We Reviewed 20 Humanoid Robot Deployments. The Evidence for Sustained Work Is Still Thin

Humanoid robots are already working in real factories and warehouses. But the evidence supporting individual deployments varies considerably.

At BMW's Spartanburg plant, Figure's humanoid robot loaded more than 90,000 parts. At a GXO warehouse, Agility Robotics' Digit moved more than 100,000 totes. These are substantial published operating records.

Elsewhere, manufacturers describe robots performing production tasks, completing factory training or entering partnerships with industrial companies. Those announcements do not all establish the same level of commercial maturity.

TechView Africa independently reviewed evidence covering 20 named humanoid-robot programmes involving six manufacturers, comparing what companies disclosed about tasks, locations, operating results and deployment stages. Our latest verification distinguishes cumulative production records from narrower performance metrics and programmes where operational evidence remains incomplete.

How TechView Africa analysed the 20 cases

Our review examined manufacturer announcements, customer statements, regulatory filings and independent reporting concerning Figure, Agility Robotics, Spirit AI, Apptronik, Hexagon Robotics and UBTECH.

We classified each named programme according to the strongest evidence we could verify, concentrating on information available by the original article's 21 September 2026 research cutoff. We revisited the supporting records on 16 October 2026 and documented remaining source limitations.

The classifications distinguish three levels of evidence:

Evidence categoryWhat it meansCases
Published cumulative outputA disclosed total of completed production or logistics tasks, supported by operating records2
Other active workplace useEvidence that robots performed tasks at a workplace, without comparable cumulative output totals5
Pilot, training or announced programmeExperimental use, factory training, validation, or a partnership without sufficient proof of routine operation13
Total20

This is a selected comparison of reported industrial programmes, not a worldwide census. Some companies provide detailed operating data; others describe partnerships or experiments without disclosing individual robot productivity.

The two clearest cumulative output records

1. Figure 02 at BMW Spartanburg

Figure 02 working inside BMW Group Plant Spartanburg. BMW says the robot accumulated about 1,250 operating hours and moved more than 90,000 sheet-metal components during the programme. Credit- BMW Group
Figure 02 working inside BMW Group Plant Spartanburg. BMW says the robot accumulated about 1,250 operating hours and moved more than 90,000 sheet-metal components during the programme. Credit- BMW Group

Figure Robotics published an unusually detailed account of its Figure 02 work at BMW's Spartanburg plant in the United States.

The company reported:

  • More than 90,000 sheet-metal parts loaded.
  • More than 1,250 operating hours.
  • Contributions to the production of more than 30,000 BMW X3 vehicles.

The robot handled a defined loading task on an automotive production line.

This is significant because the disclosure goes beyond a video demonstration. It identifies a task, a real customer facility and cumulative work performed.

It does not, however, establish that the robot could perform every task carried out by a factory worker.

2. Agility Digit at GXO

GXO
GXO

Agility Robotics reported that its Digit humanoid moved more than 100,000 totes at GXO's Flowery Branch logistics facility. The company also reported approximately 98% accuracy while on-task in subsequent operational disclosures.

The robot's work included moving and stacking containers within warehouse workflows. As with BMW, the evidence is valuable because it describes repeated physical work over an extended deployment.

These figures remain company-reported results. TechView Africa did not independently observe or audit the underlying production logs.

Five other cases with evidence of workplace activity

Our review identified five further cases with reported workplace operations but without publicly available cumulative output records directly comparable to BMW's and GXO's.

RobotCustomer or siteWhat the evidence supports
Agility DigitToyota Motor Manufacturing CanadaManufacturer-reported work at customer facilities
Agility DigitSchaefflerManufacturer-reported operational deployment
Spirit AI MozCATL ZhongzhouBattery connector insertion and production-line tasks
Spirit AI Moz1JD.com operationsCommercial production activity reported by Reuters, with limited site-level detail
UBTECH Walker SNIO manufacturing facilityAssembly and quality-inspection work reported by UBTECH

Agility's September 2026 announcement identifies Toyota and Schaeffler among its operational customers.

Spirit AI describes Moz performing testing-related work on CATL's battery production lines. Reuters separately reported commercial production deployments involving the company's humanoids.

UBTECH describes Walker S working alongside humans at NIO's final-assembly facility, including inspection-related tasks.

These reports establish more than laboratory demonstrations. However, the published evidence does not give us comparable site-specific totals for every case.

A finding our original classification needed to distinguish

CATL offers an important exception to the idea that only BMW and GXO disclosed numbers.

Spirit AI reports that Moz achieved more than 99% success in connector insertion on CATL's battery production line. It also describes an increase in daily workload.

That is quantitative performance information. But a task-success percentage and a cumulative number of successfully handled parts measure different things.

A robot achieving 99% success over a short controlled period has not necessarily demonstrated the same commercial endurance as one completing 90,000 production tasks across months of operation.

Our finding: Public reporting should distinguish task accuracy, deployment duration, throughput and total completed work instead of treating every robot-performance number as equivalent.

The remaining 13 programmes: Training, pilots and announcements

The largest category consists of industrial experiments, factory training and announced collaborations.

#ManufacturerCustomer or programmeEvidence classification
1Agility RoboticsMercado LibreCommercial agreement; site operation not independently established
2Hexagon RoboticsBMW LeipzigPilot and factory testing
3ApptronikMercedes-BenzIndustrial pilot
4ApptronikJabilManufacturing pilot and validation
5ApptronikGXOEarly-stage proof of concept
6UBTECHBYDFactory logistics training
7UBTECHGeely / ZeekrFactory training and task coordination
8UBTECHDongfeng Liuzhou MotorAutomotive task training
9UBTECHFAW-Volkswagen QingdaoFactory inspection application or trial
10UBTECHFoxconnDocumented logistics-training phase
11UBTECHAudi FAWReported factory collaboration; individual deployment details limited
12UBTECHBAIC New EnergyReported factory collaboration; individual deployment details limited
13UBTECHSF ExpressReported logistics collaboration; individual deployment details limited

BMW explicitly described its Hexagon AEON programme at Leipzig as a pilot. Apptronik and Jabil similarly announced a pilot involving factory validation.

UBTECH's annual reporting documented work and training relationships involving several automotive manufacturers, Foxconn and SF Express.

However, the last three cases in our table have weaker case-specific evidence: the annual report identifies collaborations, but does not independently establish sustained work by a named robot at each site.

Why we corrected the Mercado Libre classification

Our earlier comparison placed Mercado Libre among active deployments. On rechecking Agility Robotics' commercial agreement announcement, we found that its language described a planned integration at the company's San Antonio facility.

The announcement alone does not prove that regular robot operations had begun by our original research cutoff. We therefore moved Mercado Libre into the announced-programme category.

This changes the selected sample's classification from 2 quantified / 6 active / 12 experimental to 2 cumulative-output records / 5 other active / 13 pilot, training or announced programmes. The adjustment makes the comparison more precise. It does not mean the Mercado Libre programme failed or subsequently stopped.

What the strongest humanoid deployments actually demonstrate

Humanoid robots at work
Humanoid robots at work

The clearest commercial applications are surprisingly narrow. Figure 02 loaded sheet metal. Digit handled warehouse totes. Spirit AI's Moz inserted and checked battery connectors. Other programmes involved parts movement, sorting and inspection.

These are useful tasks, but they are not equivalent to an autonomous robot replacing an entire human occupation. The evidence suggests that industrial humanoids are making progress first in structured environments where tasks can be repeated, monitored and carefully evaluated.

The more difficult challenge is demonstrating equivalent reliability across unfamiliar locations, different physical objects and changing work requirements. That is the difference between showing that a humanoid can perform a task and proving that it can repeatedly deliver economic value.

What this means for Nigerian and African businesses

For African manufacturers, warehouses and logistics companies, the important question is not whether humanoid robots are impressive. It is whether they outperform an available alternative at an acceptable total cost.

That comparison should include specialised industrial robots, conveyor systems, autonomous mobile robots, existing workers and conventional machinery. Before investing in an industrial humanoid, buyers should demand evidence covering:

  • Throughput: How many useful tasks can it complete per hour or shift?
  • Reliability: How often does it fail, stop or require human intervention?
  • Operating cost: What are the real maintenance, energy, support and integration costs?
  • Safety: Can it work near employees without unacceptable risks?
  • Adaptability: Can it move to another valuable task without extensive retraining?

These questions matter especially where electrical reliability, spare-parts availability and specialist maintenance support may affect the economics of automation.

Our analysis of AI job vacancies across Africa found employers seeking people who can deploy, monitor and evaluate complex AI systems. Physical AI adds robotics maintenance, industrial safety and machine integration to that skills picture.

Similarly, our report on how AI is changing entry-level technology jobs explores why automation often transforms tasks before it replaces entire roles.

Our Recommendation

Businesses should evaluate humanoid robots using operational records, not demonstrations or partnership announcements alone.

The most persuasive evidence includes cumulative task counts, sustained performance over time, breakdown frequency, human supervision requirements and the economics of deployment.

BMW and GXO provide useful examples of manufacturers disclosing meaningful task-output totals. CATL adds valuable task-success evidence, though its published measures are not directly comparable. The strongest conclusion from our 20-case review is not that humanoids cannot work. It is that evidence of productive work must be evaluated more carefully than the industry's promotional language often allows.

Sources & Verification

Figure AI — BMW factory deployment

Spirit AI — CATL humanoid deployment and over 99% task-success rate

UBTECH — 2024 Annual Report documenting automotive factory trials and collaborations

Agility Robotics — Mercado Libre commercial agreement

GXO Logistics — Apptronik Apollo proof-of-concept programme

Frequently asked questions

1. Are humanoid robots already working in real factories?

Yes. BMW, GXO and CATL are among the organisations associated with documented humanoid industrial activity. However, the scale, duration and quality of published performance evidence differ between programmes.

2. Did only two humanoid robot deployments publish performance data?

No. Other programmes, including Spirit AI's CATL deployment, have reported quantitative task-performance measures. The two cases highlighted in this review are distinguished by their published cumulative output totals.

3. Does a commercial partnership mean robots are already working?

Not necessarily. A commercial agreement may precede installation, testing and routine operations. Evidence of a signed partnership should not automatically be counted as proof of an active factory deployment.

4. Will humanoid robots replace workers in African factories?

The available cases do not establish when or how extensively this will happen. Some narrow industrial tasks are being automated, but large-scale workforce effects depend on cost, reliability, regulation, adoption and the availability of alternative technologies.

5. Did TechView Africa physically inspect these factories?

No. Our investigation is an independent comparison of named, publicly documented industrial programmes using manufacturer records, customer statements, company filings and reporting. The classifications and conclusions are TechView Africa's own analysis; the underlying production figures remain attributed to their original sources.

Comment
Reader discussion

Leave a comment

Comments cannot be edited or deleted after posting. Please review your comment before submitting.

No comments yet. Start the conversation.

Found an error, outdated step or safety concern? Contact the desk.