Signals Inbox·July 16, 2026·Humanoid Robotics

Humanoids: what warehouse tasks can they do today?

Humanoids can already transfer standardized totes, sort packages and retrieve products in warehouses, but only one narrowly defined workflow has reached meaningful commercial volume. The closer a task gets to mixed inventory, exceptions and independent decisions, the weaker the evidence becomes.

Humanoid warehouse tasks ranked by current capability

Q1Which warehouse tasks are humanoids closest to doing reliably now?

Standardized tote transfer is the only warehouse task currently supported by strong commercial evidence. Package sorting and fixed-station product picking form the next tier, though neither has published operating volume comparable with Digit’s GXO deployment.

The top five tasks share three constraints. The containers are standardized, the pickup and drop-off zones are predefined, and another system determines where the goods should go. The robot mainly executes the physical transfer. It does not own inventory decisions, order planning or exception resolution.

This pattern appears across unrelated developers. Agility connects AMRs with conveyors. UBTECH places totes at an automated storage workstation. Figure trains Helix to reorient packages inside a bounded logistics zone. Apptronik focuses on transfers between storage systems, carts and shipping stations. Different hardware, same useful job: closing the physical gap between existing machines.

The score falls once individual products replace standardized containers. A warehouse may contain tens of thousands of SKUs with reflective packaging, soft bags, damaged boxes and almost identical labels. Identifying an ordered item is only the first step. The robot must also find a viable grasp and maintain competitive speed.

For now, humanoids are reliable enough for carefully constrained material flow. Their ability to handle open-ended inventory remains far less convincing.

Q2Which humanoid warehouse tasks have moved beyond demonstrations?

Tote movement has clearly crossed from demonstration into commercial operation. A smaller group of tasks, including tote loading at storage systems and merchandise picking and packing, has reached real workplace pilots. Most of the remaining warehouse task map is still supported by internal videos or product pages.

Digit’s trajectory sets the strongest standard. GXO first tested the robot in December 2023, signed a multiyear commercial agreement in June 2024 and continued operating it until the workflow passed six figures in volume. The progression contains three separate forms of evidence: customer testing, contractual commitment and accumulated output.

UBTECH’s three-week Zeekr program qualifies as a real pilot because Walker S Lite trained at an operating CTU workstation. Sanctuary’s week inside a Mark’s store also counts, although the robot divided its time among 110 activities rather than repeatedly performing one high-volume task.

Apollo illustrates an earlier stage. GXO announced a multi-phase research program with Apptronik in June 2024, while Apptronik has since built a detailed menu of warehouse solutions. Public evidence still emphasizes what Apollo is designed to do rather than what a customer has measured over months.

Across the tracker, only three of 30 tasks reach operational or scaled status, and all three are variations of tote transfer. Four more have credible workplace-pilot evidence. The remaining 23 are demonstrated or claimed. Warehouse humanoids are in commercial operations now, but only just.

Q3Can humanoids work a full warehouse shift without human help today?

No humanoid has publicly demonstrated an independent warehouse shift with the metrics needed to verify that claim.

A warehouse operator would need at least four figures: productive uptime, successful cycles per hour, human interventions and average recovery time. Humanoid developers generally publish cumulative movements, scheduled operating windows or battery features instead. Those figures show progress, but they cannot tell us whether a robot completed eight hours of useful work without engineers rescuing it.

Battery life is becoming less decisive. Atlas and Walker S2 can replace their own batteries, while fleet systems can rotate robots through charging. Continuous power still leaves grasp failures, misplaced objects, blocked paths and software faults.

Commercial support structures reveal the current reality. GXO’s Digit deployment includes Agility Arc for mapping, workflow management and troubleshooting. A robot can be commercially useful while still depending on fleet software and specialist support.

The industry has proved that humanoids can participate in long-running warehouse operations. It has not shown that one robot can independently absorb the variability of a human shift. A credible breakthrough would be a full-shift report disclosing cycles, uptime, interventions and errors, not another endurance headline.

Q4Why use humanoids when warehouses already have specialized robots?

Humanoids make sense where existing automation leaves a short but expensive manual gap. They are a weak choice when a stable, high-volume workflow can justify a machine designed specifically for that task.

Boston Dynamics’ Stretch already unloads trailers and builds case pallets using a compact wheeled base and purpose-built arm. AMRs move inventory efficiently across floors. Conveyors dominate fixed routes. Robotic arms outperform humanoids inside repeatable workcells. These machines avoid the balance, battery and control complexity created by legs and humanlike hands.

Digit’s GXO task explains where the humanoid form can still win. The warehouse already had AMRs and conveyors, but someone had to transfer totes between them. Installing another fixed system could require space, engineering work and permanent changes to the facility. Digit entered the existing human workstation and used the same totes and transfer points.

That advantage is strongest in third-party logistics. A 3PL frequently changes clients, products and workflows, making expensive fixed automation harder to justify. Apptronik explicitly targets this flexibility: Apollo is designed to use existing racks, carts, bins and workstations rather than requiring a dedicated system for every contract.

Humanoids compete mainly for the manual connections left between specialized machines. Their commercial future depends on whether that flexibility can compensate for lower speed, greater complexity and heavier support requirements.

Q5Which warehouse jobs are humanoids still far from doing?

Humanoids remain far from inventory-intensive jobs where every order creates a different sequence of perception, manipulation and judgment.

General order picking is the clearest example. Retrieving one visible product in a demonstration is already possible. A commercial picker must process hundreds of items per hour across thousands of SKUs while avoiding substitutions, damage and inventory errors. No developer currently publishes humanoid results at that level.

Packing adds another layer. The robot has to verify the order, select appropriate packaging, arrange several products efficiently, protect fragile goods, apply the correct label and recognize when something is missing. Sanctuary has shown individual packing and labelling activities, but not this complete decision chain.

Exception handling is further away. Human warehouse workers constantly resolve problems that never appear in task descriptions: torn packaging, leaking products, blocked pick faces, incorrect inventory, collapsed stacks and urgent order changes. These cases occur too irregularly to justify a specialized machine but frequently enough to disrupt autonomous operation.

Our tracker reflects the divide. Twenty-one tasks score below 55, including every task involving broad SKU coverage, independent task switching or end-to-end fulfillment. Humanoids can execute physical steps within warehouse jobs today. They remain poor owners of workflows whose next action changes with every exception.

Q6Will humanoids replace warehouse workers or mainly fill automation gaps?

During the first phase of deployment, humanoids will fill automation gaps and remove individual physical tasks rather than replace complete warehouse occupations.

Digit’s commercial assignment removes one transfer loop between AMRs and a conveyor. Employees are still needed for receiving, picking, packing, quality checks, exceptions and supervision. Apollo’s proposed workflows follow the same structure: case picking, palletizing, trailer unloading and workcell delivery are sold as separate modules.

The fragmented task model affects headcount differently from conventional replacement narratives. One humanoid may remove repetitive lifting from several workers’ shifts without eliminating any one person’s entire role. Warehouses can then reduce overtime, avoid filling high-turnover positions or reallocate people to variable tasks that remain difficult to automate.

The labor impact could become larger once robots perform several workflows with the same hardware. That possibility remains central to the humanoid business case, but the evidence is weak today. None of the live deployments in our tracker shows a robot autonomously switching between picking, packing and material movement during one shift.

The near-term unit of automation is the task, not the job. Meaningful job displacement would require several automated tasks to overlap within the same worker role and operate reliably enough that the remaining human activities no longer justify a full position.

Q7What will be the first humanoid warehouse task to reach real scale?

Transferring standardized totes between warehouse systems remains the strongest candidate, with approximately 75% probability. We assign 15% to fixed-station parcel sorting, 7% to case or product picking and 3% to every other warehouse task combined.

Tote transfer already has the only six-figure humanoid workflow publicly documented. Agility has also signed agreements with Mercado Libre and Toyota Motor Manufacturing Canada, giving the same platform a route into e-commerce logistics and automotive material flow.

The task is unusually transferable. Warehouses across industries use similar plastic totes, conveyors, AMRs and putwalls. A product-picking system must adapt to each customer’s SKU catalog, while a tote-transfer workflow can remain largely unchanged.

Manufacturing capacity is no longer the only constraint. Agility’s RoboFab is designed for peak annual capacity of 10,000 robots, Figure has produced more than 350 Figure 03 units, and Boston Dynamics is beginning Atlas enterprise deployments. Hardware fleets could grow faster than the number of validated workflows, pushing developers to replicate the few tasks that already perform reliably.

The largest threat comes from specialized automation. Once a tote route becomes permanent and high-volume, a fixed transfer device or simpler wheeled robot may offer better economics. Humanoids retain the advantage where layouts change, several transfer points share one robot or the customer cannot rebuild the facility.

The first scaled warehouse humanoid task will probably remain extremely narrow: moving a standardized container across the final few meters between systems built to automate everything around it.