Short answer
Goods-to-person station data comes from fixed workstations where totes or shelves arrive and an operator picks to an order container. Fulfillment centers with automation already log scans, tote arrivals and light or screen prompts, and often hold station cameras. It can support station-level perception, pick-step timing and exception detection, but not robot actions or full-body mobility.
The deployment problem
At a goods-to-person (GTP) station, the system brings inventory to a fixed picker. A tote or pod arrives, a screen or light says what to pick and how many, the operator takes the item, scans or confirms, and places it into an order tote. The operator's reach envelope, view and tool set barely change between shifts.
For a buyer, that stability is the draw. The camera sees the same geometry every time, and each pick has system events around it. For a humanoid program the question is narrower: which part of this station can data teach, and which part is only context? Picking cost matters because order picking is described as the most labour-intensive and costly warehouse activity, with cost estimated at as much as 55% of total warehouse operating expense (de Koster et al., 2007). Stations are where that labour concentrates in automated sites.
This page is about the station as a data source. Item-level picking from shelves is in each-picking demonstrations; pack-side recordings are in pack station camera recordings.
Task anatomy
| Step | System event usually available | Visual content |
|---|---|---|
| Source tote or pod arrives | Arrival timestamp, source container ID | Container enters the view |
| Prompt displayed | Instruction: item, quantity, destination slot | Light or screen cue |
| Item located | None, or a scan | Operator reaches into compartments |
| Item removed | Often none | Grasp, lift, possible hesitation |
| Confirm | Scan, button press or put-light acknowledgment | Item at scanner or destination |
| Place | Destination container ID | Placement into order tote |
| Release source | Source departs | Tote or pod leaves |
| Exception | Short, damaged or mismatch flag | Operator pauses, calls lead |
The grasp itself, the one step a manipulation policy needs, is usually the step with no system event. That is why the record and the video need alignment before either labels the other.
What this data can support, and what it cannot
Can support. Perception of containers and compartments, item segmentation in totes, activity recognition at the station, step durations, scan-to-scan cycle time, and exception detection. Published ARMBench material gives a sense of what a warehouse workcell dataset contains: images, videos and metadata for 235K+ pick-and-place activities on 190K+ unique objects, with tasks including segmentation in clutter, object identification and defect detection (Mitash et al., 2023). That dataset comes from a robotic pick-and-place workcell, so it differs from a human-operated station. It is useful as a template for the metadata a buyer should ask for.
Cannot support. Station video from a human picker does not provide robot joint trajectories, grip force or compliance. The operator is seated or standing at a fixed counter; a humanoid would have a different reach, hand and viewpoint, and the embodiment gap applies fully. If the station already uses a robotic arm, its logs may exist, but they describe that vendor's system and not a humanoid; see existing warehouse robot logs. A GTP station also hides much of the building: travel, congestion and aisle navigation are absent.
Illustrative data package
# Illustrative only - synthetic structure, not an available dataset
station_type: gtp_pick_station
site_id: site_021
video:
camera_a: top-down, covers source tote and scanner
camera_b: oblique, covers order tote
clock: NTP-synced, offset logged per day
events_csv:
columns: [event_id, station_id, t_utc, event_type, source_container, dest_container, sku_hash, qty]
event_types: [tote_arrive, prompt, scan, put_confirm, short_flag, tote_depart]
labels_added:
pick_segment: start/end per pick (annotator or model-assisted)
outcome: ok | short | wrong_item | damaged
excluded: [robot_state, grip_force, operator_identity]
Quality and acceptance pointers
- Time alignment. Request the clock source for cameras and the system, and the measured offset. A scan that lands several seconds off its video turns every label into noise.
- Join keys. Containers and logistic units should carry stable identifiers; the GS1 SSCC is an example of a standard key for logistic units (GS1), though many stations use internal license plates instead.
- Coverage. Check the mix of SKUs per shift, compartment types and short-pick rates. A week of one product family tells you little about generalization.
- Camera view. A ceiling camera can hide the hand inside a bin. Request sample frames with the hand in the deepest compartment.
- Held-out data. Hold out by station, site or product category, whichever your claim needs.
Supplier types
Fulfillment centers with automated storage or shuttle systems are the main holders, followed by 3PLs running these stations for clients. Station camera systems are often installed for quality or loss prevention and are owned by the operator, while the goods and order data may belong to the client. Review 3PL providers as warehouse data suppliers before approaching one.
Limitations and additional preparation
Station data was recorded for operations. It does not cover human-scale movement through aisles, tote lifts from floor level, or the effect of a humanoid's body on station design. Operators' hands may be occluded by the tote walls or the scanner. The data needs permission checks, redaction decisions that do not erase the hands, segmentation of pick intervals and clean joins to events. A buyer who wants a robot at such a station still needs robot or teleoperated data from the target hardware.
A practical sampling approach
Ask the supplier for one full shift from two different stations, not curated clips. Then run four checks. First, list the share of picks where the item is visible at grasp time, and the share where it is hidden inside a compartment. Second, compare event counts per hour with the video: do prompts, scans and confirmations all appear, and do any events lack video? Third, look at how the station behaves during a tote jam or a late arrival, since waiting states dominate real footage and can swamp a naive training set. Fourth, ask whether the layout changed during the period, because new lights, new scanners or a moved camera alter the visual domain.
These checks tell you whether the package is a perception asset, a timing asset or both. Many station datasets are strong on timing and weak on hand-level detail. Decide which you are buying before you negotiate scope, and write it down in your requirements specification.
Questions to ask suppliers
- How many stations and how many distinct station designs are covered?
- Which events does the system log per pick, and which are not logged?
- How are camera and system clocks synchronized, and can you show the offset history?
- Is the item visible when it is grasped, or hidden inside a compartment?
- Can you share a sample day with video, events and exceptions side by side?
- What proportion of picks are shorts, damages or overrides?
- Who owns the SKU and order data, and has the client approved licensing?
- Are operator identifiers present in video or events, and what notice applied?
- Does any robotic arm at the station produce its own logs?
Permissions and provenance
Order data and product images may belong to the facility's clients, and operators are identifiable on camera and in login events. Ask for written confirmation of who can grant what, and see client-owned inventory rights and worker consent and identifiable people. This is not legal advice.
Discuss your requirements
To scope station data, tell SourceX what your evaluation set needs to cover: site types, modalities, event fields and permitted uses. SourceX reviews the requirement, follows up about potential sources among US businesses, and checks the data and licensing permissions. A request does not guarantee a match. Start from the humanoid data guide or the warehouse data hub; for catalog context see supply chain and logistics data.
Tell SourceX what your evaluation set needs to cover
Sources
- Design and control of warehouse order picking: A literature review — order-picking cost share estimate.
- ARMBench: An Object-centric Benchmark Dataset for Robotic Manipulation — example of warehouse workcell dataset contents.
- Serial Shipping Container Code (SSCC) — standard logistic unit key.