What Do Humanoid Robots Actually Do in Factories? 5 Jobs Beyond the Demo [ROBs26-0824EN]
Humanoid robots are entering factories, but they are not starting by assembling an entire car.
The first credible jobs are narrower: moving a tote from an AMR to a conveyor, sequencing parts in the correct production order, loading sheet metal into a fixture, feeding a workstation, or performing an initial visual check. These tasks sit in the physical gaps between automation systems that already work well.
That distinction matters. Six-axis industrial robots already dominate stable, high-speed operations such as welding and painting. AMRs already move materials efficiently across flat factory floors. A humanoid becomes interesting when a mobile machine must also grasp, reposition, and place objects in facilities designed around human reach, carts, racks, aisles, and tools.
The first business case for a factory humanoid is not “replace a worker.” It is “close a recurring handoff that fixed robots and mobile robots leave unfinished.” Start with the workflow gap, not the robot.
Updated August 24, 2026. Commercial operation, factory demonstration, trial, and future plan are labeled separately throughout this article.
Watch: five factory jobs humanoids are taking first
Watch the full video on YouTube
Evidence map: deployment, demonstration, trial, or roadmap?
Humanoid announcements often place a production milestone and a future ambition in the same paragraph. Procurement teams should separate them before comparing vendors.
| Program | Workflow | Maturity | What the evidence supports |
|---|---|---|---|
| Figure 02 at BMW | Loading sheet-metal parts into a welding fixture | Reported production deployment | An 11-month vendor-reported deployment with operating metrics |
| Digit at GXO | Moving totes between AMRs, conveyors, and floor positions | Reported commercial operation | A repeatable intralogistics handoff integrated into a live workflow |
| Figure 03 at BMW | Dynamic part sequencing and cart manipulation | Factory demonstration | Figure describes it as the first demonstration of this logistics workflow |
| Apollo at Mercedes-Benz | Intralogistics and initial component checks | Trial | Data collection and training for specific production use cases |
| Atlas at Hyundai Motor Group | Sequencing followed by component assembly | Roadmap | Sequencing planned from 2028 and assembly expansion from 2030 |
Figure says Figure 02 loaded more than 90,000 parts, accumulated more than 1,250 runtime hours, and contributed to the production of more than 30,000 BMW X3 vehicles. Agility Robotics says Digit moved more than 100,000 totes in GXO operations. These are primary company disclosures, not independent audits.
The five jobs—and why they come before general assembly
1. Tote transfer: connecting mobility with manipulation
An AMR can carry material across a facility, but it may still need a separate system—or a person—to unload the tote and place it on a conveyor. A fixed arm can perform the transfer, but only inside a fixed work envelope. A humanoid or mobile manipulator targets the combined workflow: approach, grasp, carry, reposition, and place.
Digit’s reported GXO milestone is useful because it describes a complete operational handoff rather than a single isolated movement. Agility says Digit moved totes on and off AMRs, transferred them to conveyors, and stacked them at another floor location. The value is not walking by itself. The value is integration into an existing logistics process.
Before selecting this use case, record hourly transfer volume, tote weight, grip geometry, pickup and drop-off height, aisle congestion, and exception frequency. A robot can succeed in a staged demo and still fail when a tote is damaged, rotated, overfilled, or presented outside its expected position.
2. Part sequencing: handling variation that defeats hard-coded motion
Sequencing is a deceptively difficult manufacturing task. Components must reach the line in the correct order, but carts, bins, and parts do not arrive in mathematically identical positions. Parts may shift, rotate, overlap, or become partially occluded.
Figure’s BMW material says Figure 03 must combine perception, bimanual grasping, foot placement, balance, and precise placement while adapting to those variations. The company describes this as its first demonstration of Figure 03 performing the logistics workflow at BMW—not a scaled commercial deployment.
Hyundai Motor Group’s roadmap reinforces the sequencing-first pattern. It plans to introduce Atlas to validated sequencing processes beginning in 2028, then extend applications to component assembly from 2030. The sequence itself is strategic: validate safety and quality in a constrained workflow before expanding task complexity.
3. Machine loading: preparing the automated process
Figure 02 did not replace BMW’s welding robots. It picked sheet-metal parts and loaded them into a welding fixture. Six-axis industrial robots then completed the welding and downstream work.
The most valuable detail in Figure’s report is the KPI structure. The published requirement was an 84-second total cycle, including a 37-second loading phase. The target was greater than 99% successful placement per shift, with zero human interventions. Figure also describes placing parts within a 5-millimeter tolerance in two seconds.
This is how a factory buyer should evaluate a humanoid. Do not ask only whether it can complete the movement. Ask whether it can meet takt requirements across a shift, maintain placement accuracy, and recover without consuming the labor it was meant to save.
4. Pack-out support and line feeding: automate the handoffs first
Packaging changes with product dimensions, inserts, labels, order profiles, and quality exceptions. A claim that one humanoid will complete every packaging step is usually too broad for a first deployment.
A lower-risk entry point is the work around packaging: supplying empty containers, bringing parts to the workstation, removing completed totes, returning empties, and transferring material to the next process. These actions may be simple, but they can occur hundreds of times per shift and create repetitive lifting or walking.
The strongest pilot candidates have standardized objects, limited SKU variation, predictable pickup zones, and a short manual recovery path when something goes wrong.
5. Initial inspection: triage, not final release
Mercedes-Benz says it is testing Apollo for repetitive intralogistics, including transporting components or modules to production lines and conducting initial quality checks of components. Employees transfer production knowledge using teleoperation and augmented reality while the robots collect data for specific use cases.
“Initial” is the critical boundary. A camera-based presence, orientation, or visible-anomaly check is not automatically equivalent to final quality release. A practical design uses the humanoid to identify obvious exceptions and route them to a dedicated inspection system or a responsible human.
Do not describe a demonstration as scaled production, a trial as autonomous operation, or a 2028–2030 roadmap as current capability. There is not enough public evidence to claim that humanoids already perform complete packaging, inspection, or general assembly without supervision.
Fixed arm, AMR, or humanoid?
The right question is not whether humanoids are more advanced. It is which automation architecture matches the workflow.
| Architecture | Best at | Constraint | Typical fit |
|---|---|---|---|
| Fixed industrial arm | Speed, precision, repeatability | Fixed work envelope and redesign cost | Welding, painting, stable assembly |
| AMR | Efficient material transport | Limited complex manipulation | Pallet, cart, and tote movement |
| Humanoid/mobile manipulator | Combining movement and two-handed manipulation | Cost, speed, reliability, safety, and service maturity | Sequencing, machine tending, cross-cell handoffs |
If a plant can redesign a high-volume station around fixed automation, a dedicated system may remain faster and cheaper. A humanoid earns consideration when the plant must preserve human-oriented infrastructure and absorb moderate variation without building a new machine for every product change.
The integration stack is the hidden product
The robot is only one layer of a production deployment. The commercial system also includes:
- task-specific grippers and fixtures;
- 2D/3D perception that tolerates lighting and occlusion;
- force or tactile sensing for fragile or variable parts;
- standardized carts, racks, handles, slots, and pickup zones;
- barcode and traceability workflows;
- PLC, MES, WMS, and quality-system integration;
- safety zoning, speed limits, stop logic, and restart procedures;
- charging or battery strategy, networking, and fleet management;
- remote support, spare parts, field service, and escalation paths.
A 90% average task success rate may sound impressive. It may still be unacceptable if the remaining 10% creates long production stops. Measure the recovery burden of every failure: who intervenes, how long recovery takes, whether upstream or downstream equipment must stop, and whether the same failure repeats.
A 30-day pilot designed for procurement—not publicity
Start with one task, one object family, one shift, and one accountable process owner. Capture baseline data before the robot arrives so that “improvement” has a real comparator.
- cycles per hour and labor minutes per cycle;
- travel distance and aisle width;
- minimum and maximum payload;
- pickup and placement heights;
- object variation and orientation range;
- exceptions per shift;
- current quality errors and ergonomic exposure;
- maximum acceptable recovery time.
| Pilot KPI | Measurement rule | Decision question |
|---|---|---|
| Cycle time | Measure normal, congested, and changeover periods | Can it sustain the required pace? |
| End-to-end success | Count success only after correct final placement | Are partial motions being reported as success? |
| Interventions | Log pauses, resets, repositioning, and teleoperation | Does support labor erase the benefit? |
| Productive uptime | Exclude waiting, charging, faults, and network loss | How many hours actually created output? |
| Safety stops | Record cause, location, and recovery time | Can people and robots share the workflow? |
| Quality impact | Track misplacement, damage, and missing parts | Did throughput improve without quality loss? |
Do not scale because the robot produced a good ten-minute video. Scale only after it meets cycle, quality, intervention, safety, and recovery thresholds across complete shifts.
Where suppliers can win before humanoid volumes scale
The near-term supplier opportunity may be larger around the robot than inside it. Every deployment must be adapted to a specific plant, process, object, software stack, and service environment.
Potential winners include gripper and fixture companies, industrial vision providers, safety integrators, cart and rack manufacturers, MES/WMS specialists, simulation teams, field-service networks, and RaaS operators. These suppliers turn a general-purpose platform into a measurable production system.
The strategic asset is not simply the robot model. It is the reusable deployment package: validated workflow, task model, tooling, safety case, interface specification, exception library, service-level agreement, and operating data.
Eight questions before issuing an RFP
- Is the task repeated often enough to justify automation?
- Are payload, reach, and grip requirements within a stable range?
- Does the workflow require both mobility and manipulation?
- Must the solution use existing human-oriented racks, carts, and aisles?
- Is object and SKU variation controlled?
- Can a failed cycle be isolated without stopping the entire line?
- Is recovery fast, documented, and safe?
- Does the use case clearly reduce ergonomic risk, labor scarcity, or downtime?
If most answers are no, improve conventional automation, workplace design, AMRs, or cobots before adding a humanoid.
Frequently asked questions
Will humanoids replace industrial robots?
Not in the first wave. Fixed robots remain superior for stable, high-speed, high-precision operations. Humanoids target variable handoffs around those systems.
Are legs always better than wheels?
No. Wheels can be more efficient on flat floors. Legs become valuable when the workflow includes obstacles, human-scale infrastructure, or frequent repositioning during manipulation.
Can a humanoid perform final quality inspection?
It may support presence, orientation, and visible-anomaly checks. Final release must remain connected to the plant’s validated inspection equipment, procedures, and accountable quality organization.
What is the safest first use case?
A repetitive task with standardized objects, limited exceptions, measurable ergonomic burden, and a short recovery path—often tote transfer, sequencing, or machine loading.
Should buyers purchase or use RaaS?
RaaS can reduce upfront capital, but the contract should define productive uptime, intervention limits, service response, data access, software updates, safety responsibility, and exit conditions. Compare total workflow cost, not the monthly robot fee alone.
Primary sources and verification scope
- Figure 02 at BMW: operating metrics and KPI definitions
- Figure 03 at BMW: sequencing demonstration
- Agility Robotics: Digit moved more than 100,000 totes
- Mercedes-Benz: Apollo intralogistics and initial quality-check trial
- Hyundai Motor Group: Atlas sequencing and assembly roadmap
- Boston Dynamics: Atlas product information
The deployment figures in this article are company-reported. Where independent audit data was not publicly available, the wording identifies the source. Demonstrations, trials, and roadmaps are not presented as scaled commercial operation.
Which recurring handoff in your factory should be measured first: tote transfer, part sequencing, machine loading, pack-out support, or initial inspection?
Comments
Post a Comment