Short answer
Warehouse robot training data covers the picking, packing, unloading and handling work done in fulfillment and distribution sites, recorded as footage, item images and system events. With WMS scans and weights used as labels, it supports perception in clutter, item identification, pick verification and task timing. It contains no robot grasp actions or contact forces, and 3PL footage often shows client-owned goods.
Why warehouses are a strong source
Warehouses combine three things humanoid teams rarely find together: a lot of repetitive manual handling, wide item variety, and systems that already log what happened. Order picking is described in the operations literature as the most labor-intensive and costly warehouse activity, with an estimate cited in one review putting it at as much as 55% of operating expense (de Koster et al., 2007). That is why picking, packing and unloading dominate the task pages in this cluster.
The system side matters as much as the video. A scan, a scale reading or a pick confirmation timestamps an event and says what item was involved. Joined to footage, those events become labels nobody had to draw by hand. Logistics identifiers help with the join: the GS1 Serial Shipping Container Code is an 18-digit key that identifies a case, pallet or parcel and links it to despatch advice data (GS1).
This cluster is written for applied deployment teams who need data from real sites, and for perception engineers working on clutter. It sits in the demonstrations, objects and evaluation layers of the humanoid training data guide.
What warehouse data can support, and what it cannot
| Asset | Supports | Cannot support alone |
|---|---|---|
| Pack-station and overhead footage | Activity recognition, task timing, exception spotting | Hand-level grasp detail |
| Commissioned head or wrist capture at pick faces | Hand-object interaction, step structure, inputs for co-training | Robot actions or contact forces |
| Tote, bin and shelf images | Segmentation in clutter, item identification | Grasp outcomes, unless a robot logged them |
| WMS and labor-system events | Ground-truth timing, outcomes and labels for footage | Any visual signal on their own |
| Existing warehouse robot logs | Robot observations and failure cases in real operations | Humanoid actions, unless the robot is a humanoid |
| Layouts and facility data | Deployment planning, simulation environments | Task demonstrations |
Research datasets show what is achievable when a site instruments one cell well. ARMBench, collected in a warehouse pick-and-place workcell, holds images, video and metadata for more than 235,000 pick-and-place activities on more than 190,000 unique objects, with benchmark tasks for segmentation in clutter, identification and defect detection (Mitash et al., 2023). MetaGraspNet pairs 217,000 synthetic bin images with a much smaller set of more than 2,300 real annotated images across five difficulty levels (Gilles et al., 2022), a common pattern when labeled real clutter is scarce. Scene datasets such as LOCO cover the larger objects of logistics: pallets, small load carriers, stillages and forklifts (Mayershofer et al., 2020).
What the data cannot do is close the loop on manipulation. Even well-labeled picking footage of people has no robot actions, and the lessons from robot picking competitions point to control, not perception alone: teams in the first Amazon Picking Challenge concluded that reactive control must be combined with deliberative planning (Correll et al., 2016). For the general boundary, see human demonstrations vs teleoperated robot demonstrations.
Sourcing routes
3PL and fulfillment operators. Third-party logistics providers run many sites with many clients, which makes them plausible suppliers of footage, records and access for commissioned recording. It also means much of the inventory on camera belongs to their clients, whose contracts may restrict how images of their goods and data about their orders are used. Read 3PL providers as data suppliers and client-owned inventory in 3PL footage before scoping.
System records. WMS, labor-management and conveyor systems hold scans, weights, dimensions and exceptions. They are rarely valuable alone but turn footage into labeled data; see pick confirmation as labels.
Commissioned recording. Existing cameras usually point at stations, not hands. Recording pickers or packers with head or wrist cameras gives hand detail and step structure, at the cost of site disruption, safety review and client permission. If no existing asset fits, SourceX can scope a commissioned collection of new recordings.
Existing robot logs. Sites that already run picking or mobile robots may hold robot observations and failure cases, though the robot vendor often controls them; see existing warehouse robot logs. SourceX's role is sourcing, rights review, licensing and delivery coordination; it can also arrange annotation and engage specialist capture partners, scoped case by case.
How to evaluate warehouse supply
- Client scope. Which clients' goods appear, and whether each client's contract allows AI training use.
- Camera placement. Whether the camera sees hands and the pick face, or only the station from above.
- Clock alignment. Whether video and system events share a time source, and the measured offset on a sample.
- Item mix. SKU count, size and packaging range, and whether the site is e-commerce eaches or B2B cases.
- Exceptions kept. Short picks, damages and mispicks retained rather than cut.
- Volume variation. Peak and off-peak periods, since pace and clutter change with volume.
- Labels available. Which events exist per task: scan, weight check, dimension capture, pack confirmation.
Illustrative: one pack-station clip joined to system events
| Time (synthetic) | System event | What it labels in the clip |
|---|---|---|
| 14:02:11.3 | Order ORD_88120 released to station PS_07 | Task start |
| 14:02:19.8 | Item scan SKU_40417 | First item handled |
| 14:02:31.2 | Item scan SKU_11902 | Second item handled |
| 14:02:44.6 | Carton C_14 selected | Box selection step |
| 14:03:02.0 | Scale reads 1.84 kg, within tolerance | Pack verified |
| 14:03:09.5 | Label printed, carton sealed | Task end and outcome |
Before trusting rows like these, confirm the clock offset between camera and WMS on a sample; a two-second drift moves every label to the wrong action.
Key pages in this cluster
| Group | Page | What it answers |
|---|---|---|
| Suppliers and rights | 3PL providers as data suppliers | The most plausible warehouse supplier and its rights issues |
| Suppliers and rights | Client-owned inventory in 3PL footage | Permissions for goods that belong to clients |
| Suppliers and rights | Pack-station camera recordings | Evaluating existing station footage |
| Suppliers and rights | Existing warehouse robot logs | Robot data already held at sites |
| Tasks | Each-picking demonstrations | Picking single items from totes and shelves |
| Tasks | Packing demonstrations | Box selection, void fill and sealing |
| Tasks | Trailer and container unloading | Unloading mixed freight |
| Tasks | Tote and container handling | Moving and stacking totes |
| Tasks | Case picking and mixed palletizing | Building mixed pallets |
| Tasks | Depalletizing mixed cases | Breaking down mixed pallets |
| Tasks | Parcel sortation hand tasks | Manual sorting of parcels |
| Tasks | Decanting cases into totes | Transferring inbound goods |
| Tasks | Returns processing workflow | Inspecting and dispositioning returns |
| Images and labels | Bin-picking image datasets | Images of cluttered totes |
| Images and labels | Pick errors and mispicks | Error examples and how they are detected |
| Images and labels | Pick confirmation as labels | Scans and weights as automatic labels |
| Images and labels | Rack, shelf and location label images | Location recognition |
| Images and labels | Dimensioning and weighing records | Measured item data from stations |
| Planning and evaluation | Warehouse task taxonomy | A consistent vocabulary for tasks |
| Planning and evaluation | Warehouse pick evaluation protocol | Pick-and-place evaluation design |
| Planning and evaluation | E-commerce vs B2B distribution | Why the two produce different data |
| Planning and evaluation | Warehouse layout and facility data | Layouts for deployment planning |
| Planning and evaluation | Warehouse onboarding materials | Training materials as task specifications |
Neighboring clusters
Item recognition beyond the warehouse, including catalog images and dimensions, is in product and object data; see product catalog data for robot perception. Package damage and condition grading are in inspection and failure data. Lifting, carrying and cart pushing outside warehouse contexts are in workplace task demonstrations. For catalog context, see supply chain and logistics operations records and SourceX's logistics industry page.
Discuss your requirements
Describe the tasks, item mix, site type and labels you need, and whether you need existing footage, new recording or both. Send the requirement to SourceX: it looks for US businesses that hold suitable footage and records, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. SourceX reviews each requirement and follows up about potential sources; a request does not guarantee a match.
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 labor-intensive and costly warehouse activity.
- Serial Shipping Container Code (SSCC) — identifier for logistic units.
- ARMBench: An Object-centric Benchmark Dataset for Robotic Manipulation — warehouse workcell pick-and-place data.
- MetaGraspNet: A Large-Scale Benchmark Dataset for Scene-Aware Ambidextrous Bin Picking — synthetic and real bin-picking images.
- LOCO: Logistics Objects in Context — logistics scene understanding.
- Analysis and Observations from the First Amazon Picking Challenge — lessons on control and planning in robot picking.