Put Walls and Order Consolidation

The deployment problem

A put wall (also called a sort wall or order consolidation wall) is a rack of cubbies. Items picked in batches arrive at the wall, and each item goes into the slot belonging to its order. When a slot's order is complete, a light signals the packer to remove it. The structure is fixed, slot positions are known and the system tells the operator where to go.

That makes it a friendly task for robot data. Compared with open-floor picking, the geometry is repeatable and the correct answer for each put is known from the system. Order picking and batching are well-studied warehouse design topics; a literature review covers layout design, storage assignment, routing, order batching and zoning (de Koster et al., 2007), and consolidation is where batch picks are split back into orders.

This page is about the put wall as a task and data source. Picking itself is in each-picking demonstrations; the packing step that follows is in packing demonstrations; and handling returns to totes is in tote and container handling.

Task anatomy

What this data can support, and what it cannot

Can support. Detection of lit slots and slot occupancy; hand-to-slot targeting from video; step segmentation; error detection (wrong slot, double put); cycle-time and throughput statistics; and training perception to map a displayed cue to a physical target. The system record provides labels for free: slot ID, item ID, time.

Cannot support. The put is only one part of the manipulation. The record does not capture how the operator grasped the item, how they avoided other items in the slot or what happened to bags that jammed. There are no robot joint states or actions. Operators reach into cubbies at heights and angles suited to a human body; a humanoid with a different reach and viewing angle needs data from that body, as discussed in the embodiment gap. Put-to-light data teaches cue following; it does not teach how to handle awkward items. For that, see what human demonstration video can and cannot do.

Illustrative record and video package

# Illustrative only - synthetic fields
wall_id: pw_02
site_id: site_028
slots: 96   # example geometry, stated by supplier
video:
  cameras: [wall_front_wide, operator_side]
  clock: ntp, offset logged
events:
  - {t_utc: 2026-04-01T08:10:05.2Z, type: item_scan, item_token: i_9c2}
  - {t_utc: 2026-04-01T08:10:05.9Z, type: slot_lit, slot_id: B14}
  - {t_utc: 2026-04-01T08:10:08.4Z, type: put_confirm, slot_id: B14}
  - {t_utc: 2026-04-01T08:10:41.0Z, type: order_complete, slot_id: B14}
labels_added: [hand_target_slot, put_outcome, exception_type]
excluded: [operator_name, customer_data, robot_state]
holdout: by wall and by site

Capture and labels

Existing assets: wall cameras, put-to-light logs, WMS order consolidation records, and exception reports. A useful addition is a camera view on the operator that shows the hand entering the slot. Labels a buyer can request or derive:

• slot ID at each put, from the system;
• the lit-slot state in each frame;
• whether the put matched the cue (right slot, wrong slot, corrected);
• item class, coarse rather than by customer product name;
• slot fill state at completion.

Quality and acceptance criteria

• Cue-to-video sync. Confirm that the light state in video matches the logged light state; sample across the day.
• Slot visibility. Check whether upper and lower rows are both visible, and how often the operator's body hides a target.
• Wall design variety. Slot sizes, light types and layouts differ; a set from one design may not transfer.
• Error balance. A set with no wrong puts cannot train error detection.
• Order variety. Single-line and multi-line orders, small and bulky items.
• Splits. Hold out whole walls and sites.

Supplier types

Fulfillment centers and 3PLs that run batch picking, especially for e-commerce, use put walls. Some purchase them from automation vendors, who may hold cloud-side logs under client contracts. The goods and order records often belong to client brands; see client-owned inventory rights and 3PL providers as warehouse data suppliers. Related operations are described in kitting in fulfillment.

Limitations and additional preparation

Put wall footage shows structured, system-guided work, so it does not represent the open-ended choices a general-purpose humanoid will meet. The camera may be mounted above the operator, hiding the hand. Records need time alignment and join keys; identifiers for operators and customers need removal or tokenizing. Cues in the video may be colored lights whose appearance depends on camera exposure. A policy trained on this data still needs robot-side data for reaching, placing and recovery in cubbies, collected on the target hardware.

Evaluating a put wall sample

A good sample day includes a quiet period and a peak period, because operator behavior and error rates change with load. Ask the supplier to show three things. First, a clip where the lit slot and the logged slot agree, and one where they do not. Second, an example of an overfull slot and how the operator resolved it. Third, a clip where an item is too large for the slot or does not fit the order and the operator escalates.

Also check how the wall assigns slots: static assignment by order, dynamic assignment per item, or batch-based. The assignment logic changes what the system record can label. Static walls give a fixed map from order to slot; dynamic walls need the assignment table to interpret each put. Ask for that table with the events, or the labels lose meaning. Finally, ask how the wall's lighting appears on camera; saturated indicator lights can wash out the slot edge.

Questions to ask suppliers

1. How many put walls and which designs (slot count, light type, depth) are covered?
2. Which events are logged: scan, slot assignment, confirm, completion, exception?
3. Is the logged light state synchronized with video, and how was that checked?
4. Can the camera see the hand inside the slot, or only the wall face?
5. What share of puts are errors or corrections?
6. How are overfull slots and missing items recorded?
7. Which item types and sizes pass through the wall?
8. Who owns order and item data, and have clients approved licensing?
9. Are operators identifiable in video or events, and what notice applied?

Permissions and provenance

Order contents, item images and operator identity are all present in this kind of data. Ask who can license each and whether any client contracts restrict use. See worker consent and identifiable people. Blurring does not by itself resolve a permission question. This is not legal advice.

Discuss your requirements

If you know the wall designs, item classes and event fields you need, describe the requirement to SourceX. SourceX reviews it, follows up about potential sources among US businesses, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match. Start at the humanoid training data guide or the warehouse and fulfillment hub; for catalog context see video recordings.

Discuss your humanoid training-data requirements with SourceX

Sources

• Design and control of warehouse order picking: A literature review — scope of order picking design topics including batching.