Short answer
E-commerce fulfillment is dominated by each-picking of many small, varied items into individual orders, while B2B distribution moves cases and pallets to stores or business customers. The task mix, object sets, packaging and exceptions differ, so data from one does not stand in for the other. Specify the operation type, unit of handling and order profile before scoping any warehouse request.
Why the label "warehouse" is too coarse
A request for "warehouse data" can describe a parcel fulfillment center, a grocery distribution hub, a cross-dock or a spare-parts depot. These share buildings and some equipment but differ in what a worker lifts, how many different things pass through, and what an error costs. A model trained on one profile can meet an entirely different distribution of objects, packaging and motions at another.
The economics explain why buyers care about the picking-heavy profile. A review of warehouse order picking notes that order picking is the most labour-intensive and costly warehouse activity, with its cost estimated at up to 55% of total warehouse operating expense (de Koster et al., 2007). That figure comes from the review's abstract and predates modern automation, so it motivates attention to picking, not a forecast for any site. It also does not tell you which operation type your robot will join.
How the two operation types differ
| Dimension | E-commerce fulfillment | B2B distribution |
|---|---|---|
| Typical unit of handling | Each (single item), small multi-line orders | Case, inner pack, layer, pallet |
| Object variety | Very high; many SKUs, soft goods, polybags, irregular shapes | Lower per SKU but wide pack formats; cartons, shrink-wrapped goods |
| Order profile | Many small orders, short lines | Fewer, larger orders by store or customer |
| Dominant tasks | Each-picking, put wall, pack station, singulation | Case picking, mixed palletizing, depalletizing, trailer loading |
| Handling risk | Item damage, mis-picks, bag slip | Load stability, crushed cartons, label orientation |
| Exceptions | Unreadable barcodes, missing items, damaged goods | Mixed-SKU pallets, pallet damage, short shipments |
| Returns | High, see returns processing workflow | Lower, but chargebacks and shortages |
Real sites mix these. Many third-party logistics operators run both under one roof, which is why the 3PL providers as data suppliers page asks what share of each a site handles.
What this means for data
Each-picking data centers on grasping varied small items from totes or shelves, placing them into containers, and confirming the pick. The relevant sets are each-picking demonstrations and bin-based image datasets, and the major open resources come from picking workcells. ARMBench, for instance, records more than 235,000 pick-and-place activities on over 190,000 unique objects in a warehouse workcell (Mitash et al., 2023). It shows the object-variety scale of each-picking in one operation, and it is a robot-workcell dataset, not a humanoid dataset.
Case and pallet data centers on lifting cartons of different sizes, building stable loads, reading labels and handling shrink wrap. The relevant pages are case picking and mixed palletizing and depalletizing mixed cases. Weight, compliance of cartons and stacking logic matter more than fine finger-level manipulation.
Shared layers are the same in both: the barcode and label workflows, the warehouse management system records that tell a worker what to pick, and the exception handling. These can be sourced once and reused, subject to each site's permissions.
Illustrative scoping brief
# Illustrative only - values are synthetic
request_id: wh_scope_v0
target_deployment:
operation_type: ecommerce_fulfillment # or b2b_distribution, or mixed
handling_unit: each # each | case | layer | pallet
object_mix: [rigid_boxes, polybags, bottles, apparel]
container_types: [tote_A, shelf_bin_B]
data_needed:
tasks: [each_pick, put_wall, pack]
modalities: [video, station_images, wms_pick_records]
exclusions: [case_picking, trailer_unloading]
evaluation:
holdout: one site; one SKU family; one shift pattern
Choosing between them
Start from the deployment, not the data supply:
- If the target job is picking single items into orders, source fulfillment data, and check object variety and container types first.
- If the job is moving cartons and pallets, source distribution data and check pack formats, weight range and pallet patterns.
- If the target is a mixed site, specify the share of each, and plan separate held-out sets.
- If you are undecided, use the shared vocabulary in the warehouse task taxonomy to list the tasks and map them to the operation types that perform them.
Limitations of treating one type's data as the other's
Fulfillment footage does not show pallet building, and distribution footage does not show polybag grasping. Object sets, tote geometry, lighting, shelving height and worker posture differ. Even within a type, a site's product category (apparel, grocery, electronics) changes grasp difficulty considerably.
None of this data contains robot actions or forces unless it came from robot cells. Warehouse workers' video supports visual and step understanding, and robot-side logs exist for some deployments; see existing warehouse robot logs. Neither makes a humanoid controller on its own.
Worked example: two requests
Request one comes from a team building a humanoid for an apparel and consumer-goods fulfillment center. Its handling unit is the each, its object mix is soft and bagged goods, and its tasks are tote-to-tote picking, put-wall induction and packing. The team needs station video, tote and shelf imagery, and WMS pick records linked to timestamps. Case-handling footage from a grocery distribution center would add little, and could mislead evaluation by raising apparent diversity without touching the target objects.
Request two comes from a team targeting a regional distribution center. Its handling unit is the case and the layer, its objects are cartons, shrink-wrapped goods and mixed pallets, and its tasks are depalletizing, case picking and trailer loading. The team needs data that shows weights, carton deformation, pallet patterns and label orientation. Fulfillment footage of polybag picking would be the wrong match.
A third team may need both, such as a third-party logistics operator. That request should list each operation type separately, with its own held-out set, rather than blending counts into a single total that hides the imbalance.
Pick-face design can differ between the two operation types as well; see pick face types.
Questions to ask suppliers
- What share of volume is each-picking, case picking, pallet handling and returns?
- What is the SKU count, and how does it split across rigid, soft and irregular items?
- Which container and shelf types appear at the stations that would be recorded?
- Are orders single-line, multi-line or store replenishment?
- Which tasks does the supplier's footage cover, and from which camera positions?
- Do records link WMS pick tasks to video timestamps?
- Can the supplier break out data by product category and site area?
- Which client contracts restrict recording or reuse of the inventory shown?
Discuss your requirements
If you are defining which warehouse type your robot data should come from, describe the operation type, handling unit and tasks to SourceX. SourceX reviews the requirement and follows up about potential sources among US businesses; a request does not guarantee a match. For wider orientation see the warehouse data hub, the humanoid training data guide, and the catalog's supply chain and logistics data.
Discuss your humanoid training-data requirements with SourceX
Sources
- Design and control of warehouse order picking: A literature review — order-picking cost share and scope of the picking literature.
- ARMBench: An Object-centric Benchmark Dataset for Robotic Manipulation — scale of pick-and-place activity and object variety in one warehouse workcell.