Industrial deployment is the installation of robots in factories and warehouses, where each cell has to pay back its cost. IFR counts a record 5 million industrial robots operating worldwide in 2025, after more than 600,000 installations that year, 59 percent of them in ChinaSource 21.

Orange KUKA industrial robot arms with welding guns surround a bare metal car body shell on a body-shop line at the BMW plant in Leipzig, with more robots in the hall behind.
KUKA robots spot-welding BMW 3 Series bodies at the Leipzig plant in 2005. Fixed cells repeating one motion like this make up most of the installed base.
Photo: BMW Werk Leipzig / Wikimedia Commons. Licence: CC BY-SA 2.0.

The economics you can drag

Drag the time to fix each jam and watch the payback change. Payback here divides cell capital by modeled monthly labour value, leaving out the damage and throughput savings in EVST's formulaSource 13.

Cell economics calculator
Slow jam fixes delay when a picking robot pays off
How often a pick fails
Assumption: seconds of human attention per failed pick; this dial is the whole argument.
Adjust more
Assumption: $80k is a chosen example, not a sourced arm-price quote. EVST gives complete-cell budgets, not fixed list prices.
Assumption: 2.5x is chosen within EVST’s 2-3x complete-cell guidance, not a measured cell.
Assumption: a paced piece-picking cycle; drag it to match any quoted cell.
Assumption: availability net of maintenance and faults; 95% is a working figure, not a vendor claim.
Assumption: the policy headline number; the demonstration teaching step drops this to 99.
Assumption: fully loaded picker wage; set it to your own facility number.
robot pickinga person fixing jamsstopped for upkeep

Pays back inside 24 months

The poise rests where the beam balances: the months to pay backPays for itself in 11.6 months

Illustrative: chosen inputs, not a measured cell.

Every failed pick pulls a person over, and how long that takes decides when the robot earns back its price.
How this was made

The calculator values a robotic pick cell by the picker wages its output replaces. Cell cost is robot price times the integration multiple ($200,000 now). Each failed pick is a jam a person clears, so every good pick carries the jam rate times the clearing time on top of its cycle. Payback divides cell cost by the monthly labour value of the modeled picks; the verdict compares it with a chosen 24-month horizon.

Current case: payback 11.6 months; capital cost per pick 0.008 USD over 5 yr. Cell cost $200,000; 569 modeled picks per elapsed hour and 415,062 per 730-hour month; jam rate 0.10%.

With 15 seconds to fix each jam, it pays back in 11.6 months when 1 pick in 1,000 fails and 11.8 months when 1 in 100 fails.

Where the elapsed hour goes: productive 3411 s, jam clearing 9 s, downtime 180 s of 3600 s.

  • Time to fix each jam (jam-clearing time, default 15 seconds): Assumption: seconds of human attention per failed pick; this dial is the whole argument.
  • Robot price (robot cost, default $80k): Assumption: $80k is a chosen example, not a sourced arm-price quote. EVST gives complete-cell budgets, not fixed list prices.
  • Whole cell cost, in robot prices (integration multiple, default 2.5x): Assumption: 2.5x is chosen within EVST’s 2-3x complete-cell guidance, not a measured cell.
  • Seconds per pick (cycle time, default 6.0 s): Assumption: a paced piece-picking cycle; drag it to match any quoted cell.
  • Share of time the cell runs (uptime, default 95.0%): Assumption: availability net of maintenance and faults; 95% is a working figure, not a vendor claim.
  • Picks that succeed (per-pick success, default 99.9%): Assumption: the policy headline number; the demonstration teaching step drops this to 99.
  • Hourly wage of the picker it replaces (displaced wage, default $25/h): Assumption: fully loaded picker wage; set it to your own facility number.

All seven defaults and slider ranges are authored assumptions; EVST’s guide supplies context, not exact inputs; none is a measured deployment result. The chart’s lines hold every other input at its current value and vary only the clearing time, at 99.9% (1 failed pick in 1,000) and 99% (1 in 100) per-pick success. The hour bar spends 3,600 elapsed seconds on productive cycles, jam clearing and downtime.

Months to pay back rise with jam-clearing time. At 99% per-pick success they go from 11.8 months with 15-second clearing to 17.3 months with 300-second clearing; at 99.9% they reach 12.1 months at 300 seconds. The current case pays back in 11.6 months.

Months to pay back by jam-clearing time
jam-clearing secondspayback at 99.9% successpayback at 99% success
511.511.6
1511.611.8
6011.712.7
12011.813.8
18011.915.0
24012.016.1
30012.117.3
Picking-cell economics
Same 99% success; slow fixes eat 19 minutes an hour
Adjust more
outer ring: robot pickingmiddle ring: a person fixing a jaminner ring: stopped for upkeep

beat 4 / 4 Pays for itself in 11.8 months with quick fixes, 17.3 with slow ones

Illustrative: chosen inputs, not a measured cell.

A robot that rarely fails can still pay off slowly if every failure needs a person for minutes.
How this was made

Both hours come from the cell economics calculator above, with its seven authored defaults except per-pick success, set to 99%: robot cost 80,000 USD, integration multiple 2.5x, a 6-second cycle, 95% uptime and a 25 USD hourly wage. Only the jam-clearing time changes: 15 s against 300 s. Downtime is the 5% of the hour the cell is not up.

Quick fixes: 83 s of jam clearing per elapsed hour, 556.1 picks/h, payback 11.8 mo. Slow fixes: 1140 s of jam clearing, 380.0 picks/h, payback 17.3 mo.

The drawing is a control desk for the two cells. Each door carries a 60-minute circular chart recorder whose paper turns once an hour; the pen rides the outer ring while the robot picks, the middle ring while a person clears a failed pick and the inner ring while the cell is down for upkeep. Each hour is laid out as runs of 100 good picks, each followed by one fix, with the last partial run split in the same ratio and the downtime at the end, so the inked arcs add up to the calculator's time split. The counter under each chart counts good picks (556 and 380 in the hour) and the timer beside it shows how long each fix takes.

Illustrative cell, not observed economics or a financial forecast. The seven authored inputs and the 730-hour month, 60-month amortization and 24-month verdict assumptions are stated with the calculator above.

This worked example has no measured deployment behind it, and each of its seven inputs is an assumption, labelled beside its control. The model also assumes 730 hours a month and a 60-month amortization period, counts modeled picks per elapsed hour, including downtime, and judges payback against a chosen 24-month horizon, the upper end of EVST's estimate for multi-shift palletising cellsSource 13.

At 15-second jam clearing, cutting per-pick success from 99.9 to 99 percent moves payback from 11.56 to 11.82 months, an increase of 2.2 percent. At 300 seconds the same cut moves payback from 12.11 to 17.30 months, and modeled output falls 30 percent, from 542.9 to 380.0 picks per elapsed hour. EVST, a vendor that quotes complete cells, publishes no fixed list prices and puts the robot at a third to half of total cell costSource 13.

What the installed base automates

5 million
operational stock
industrial robots worldwide, 2025
600,000+
2025 installations
up 11% on 2024
59%
China share
of global installations in 2025

IFR's World Robotics 2026, released on September 24, 2026, counts a record operational stock of 5 million industrial robots in 2025, up 9 percent, after more than 600,000 installations that year, up 11 percent. China installed 354,000 units, 59 percent of the total, and the United States, with almost 38,500 installations, passed Japan to become the second-largest marketSource 21.

A year earlier, World Robotics 2025 counted 4,663,698 robots in operation and 542,076 installed in 2024, above 500,000 in each year from 2021Source 1. China then held 2,027,190, about 4.5 times second-ranked Japan's 450,530, and took 295,045 installations, 54 percent of the totalSource 1 Source 2.

Strip the installed base down by task and most of it is repetitive motion in a fixed cell, with little or no perception worth the name. OSHA's Technical Manual lists uses including arc and resistance welding, painting and spraying, machine-tool loading and unloading, assembly, materials handling and packagingSource 16. In A3's North American order data for 2025, collaborative robots were 7,212 units worth $241 million, 19.6 percent of the 36,766 robots ordered and 10.7 percent of the $2.25 billion in order value. A3 began reporting them as a distinct category in Q1 2025Source 3.

Learned systems enter mostly through logistics. Automated storage and retrieval puts high-bay racking and shuttles where aisles and forklifts used to be, and AutoStore reports about 2,000 systems in 68 countriesSource 22. Autonomous mobile robot fleets move cases and totes, and transport and logistics robots sold 117,500 units in 2025, 47 percent of the almost 250,000 professional service robots shipped that year, a total up 24 percent on 2024Source 23. Goods-to-person stations bring the shelf to a stationary worker, which sets the productivity bar for automated piece picking.

In October 2023 Amazon reported over 750,000 robots working collaboratively with employees; Sequoia was operating at one Houston fulfillment center, and testing Digit for tote recycling was plannedSource 5. Its fleet overview, updated in June 2026, reports more than one million robots deployed since 2012, a cumulative total and not a simultaneous active-fleet count, with next-generation Proteus still a lab pilotSource 6. In September 2026 Amazon said its robots work across more than 300 facilities and assist employees with 75 percent of the customer orders it deliversSource 24. Its DeepFleet model was set to improve fleet travel efficiency by 10 percentSource 25.

Amazon's Vulcan picks and stows approximately 75 percent of the types of items stored at its fulfillment centers, at speeds comparable to front-line employees. It hands items it cannot move to employees, and the post reports no pick-success rate or independently validated reliabilitySource 7. An Amazon Robotics paper on stowing into fabric pods does report outcomes: over 100,000 attempts in a fulfillment center, 85.86 percent succeeded, 9.31 percent were unproductive and 4.83 percent were defectsSource 26.

Symbotic's FY2025 Form 10-K reports about $22.5 billion of backlog, meaning unperformed obligations under existing contracts and not revenue already earned, at September 27, 2025, against $2.246922 billion of revenue for the year. Its Walmart agreement, expanded in May 2022, covers all 42 regional distribution centres, and Symbotic had 50 systems in deployment and 48 operational systems under software maintenance and support contractsSource 4.

In November 2025 Kroger said it would close three automated fulfillment centres, in Wisconsin, Maryland and Florida, while monitoring the remaining facilities, and expected about $2.6 billion of impairment and related chargesSource 15. This is Money counted five remaining sitesSource 14. On December 5, 2025 Ocado and Kroger agreed a one-off $350 million payment that also covers Kroger's decision not to proceed with Charlotte. Ocado expects the closures to cut its FY26 fee revenue by about $50 millionSource 27. Ocado Intelligent Automation sells the same grid-robot warehouses outward into healthcare, third-party logistics and industrial partsSource 17.

Boxy white robots with blue tops run on the rails of an aluminium storage grid that fills an Ocado warehouse hall; one robot on the left is blurred by motion.
Bots on the storage grid of an Ocado warehouse in 2017. They run on top of the grid and fetch bins from the stacks inside it.
Photo: Techwords / Wikimedia Commons. Licence: CC BY-SA 4.0.

Humanoids in plants

IFR counts about 7,000 full-size humanoids sold in 2025 and says most applications remain specialized and often require human teleoperationSource 23. Agility says Digit 4 has logged more than 65,000 operational hours and moved 100,000 totes at GXO's Flowery Branch site at about 98 percent accuracy while on-taskSource 28. Hyundai is training Atlas to sequence automotive parts at its Robotics Metaplant Application Center and expects component assembly by 2030Source 29.

Jonathan Hurst, Agility's cofounder and chief robot officer, wrote in IEEE Spectrum in October 2026 that in Digit 4's commercial deployments the robots had to work behind physical barriers, kept apart from people, to meet insurance and regulatory demands, which limited their usefulnessSource 38. As a person approaches, Digit 5 is programmed to stop, set down its payload, sit on the floor and cut power to its motors before the person can touch it, which he calls "the only approach that makes sense as a first step"Source 38. Its safety sensors and computers, larger battery and stronger motors added bulk, and Digit 5 can no longer run or jump as earlier versions couldSource 38.

Figure's 11-month BMW deployment set an 84-second cycle requirement, a target above 99 percent success per shift and a goal of zero interventions. It reports 1,250+ hours of runtime and 90,000+ parts loaded, without achieved rates against those targetsSource 30. Each of these figures comes from the company itself. Unitree's IPO prospectus adds the financial side: first-quarter 2026 revenue growth slowed to 68.49%, from 332.64% a year earlier, and adjusted net profit fell 52.55% year on yearSource 31.

The integrator's multiple

A robot arm is a minority of what a cell costs. EVST puts complete palletising cells at two to three times the arm's priceSource 13. Its July 15, 2026 commercial palletising-cell guide lists application-specific end-of-arm tooling, fencing, light curtains or area scanners, interlocked gates, safety-rated PLCs, installation, PLC/HMI integration, commissioning, pattern programming, software licences and any vision-guided pick logic for mixed loads. The surrounding-line work depends on what is already installedSource 13.

Orange KUKA robot arms inside a wire-mesh safety fence lift yellow crates of bread from a roller conveyor and stack them in tall piles, in a bakery hall.
KUKA robots palletising bread crates at a German bakery in 2005. The grippers, conveyors and safety fence around them are part of the cell's cost.
Photo: KUKA Roboter GmbH / Wikimedia Commons. Licence: Public domain.

OSHA's Technical Manual, discussing ANSI/RIA R15.06-2012, says the systems integrator completes and documents the application's risk assessment before commissioning and gives the results to the employer. Manufacturers or employers may also act as integrators. The employer retains responsibility for a safe workplace, and a risk assessment alone does not establish that the application protects workersSource 16.

In its April 16, 2026 investor outlook, Bessemer predicts that near-term value will accrue to full-stack, vertically integrated players ahead of pure-play foundation-model companies. It says deployment requires domain-specific data collection, target-environment fine-tuning, hardware integration and operational infrastructure. It also stresses expensive data collection and foundation models not yet general enough to work out of the boxSource 20.

The metrics that decide a deployment

Cycle time is the time required to produce a part or complete a process, measured for that operationSource 11. Takt time is available production time divided by customer demand, and its purpose is to match production with demandSource 10. NASA distinguishes inherent availability, computed from mean time between failures and mean time to repair without administrative, logistics or preventive-maintenance time, from operational availability, which includes them. Uptime is time in an operable stateSource 12.

The calculator above reports capital cost per modeled pick: robot price times the integration multiple, divided by modeled lifetime picks over a fixed amortization period. It does not include running costs. The payback period is total system capital cost divided by monthly labour, reduced-damage and throughput value. EVST's guide estimates 12 to 24 months for multi-shift palletising cells and 24 to 36 months for single-shift or lower-throughput operationsSource 13.

The jam rate is the lesson

A separate authored example has no measured intervention rate behind it. At 99 percent per-pick success, with every failure needing a ten-second clearing, overhead is (1 − 0.99) × 10 = 0.1 seconds per attempted pick. Amazon's stow system is designed to stow 80 percent of warehouse items at 300 units per hour. A no-stow turnaway, when a pod leaves with nothing stowed, consumes at least 12 secondsSource 26.

The intervention rate and clearing cost together decide whether a cell ships. Physical Intelligence reports that training on deployment data, including expert corrections during autonomous runs, more than doubled throughput and roughly halved the failure rate on some of its hardest tasksSource 32. Ken Goldberg proposes combining model-based engineering with model-free learning so robots can perform useful work, collect real-world data, and use those data to improve performance and learn adjacent skills. Goldberg does not claim that engineering removes the need for learning, and he expects model-free AI eventually to enable fully general-purpose robotsSource 19.

The Reliability Gap tracks the demo-to-production gap and selected deployment examples; Autonomous Vehicles covers the same question for driverless cars. Morgan Stanley's 2026 note on the humanoid industry's "PR problem", as reported by CNBC, describes investors looking for real-world return on investment over polished demonstrationsSource 18. Figure says its livestreamed logistics run lasted 200 hoursSource 33, but it was one task at one site, with no published success rate.

Brownfield versus greenfield

A greenfield site is designed around the automation: floor loading, ceiling height, power, network and traffic lanes exist because the system needed them. Amazon's Shreveport fulfillment center was built with 10 times more robotics, and its Sequoia can hold more than 30 million itemsSource 34. A brownfield deployment must fit the automation into a building that already runs.

OSHA describes safeguards for non-collaborative robot applications during automatic operation: guards such as fences and barriers, interlocked guards and presence-sensing devices. Teaching inside the safeguarded space uses reduced-speed manual mode with an enabling device, and maintenance follows lockout/tagout procedures under 29 CFR 1910.147 or 29 CFR 1910.333Source 16. Symbotic says it can install its systems in phases while the existing warehouse keeps operatingSource 4. Safety and Assurance covers the wider obligations.

The labour evidence, disagreement intact

Daron Acemoglu and Pascual Restrepo find that one more robot per thousand workers cuts the US employment-to-population ratio by about 0.2 percentage points and wages by 0.42 percent. Losses concentrate in commuting zones exposed to industrial automationSource 8.

In Germany, robot-driven displacement in manufacturing was fully offset by new jobs in services, and the cost fell mostly on young workers entering the labour forceSource 35. Spanish manufacturers that adopted robots raised output by 20 to 25 percent within four years and created net jobs at a rate of 10 percentSource 36. French adopters also grew employment, at their competitors' expense, and the overall effect on industry employment was negativeSource 37.

The MIT Task Force on the Work of the Future, co-chaired by David Autor and David Mindell with Elisabeth Reynolds as executive director, found in 2020 no compelling evidence of a jobless future. It describes automation displacing some tasks and creating new work, shaped by economic incentives, policy choices and institutionsSource 9.