container, the belt, or both?
- How are no-reads and jams logged, and with what timestamps?
- What clock source links video, scanner and sorter events?
- Are belt speed and item gap recorded?
- Do packages show recipient addresses or names, and how are those handled?
- Who owns the items and shipping data, and who can grant a license?
- Are any recordings from robotic induction cells, and who owns those logs?
Permissions and provenance
Parcels carry names, addresses and shipper information, and operators are identifiable in video. Ask what redaction is applied and whether it preserves the barcode area your model needs. See worker consent and identifiable people. Redaction does not by itself settle permission questions. This is not legal advice.
Discuss your requirements
If you know the item classes, station types and event fields you need, discuss your humanoid training-data requirements with SourceX. SourceX reviews the requirement and follows up about potential sources, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match. Begin at the humanoid training data guide or the warehouse and fulfillment hub, and see video recordings in the catalog for related formats.
Discuss your humanoid training-data requirements with SourceX
Sources
- ARMBench: An Object-centric Benchmark Dataset for Robotic Manipulation — warehouse pick-and-place dataset scope and tasks.
- ZeroWaste Dataset: Towards Deformable Object Segmentation in Cluttered Scenes — deformable object segmentation in clutter.