Short answer
Decanting is the manual step of opening inbound cases and moving the units into totes that automated storage systems can accept. Data for it is mostly video of workers at induction stations plus case and item records. It can show the sequence and variation, but a humanoid still needs robot-specific action data to learn cutting, grasping and tote loading.
Why decanting is a distinct task
Automated storage and picking systems work on standard containers, not on the mixed cases suppliers send. Someone has to open a case, check what is inside, move the units into a tote (sometimes dividing it into compartments) and release the tote to the system. Buyers should check whether it is still performed by people at their target sites, which would make it a candidate for humanoid or mobile-manipulator pilots.
Order picking draws much of the cost attention; it is described as the most labour-intensive and costly warehouse activity (de Koster et al., 2007). Decanting feeds that process, and its quality sets what picking later encounters. A unit placed in the wrong tote compartment or damaged during opening becomes a downstream exception. For nearby tasks, see receiving and inbound checks and tote and container handling.
Task anatomy
A decant cycle usually breaks into these steps. Names vary by site, so treat this as a template to align against your own taxonomy.
- Case presentation. A case arrives by conveyor, cart or pallet and is placed at the station.
- Identification. The worker scans the case label, or the system reads it, to link units to a SKU and quantity.
- Opening. Tape cut with a blade or a case cutter, flaps folded, or a sleeve removed. Tool use and blade control are the highest-risk part.
- Inspection. Check for damage, wrong item, expiry or count mismatch.
- Unit transfer. Grasp units singly or in layers and place them into the tote, respecting orientation and tote-compartment rules.
- Confirmation. Scan or confirm quantity, then release the tote.
- Waste and exceptions. Flatten cases, dispose of dunnage, route damaged units to a separate flow.
Failure modes worth capturing include crushed or collapsed case walls, shrink-wrap that tears unevenly, bagged units that stick together, units that exceed tote height, and items that resist standard compartments. Each is a distinct visual and physical event.
What data helps each part
| Task part | Helpful data | Notes |
|---|---|---|
| Case identification | Images of case labels from the station camera; case master data | Label placement and print quality vary |
| Opening | Close-up video of the cutting action from a hand-level view | Needs hands and blade visible; safety review required before any recording |
| Unit identification | Item images with SKU labels; see product catalog data | Catalog images rarely match packed-in-case appearance |
| Transfer into tote | Overhead and side video of the tote; tote and compartment IDs | Shows packing density and orientation choices |
| Exceptions | Annotated clips of damaged cases, count mismatches | Rare in routine footage |
| Cycle timing | Station scan events, tote release time | Useful as weak labels, see pick confirmation as labels |
Station camera footage already exists in many automated fulfillment centers, often for safety or quality purposes. If the recordings were not designed for training, expect fixed overhead views, limited close-ups of the hands and no calibration records.
What the data can support, and what it cannot
Human decanting video can support task and step understanding, visual pretraining on cases, totes and goods, evaluation of step segmentation, and cycle-time baselines. If hands are tracked in 3D, published work on egocentric human data suggests it can be combined with robot data for co-training; the human demonstration video guide summarizes those results.
It cannot provide robot joint states, commanded actions, grip force or cutting force. It does not show how a humanoid's hand should grasp a layer of bagged units, and a worker's reach and wrist freedom differ from a robot's. See the human-to-humanoid embodiment gap. Tool-use steps such as cutting tape need either robot teleoperation data or an instrumented rig.
Sourcing routes
Existing assets. Automated fulfillment centers and the 3PLs that run them may hold station camera footage and WMS event logs. Rights are complicated: inbound goods belong to clients, as covered in client-owned inventory rights, and workers are visible.
New collection. Because the station layout, camera positions and hand visibility are rarely suited to training, a buyer may prefer a designed recording: fixed overhead and side cameras plus head-mounted or wrist views, a task script covering your case types, and written worker consent. Costs include site access, safety review and reset time between cycles. Whether this is feasible depends on the site and is discussed case by case.
Specialist capture. If you need force, tool pose or robot-frame data, ask providers who operate teleoperation or instrumented-glove rigs. That is a different supplier type from the facility itself.
Illustrative recording specification
# Illustrative only - synthetic values the buyer would set
task: decant_case_to_tote
sites: site_017, site_021
case_classes: [rigid_carton, shrink_bundle, bagged_units, mixed_inner_packs]
cameras:
overhead: 1 fixed, tote fully visible
side: 1 fixed
hands: 1 head-mounted or 2 wrist (optional)
sync: shared clock across cameras; scan events logged with timestamps
labels: step_segments, case_class, sku_id, exception_type, outcome
exceptions_target: include at least a stated share of damaged or mismatched cases (buyer-defined)
consent: operator notice and written consent, pseudonymous operator_A12
Limitations and additional preparation
Station video from one facility reflects one tote design, one conveyor layout and one SKU mix. A model trained on it may not generalize to another site's tote geometry. Sensor sync, camera calibration and hand visibility are rarely documented in existing footage. Cutting is a safety-sensitive step: recording should follow the site's own safety rules, and the footage should not be read as approved method for a robot. Hold out whole sites for evaluation rather than random clips, so a test measures transfer.
Evaluation pointers
Define success per case class, not as one average. A decant of rigid cartons and a decant of bagged units are different problems, and a pooled success rate can hide failure on the harder class. Useful measures for a decanting trial include units placed per case without damage, correct tote or compartment assignment, handling of mismatched counts, and recovery after a stuck or collapsed case. Footage that records these outcomes lets you build an evaluation set from real variation instead of a single clean path.
If the same site hosts later picking from the totes, ask for linkage between tote IDs and downstream pick events, so decant quality can be traced to later exceptions. That link is a records question as much as a video one.
Neighboring tasks include goods-to-person station tasks downstream and box erecting and carton folding for the empty cases.
Questions to ask suppliers
- Which case types and unit packaging forms appear in the footage, and in what proportions?
- Are hands, cutting tools and the open case visible, or only the station from overhead?
- Which camera angles exist, and can you share frame rate and resolution per camera?
- Do you hold scan or WMS events with timestamps that can be aligned to the video?
- Are damaged cases, short counts and wrong items included or edited out?
- Who owns the goods shown, and have clients agreed to this use?
- What notice did decant operators receive, and can you provide pseudonymous operator IDs?
- Can the site support a new recording with a scripted case mix, and under what safety rules?
- Are tote and compartment types documented with dimensions?
Discuss your requirements
If you are scoping decanting or induction data, describe the case types, stations and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources; it checks the data and the supplier's licensing permissions before anything is contracted, and a request does not guarantee a match. For the wider map, start at the humanoid training data guide or the warehouse and fulfillment hub. The site's page on video recordings describes the data type, and item singulation and induction covers the neighboring step.
Describe the workplace demonstration data your robotics project needs
Sources
- Design and control of warehouse order picking: A literature review — order picking as the most labour-intensive and costly warehouse activity.