Assembly Automation Robots 2026 — Part Tolerance Decides, Not Robot Accuracy

Once a plant has worked through material handling and palletizing, assembly is usually the next process on the list. But if you carry over the decision criteria that worked for material handling, an assembly automation robot project will stall. Material handling is a “place” motion, so it succeeds when you improve positioning accuracy. Assembly is a “mate” motion, and if the mating part moves, nothing goes in. This article uses a Japanese-owned component maker in Chonburi, Thailand as a model, follows the investment breakdown and payback period in hard numbers, and ends with the items you must settle on paper before you issue an RFQ.

Fix the model site and production conditions first

The main reason automation discussions turn into a shouting match is that every participant is assuming a different starting point. One person says three operators come out; another says only one. Neither is lying — they are picturing different line configurations. So this article fixes the model below at the outset and never changes it. When you substitute your own plant’s conditions, replace the rows of this table one at a time with your own values and the same procedure will give you your own answer.

Every row marked “Assumption” is a number placed here as a starting point for your own study. It is not backed by a statistic or a public document. Only the row marked “Sourced” comes from published material. Throughout this article, assumptions and sourced figures are kept strictly separate. The most dangerous thing in an automation business case is an experienced person’s rule of thumb quietly getting transcribed into an internal document as a hard number, until nobody at the approval stage can say where it came from.

From the fixed assumptions, let us derive the unit rates used later. Dividing 180,000 baht per person per year by 250 days and 8 hours per shift gives a fully loaded labor rate of 90 baht per hour. The Bangkok minimum wage is 400 baht per day, which over 8 hours is 50 baht per hour. The gap between 90 baht and 50 baht is statutory benefits, overtime, and indirect cost. That 90 baht per hour is an order of magnitude below the labor rate a Japanese plant would use in the same calculation, and that single fact drives every chapter that follows.

Volume is worth pinning down too. Two shifts of 16 hours at a 30-second cycle gives a theoretical daily output of 16 hours x 3,600 seconds divided by 30 seconds = 1,920 pieces, and 480,000 pieces over 250 days per year. That is a ceiling that assumes no downtime and no changeover, and a real machine will fall short of it. It is here so that you have a feel for how much improvement headroom exists per piece when we get to the cost chapters.

What makes assembly fundamentally different from material handling and palletizing

Strip a palletizing or case-packing robot down to essentials and it performs one motion — put the workpiece at a coordinate in space. The target coordinate is owned by the equipment, and if the workpiece is slightly off, the job still succeeds as long as the position settles after placement. That is why this class of automation came together by tightening repeatability, gripper force, and cycle time. Material handling and palletizing spread across the world first precisely because the target does not move.

Assembly is different. In assembly, the target coordinate is not owned by the equipment — it is owned by the mating part. The hole position is set by the mating part’s dimensions, that part sits in a fixture, the fixture has its own mounting error, and the part itself has dimensional scatter. However precisely the robot returns to the same position, if the target itself shifts slightly every cycle, the part will not go in. That is the decisive difference.

The second difference is the cost of failure. Miss a placement in material handling and you place it again. A missed grip is easy to detect and cheap to retry. Most assembly failures, by contrast, are irreversible. A press-fit cannot be undone once it is pushed home. Drive a screw in crooked and you destroy the mating thread, scrapping the whole part. Gall an insertion and a damaged part flows downstream to the next process. So in assembly automation, “detecting that it failed” carries the same weight as “not failing”.

The third difference is inspection. You can see whether something was placed, so a sensor or camera is enough to confirm completion in material handling. But you cannot see whether a fit is good. A part that looks seated but has not reached the specified depth, or that was pushed in at an angle, is indistinguishable from the outside. So assembly judges completion using physical quantities captured during the process — insertion force and displacement, fastening torque and rotation angle. This difference is what skews the cost structure described later.

The practical consequence is that the first step in an assembly automation study is not robot model selection. It is measuring how much the mating part actually moves, and deciding where that movement gets absorbed. Do you tighten the drawing tolerance and absorb it on the part side, absorb it through the fixture’s locating scheme, or give the robot a compliance mechanism to absorb it? Any project that starts model selection before that three-way split is settled ends up being rebuilt during commissioning. The first decision that belongs to the buyer is how tolerance absorption is allocated.

What the numbers say about assembly automation

To get away from anecdote, look at published statistics. The following rows are extracted from the Japan Robot Association’s shipment statistics by application for the October to December 2025 quarter, on a member-company basis.

Two things stand out. First, assembly accounts for 12.2% of units but only 4.7% of value. Dividing total shipment value by total units gives roughly 4.72 million yen per unit, while dividing assembly value by assembly units gives roughly 1.79 million yen — under 40% of the all-application average. Robots used for assembly are visibly cheaper than those used for welding or handling, because they carry less payload, reach shorter, and are physically smaller machines.

One caveat has to travel with that figure. This is the manufacturer’s shipment value, not the price of an installed system after it has passed through a system integrator. It is a different quantity from the cost of equipment standing on your shop floor. Read “1.79 million yen per unit” as “you can buy an assembly robot for 1.79 million yen” and the entire cost discussion below collapses. What the statistic actually shows is not an absolute price but a relative relationship between applications — in assembly, the arm itself is relatively cheap.

Second, look at how small threaded fastening is. Quarterly shipments were just 43 units, 0.08% of total shipments, and down 43.4% from 76 units in the same quarter a year earlier. Within the 6,341-unit assembly subtotal, general assembly takes 5,957 units while threaded fastening stays in double digits. This is not because the world’s screwdriving is un-automated. The opposite is true — screwdriving automation is built with special-purpose machines and automatic driver units rather than articulated robots. If you are driving from one direction, in one posture, on a single axis, there is no reason to reach for an articulated arm.

For background, the global picture. New industrial robot installations reached 542,076 units in 2024, the operational stock reached 4,663,698 units for a 9% increase, and robot density in manufacturing stands at 177 units per 10,000 employees. In Japan, full-year 2025 figures show orders of 1.0456 trillion yen, up 25.7% year on year, production of 945.3 billion yen and total shipment value of 993.8 billion yen. The market as a whole is growing. But inside that growth, the structure in which assembly takes a large share of units and a small share of value has not changed.

So how should a buyer use these numbers? The conclusion is blunt — do not run quotation comparisons on the robot model number. The arm is a small fraction of an assembly cell’s capital cost, and almost all of the price difference between bidders is generated by everything else. Comparing bids on the basis that “supplier A’s robot is slightly cheaper” tells you nothing. As the following chapters show, the money moves in part feeding, fixtures, compliance mechanisms, and fastening tools. Specify a quotation format, from the buyer’s side, that forces those items to be broken out separately from the arm. That is the first practical move this statistic supports.

Assembly motions fall into four classes, and difficulty changes by an order of magnitude

Talking about “assembly automation” as one thing gets you nowhere. Assembly motions split into four classes by how hard they are for a machine, and each needs different equipment and different physical quantities to monitor. The model line’s work — 1 press-fit, 2 insertions and 4 threaded fasteners — spans three of the four.

Insertion is the easiest of the four, but only when there is a chamfer. If the mouth of the mating hole is chamfered, the sloped face mechanically corrects a certain amount of misalignment. Without a chamfer in the design, the two end faces butt against each other the moment there is offset, and pushing harder galls the surfaces. The important point is that equipment cannot decide whether a chamfer exists. The drawing decides. In other words, the difficulty of automating an insertion process has already been fixed by the design department, on the drawing, before the equipment is ordered.

Fitting, and press-fitting in particular, is defined by being irreversible. Once it is pushed home it does not come out. So a press-fit needs control that records the relationship between load and displacement as it pushes, and stops the moment the curve departs from the expected shape. Judge on load alone — “it reached the specified load, so it passes” — and you will pass a defective part that was pushed in tilted and therefore built load early. Judge on displacement alone and a cracked part that reached the specified depth still passes. Reading load and displacement as a pair is the minimum condition for automating a press-fit.

Fastening gets a chapter of its own below, but its position in the list is worth noting — it is the only one of the four classes with an industry-wide framework for pass/fail judgment. Acceptance criteria for insertion and press-fitting are set by each company, whereas fastening has category classifications such as those in VDI/VDE 2862, with monitoring requirements defined according to criticality.

Joining carries an extra issue in a Thai plant. Adhesive curing and hot-melt behavior depend on temperature and humidity, so in a building without air conditioning, or during the high humidity of the rainy season, the parameter set used at the Japanese parent plant may simply not apply. “We brought the same equipment over from Japan and the bond will not stabilize” is a genuinely common enquiry, and it comes disproportionately from joining processes.

One practical rule follows from the four classes. Do not put all four into a single cell. Load insertion, press-fitting and fastening onto one robot and every tool, sensor and monitoring routine each of them needs concentrates in one machine, so any one of them causing trouble stops the whole line. Split motions of different difficulty into separate stations, or at the very least keep their monitoring systems independent. Simply requiring the specification to include a one-page allocation table showing which class goes to which cell prevents this design error.

Robot accuracy does not help — the three-layer tolerance stack and where compliance becomes mandatory

Robot catalogs carry a “repeatability” figure, and the number printed there is very small. Reasoning from that number that “it is this accurate, so it must be able to enter a tight clearance” is the single most common misreading in assembly automation.

First, repeatability describes how consistently the arm returns to the same place when given the same command. It says nothing about how close the arm gets to the commanded coordinate in absolute terms. Absolute positioning accuracy is normally an order of magnitude worse. And the catalog figure is measured under specified temperature conditions. In a Thai plant, the ambient temperature inside a building without air conditioning can swing a long way between early morning and early afternoon, and the arm’s structural members grow with heat. Equipment that ran without complaint at the Japanese parent plant, and then shows a rise in insertion failures only in the afternoon after being installed in Thailand, often has exactly this cause.

Even so, the robot’s own error is not what actually causes assembly problems. Errors stack.

On the shop floor, these three layers do not contribute equally. If the arm tip deviation is 1, the locating error of a workpiece in a fixture is typically several times that, and the dimensional scatter of the mating part several times larger again. In other words layer 3 dominates, and robot accuracy effectively does not matter. Yet in the early stages of a study, the discussion tends to concentrate on layer 1 — the layer with the least room to move.

There is also a statistical trap. Combine errors as a root sum of squares and the total comes out small. But assembly pass/fail is decided not by the mean of the combined error but by the worst case. If a part fails to go in only at the instant all three layers deviate in the same direction, what you have built is a machine that stops occasionally. On a line running 480,000 pieces a year, an event with a probability of 0.1% happens several hundred times a year. Measured in downtime, that is not negligible. So design reviews should be a binary judgment — does it hold at worst case, and if not, is the residual absorbed by a compliance mechanism?

Compliance mechanisms divide into passive and active. Passive means giving the end effector mechanical softness so that position self-corrects along the mating chamfer or taper. It adds little cost and responds fast. Active means using a force/torque sensor so the robot detects contact force and corrects position along the force vector. It is far more flexible, but it costs the price of the sensor and the control that goes with it, and it lengthens cycle time. The model line’s investment breakdown carries “compliance mechanism (2 force/torque sensor sets)” at 700,000 baht in the base configuration and 300,000 baht in the lean configuration precisely because the passive-versus-active choice shows up directly as a price difference.

The boundary the buyer has to decide is therefore clear. How far do you tighten the drawing tolerance, and from where do you absorb the rest with a compliance mechanism? Tightening the drawing raises the part price, and depending on the supplier’s process capability it may be effectively impossible. Absorbing it with a compliance mechanism raises capital cost and lengthens the cycle. This comparison can be done on paper. Take the increase in part unit price from tightening the tolerance, multiply it by the annual volume — 480,000 pieces on the model line — to get an annual cost increase, and set it beside the cost difference of the compliance mechanism, which is 400,000 baht in this article’s estimate. On a part running hundreds of thousands of units a year, a small rise in unit price can exceed the equipment cost difference in a single year of production. Doing that comparison on paper before entering model selection is the correct order of operations.

A natural reaction is to propose correcting posture scatter with a camera, but this needs care. Vision can fix where a part is and which way it is facing. It cannot fix scatter in the part’s own dimensions. If a hole diameter has drifted to the bottom of its tolerance band, no amount of positional accuracy will get the mating part in. The scope and limits of vision are covered in robot vision implementation in practice. For layer 2, since the locating scheme and the outsourcing terms for fixture manufacture drive both cost and accuracy, read the guide to outsourcing jig design and manufacture alongside this article. Guarding and risk assessment are outside this article’s scope — if you are considering a configuration where assembly happens in shared space with people, see collaborative robot implementation in practice.

Fastening quality assurance is not “it was tightened”, it is torque times angle

Of everything in assembly, threaded fastening has the most clearly defined quality assurance requirements. It is also the process that attracts the most misunderstanding.

The misunderstanding sits at the center of the judgment “it reached the specified torque, therefore the joint is good”. Torque alone cannot distinguish the following defects.

What separates them is rotation angle. When the fastener seats against the bearing surface, torque that had barely risen until then climbs sharply. Record torque and rotation angle together as a waveform and you can read, in one picture, where that transition — the seating point — falls, how long the fastener ran free before reaching it, and how many degrees it turned after seating.

A gap under the head is the state where debris or a lifted part remains between the bearing surface and the component while the specified torque is reached. The seating point itself still appears, but the rotation angle from seating to specified torque is shorter than on a good part, because the joint has not compressed to its proper spring length. Cross-threading and galling produce an abnormal torque rise before the seating point, because the threads bind partway through what should be the free-running segment. Over-tightening looks normal up to the seating point, but the angle turned after seating exceeds specification. Double tightening leaves almost no free-running segment, so torque climbs from the moment tightening starts. Only torque and rotation angle taken as a pair make a fastening pass/fail judgment possible. A missed fastener — a location that was never tightened at all — is not caught by waveform shape at all, but by counting whether the number of fastening cycles per product reached the specified count.

For evaluating the performance of the fastening tool itself, there is ISO 5393, a standard that defines a performance test method for tools used on threaded fasteners. It specifies a test method for measuring how much the actual delivered torque scatters when tightening at the same set torque, and it is referenced as a common yardstick for comparing tool repeatability.

The framework that decides how much control to demand is VDI/VDE 2862. This guideline classifies fastened joints into three levels of criticality.

Here is the crux for the buyer. The category is decided by the product design department, not by the equipment maker and not by the system integrator. In practice, though, most enquiries go out with no category marked on the drawing. The integrator then either plays safe and quotes every location as if it were category A, or reads the silence as no requirement and quotes a torque-control-only configuration. Bids from the same drawing come back nearly a factor of two apart.

So three things must be settled on paper before an RFQ goes out. First, assign a category of A, B or C to each of the four fasteners. Second, for the A and B locations, decide how many years the fastening waveform must be retained linked to the individual product unit. Third, decide whether that retention lives locally on the equipment or in a higher-level system. The model line’s investment breakdown carries 600,000 baht for “verification (fastening data collection and higher-level system integration)” because that third decision maps straight onto money.

Finally, back to the question of whether a robot should do the screwdriving at all. The reason quarterly shipments for threaded fastening stay at 43 units is that on most shop floors a special-purpose machine is the more rational answer. Three questions separate the cases. Is tightening done from one direction or several? Is there one product variant or many? Is workpiece posture fixed or does it change? If all three fall on the simple side, building a single-axis automatic driver unit rather than an articulated robot will be cheaper, faster and more stable. The terms for ordering a special-purpose machine are set out in the guide to ordering special-purpose machine design and build. Conversely, if you tighten from several directions, or if changeovers are frequent, the robot’s flexibility outweighs its cost.

The money goes into “making parts mate”, not into the robot

Now translate all of this into money. The model line’s capital cost is presented in two configurations — base and lean. The lean configuration is what you get if you narrow the product mix, tighten the tolerances on the part side, and clean up part feeding. All amounts are in baht and all are assumptions.

The first thing to read in this table is the arm’s share. In the base configuration the robots are 11.7% of the total, and even in the lean configuration only 14.6%. The statistic in the previous chapter — assembly taking 4.7% of all-application shipment value — is a different quantity at a different scale, but the tendency it revealed, that in assembly the arm is relatively cheap and cost migrates to everything around it, shows up in exactly the same shape inside a single line’s quotation. Assembly automation cost is not the cost of buying a robot. It is the cost of making parts mate.

The next thing to read is where the 1,950,000 baht gap between the two configurations comes from.

What deserves attention is that none of these three items moves through robot selection. What moves them is a production-side decision to narrow the product mix and a design-side decision to tighten tolerances. Cutting capital cost by twenty percent is not produced by a negotiation between purchasing and the equipment maker. It comes out of production engineering, design and purchasing sitting at the same table and cleaning up the product mix and the drawings.

At the same time, some items do not move regardless of configuration. Locating fixtures and pallets at 900,000, fastening tools at 850,000, verification at 600,000, and safety, frames, panels and installation at 1,250,000 are identical in both. These are fixed costs of “doing this process at all” and narrowing the mix does not make them disappear. Together they are 3,600,000 baht, close to half of the lean configuration’s 7,850,000 baht total. This is why assembly automation is hard to structure as a “start small” project. As long as the same process stays in scope, cutting units or variants leaves fixtures, safety frames, panels and installation exactly where they were.

Three procurement practices follow. First, require the quotation broken into the eight items above. A single lump-sum bid gives you nothing to compare. Second, run the comparison on the seven items excluding the arm. Third, for part feeding and the compliance mechanism, require a conditional second figure showing how far the price falls if the mix is narrowed in a specified way. Simply specifying these three as the RFQ format transforms how comparable your bids are.

Payback — where assembly automation works at Thai labor rates

With costs established, calculate benefit and payback. The baseline is fixed at “current state, two shifts, six people” and does not change partway through.

The arithmetic convention needs stating explicitly. Capital cost goes only into the numerator of the payback calculation. Annual running cost does not enter the numerator, it is subtracted from annual benefit. The 240,000 baht of annual running cost above is not included anywhere in the eight capital items of the previous chapter. Stacking both the 9,800,000 baht capital cost and the 240,000 baht running cost on the numerator side is a double count, and it is a common error in internal documents. The numbers in this article were written after confirming that the sum of the investment breakdown matches the numerator of the payback formula, and that running cost appears exactly once, on the benefit side.

Case 3 is a conditional scenario. It holds only if demand grows to the point where a third shift becomes necessary. Its increment has exactly two components — the benefit of covering with zero people what would otherwise require 3 additional operators, or 540,000 baht, to staff a third shift; and a deduction of 120,000 baht for the increase in running cost. Adding 540,000 baht to the base net of 815,000 baht and subtracting 120,000 baht gives 1,235,000 baht. It matters that quality benefits are not scaled up for the third shift. Scaling them would improve the number, but it would count the same demand twice. Unless demand grows, the 6.36 years of case 3 never appears.

And here is the number this article most wants to convey. Taking labor reduction alone, annual net is 720,000 minus 240,000 of running cost, or 480,000 baht. Payback is 20.42 years in the base configuration and 16.35 years even in the lean configuration. Both exceed the useful life of the equipment. At Thai labor rates, assembly automation does not pay back on labor reduction alone.

The reason is already in the numbers from the assumptions chapter. At a fully loaded rate of 90 baht per hour, or 50 baht per hour on a minimum-wage basis, removing one person yields only 180,000 baht of annual benefit. In a Japanese plant, the same one-person reduction generates several times that benefit, and a payback table built on labor reduction alone stands up. The same equipment, the same effect and the same formula flip to the opposite conclusion because the labor cost in the denominator is different. Bring an investment hurdle approved at the Japanese parent plant to Thailand unchanged and virtually every assembly automation project will look unprofitable.

So what makes it work? Quality benefit. The 245,000 baht of reduced response cost for escaped fastening defects plus the 90,000 baht of reduced scrap from insertion defects total 335,000 baht, close to half of the 720,000 baht labor reduction. Without that quality benefit, payback exceeds 16 years in any configuration. The sensitivity analysis confirms it.

Sensitivity analysis has its own convention. Move only the items the broken assumption actually touches, and never apply a blanket factor to all benefits. Discounting labor reduction as well inside the scenario where quality benefit goes to zero multiplies two unrelated uncertainties together and drives the conclusion excessively negative. Likewise, in the scenario where one fewer person comes out than expected, only the labor line moves, from 720,000 to 540,000. Quality benefit does not depend on headcount, so it is left alone.

Now land these numbers in practice. Since payback runs to double digits, assembly automation cannot be approved on payback period alone. Inflating the quality assumptions to get an approval through is the worst possible response. Instead, put three further items alongside the payback figure. First, the risk that on products containing category A or B joints, the absence of records itself becomes a failure to meet a customer’s supply conditions. Second, the outlook for recruiting and holding operators for the second and third shifts. Third, the actual recorded cost of handling a single escaped defect. Most plants have real data for the third. Simply replacing this article’s assumed 350,000 baht with your own recorded figure changes how persuasive the payback table is.

Five questions that separate processes suited to automation from those that are not

These are screening questions to apply before the investment study. If you cannot answer all five, it is too early to issue an equipment enquiry.

Question 1 — can the mating part’s tolerance be tightened on the drawing

Compare the current drawing tolerance with the measured scatter of parts actually being delivered. If the measured scatter is wider than the drawing, you have a supplier process capability problem first. Tightening a drawing that is not being met anyway achieves nothing. If the answer here is “no, it cannot be tightened”, budget on the assumption that the difference is absorbed by a compliance mechanism.

Question 2 — can parts be presented in the same posture every time

Do they come out oriented from a feeder, or arrive loose in a box? Parts whose posture is undefined make feeding cost jump. This is why part feeding was the largest variable in the previous chapter’s gap analysis at 1,100,000 baht. Depending on the part geometry, changing the part shape can be cheaper than engineering the feed.

Question 3 — can the machine detect a failure

Check whether the failure mode of that process is distinguishable through insertion force, displacement, torque or rotation angle. Automate a process whose failure you cannot distinguish and defects flow undetected to downstream operations. There are real cases where the operator used to catch it by feel. If you cannot identify the physical quantity that corresponds to that feel, either take the process out of scope or add 100% inspection downstream.

Question 4 — how many variants and how often do you change over

Count the number of variants and the number of changeovers per day. Whether a changeover means swapping fixtures or only switching programs makes a large difference to real utilization. As the previous chapter showed, whether the mix can be narrowed moves capital cost itself by twenty percent. This has to be agreed with production planning first.

Question 5 — does recording actually reduce the cost of escaped defects

Confirm that recording fastening and press-fit waveforms genuinely lowers cost. There are mainly two routes — scrap reduction from not passing defects downstream, and reduced response cost because root cause is isolated faster when something does escape. Check whether your company holds actual figures for these two before the investment study begins.

Issues specific to Thailand and ASEAN

Install the same equipment in Japan and in Thailand and you will not reach the same conclusion. Here are the points where the answer diverges.

Tolerances on locally sourced parts are the biggest issue. A supplier the Japanese parent plant has worked with for decades will consistently produce inside the drawing tolerance. Parts switched to local sourcing in Thailand may swing across the full width of the tolerance band while still conforming to the drawing. The drawing is identical, but the scatter of parts that actually arrive is not. That difference hits layer 3 of the stack described earlier, head on. Most of the reasons that equipment which worked in Japan does not work in Thailand live here. The countermeasure is to collect measured data on incoming parts in parallel with the equipment study. Measure at least several hundred pieces, ideally across multiple lots, and feed that measured width into the equipment specification as an input.

Labor cost levels are as described earlier. The Bangkok minimum wage rose to 400 baht per day effective 1 July 2025, up from 372 baht. In the four provinces and one district where manufacturing is concentrated, 400 baht has applied since January 2025. Converted over 8 hours that is 50 baht per hour, and even on a fully loaded basis including statutory benefits and overtime, the rate is 90 baht per hour. At this level the conclusion of this article is that labor reduction alone does not pay back, and the justification for the investment has to shift to quality and to workforce availability.

Investment incentives also belong in the decision. The Thailand Board of Investment’s statistics for the first quarter of 2026 record applications in the machinery, automation and robotics sector of 8,081 million baht across 38 projects. Applications under the smart and sustainable industry measure, which targets productivity improvement at existing plants, numbered 61 projects worth 7,071 million baht. On a project like assembly automation where payback runs to double digits, whether an applicable incentive exists can decide the investment. Check the eligibility conditions and application timing before the equipment specification is frozen. Discovering after the specification is locked that it does not meet the incentive conditions leaves you choosing between changing the equipment and abandoning the application.

Maintenance and commissioning skills are another difference from Japan. Assembly cells stop more often than conveyance equipment, and recovery requires judgment. If nobody has decided in advance who makes which decisions when a force/torque sensor fault or an expired tool calibration comes up on the night shift, the line stays down until morning. Preparing the operator screens, alarm wording and daily check procedures in Thai is mandatory. Procedures available only in Japanese and English do not function on the night shift. The division of responsibility and the criteria for selecting an integrator are covered in detail in how to select a robot system integrator.

Environmental conditions are easy to overlook. In a building without air conditioning, the daily temperature swing shows up as thermal displacement in the arm. High humidity in the rainy season changes the behavior of adhesives and powders. In dusty processes, force/torque sensors and the moving parts of feeders are affected. Rather than reusing a Japanese equipment specification verbatim, measure the temperature, humidity and dust conditions at the installation location and write them into the specification.

Ten items to settle on paper before you issue an RFQ

Here is everything above, condensed into the items to decide before an enquiry goes out. Fill in these ten and your bids become comparable.

Filling in this table is not something an equipment maker can do for you. The buyer fills it in. Measured tolerances, fastening categories and baseline recorded figures in particular exist only inside your own company. Put differently, if you issue an enquiry holding those three, the scatter in what integrators assume drops sharply and the bids finally sit on common ground.

Frequently asked questions

How much does it cost to automate assembly with robots

For the model line in this article — 1 press-fit, 2 insertions and 4 threaded fasteners at a 30-second cycle — the estimate is 9,800,000 baht for the base configuration and 7,850,000 baht for a lean configuration with a narrowed product mix and tightened tolerances. Both are assumptions. What matters is the composition rather than the total. Robot arms are only 11.7% of the base configuration and 14.6% of the lean one. Most of the cost is part feeding, fixtures, the compliance mechanism, fastening tools, verification, installation, and integrator engineering hours. When you take quotations, require them broken into those seven items plus the robot arms, eight in total.

Should screwdriving be done by a robot or a special-purpose machine

If tightening is from one direction, on one variant, with fixed workpiece posture, a single-axis automatic driver unit will be cheaper, faster and more stable. The reason quarterly shipments for threaded fastening in the Japan Robot Association statistics stay at 43 units, 0.08% of total shipments, is that this judgment is being made widely on real shop floors. Conversely, if you tighten from several directions, or if changeovers are frequent and you do not want to own many fixtures, an articulated robot’s flexibility outweighs its cost.

Can press-fitting be automated with a robot

Yes, but you need a separate mechanism to generate the insertion force and a means of recording load and displacement. Press-fitting is irreversible, so judging pass/fail purely on whether the specified load was reached will pass a defective part that was pushed in at an angle. Reading load and displacement together as a waveform is the minimum condition. Also, consolidating press-fitting, fastening and insertion onto a single machine concentrates all the monitoring in one place and makes stoppages cascade. Split the processes, or design the monitoring systems to be independent.

Does a robot always need a force/torque sensor

Not necessarily. If the mating hole has a chamfer or taper and the part-side tolerance is tight enough, a passive compliance mechanism that gives the end effector mechanical softness is sufficient. In the model line estimate, 2 active force/torque sensor sets are placed at 700,000 baht and a lean configuration weighted toward passive compliance at 300,000 baht, a difference of 400,000 baht. The order of decisions is to take measured part scatter first, confirm on the drawing whether a chamfer exists, and only then consider a force/torque sensor for whatever cannot be absorbed.

Should we automate a whole assembly line or a single process first

Starting with a single process looks natural, but assembly automation carries a large fixed cost that deserves attention. Locating fixtures and pallets at 900,000, fastening tools at 850,000, verification at 600,000, and safety, frames, panels and installation at 1,250,000 total 3,600,000 baht, close to half of the lean configuration’s 7,850,000 baht. Of those, fixtures and the safety, frames, panels and installation group barely shrink when you cut units or variants, as long as the same process stays in scope. Fastening tools and verification do become unnecessary if you take screwdriving out of scope, but most of the quality benefit disappears with them. A “small trial” configuration is therefore proportionally expensive. If you do start small, pick one process whose failure can be detected through a physical quantity, that has few variants, and for which your company holds recorded escaped-defect cost.

Are articulated robots suited to assembly

There are conditions where they are and conditions where they are not. They suit cases needing access from several directions, cases with frequent changeovers where you do not want to add fixtures, and cases where future process changes are expected. They do not suit single-direction, single-posture, single-variant work, where a special-purpose machine will be more stable. And even if you choose an articulated robot, as this article has repeated, the outcome is decided by the mating part’s tolerance and the stability of feeding posture, not by the robot’s repeatability.

Summary

What decides the success of assembly automation is not robot accuracy. It is the mating part’s tolerance and the quality assurance of fastening. Material handling and palletizing are “place” motions and succeed on positioning accuracy, but assembly is a “mate” motion where the target itself lives on the mating part. The structure visible in the Japan Robot Association statistics — assembly at a 12.2% share of units against a 4.7% share of value, meaning the arm is cheap and cost migrates to everything around it — reappeared in exactly the same shape inside the model line quotation, as an 11.7% arm share.

At Thai labor rates, payback on labor reduction alone comes to 20.42 years in the base configuration and 16.35 years even in the lean one, exceeding the useful life of the equipment. Including quality benefit, it is 12.02 years for the base configuration and 9.63 years for the lean one. Only with the added condition of demand growth requiring a third shift does it reach 6.36 years. Assembly automation is therefore not a project to be approved on payback alone. Judge it alongside the risk that fastening records become a supply condition, the outlook for securing operators, and the recorded cost of handling escaped defects.

And there is work to do before that judgment. Take measured scatter of the mating part across multiple lots. Assign a fastening category to each individual screw. Produce recorded figures for your current escaped-defect cost and scrap cost. These three exist only inside your own company, and no integrator can supply them. Issue an enquiry holding all three and, for the first time, your bids become comparable.

TOMAS TECH supplies Japanese-owned plants in Thailand with production management systems combined with shop-floor automation. On assembly automation, we are happy to help simply with the triage — which processes are suited to automation and which ones need the part-side tolerances cleaned up first. It is fine if you do not have measured data yet. Get in touch through the contact page.

References

• World Robotics 2025 Executive Summary (IFR)
• Manipulator and Robot Statistics, Orders, Production and Shipments by Application, Oct-Dec 2025 (Japan Robot Association)
• Manipulator and Robot Statistics, Full-Year 2025 Results (Japan Robot Association)
• VDI/VDE 2862 Explained (DEPRAG)
• ISO 5393:2017 Rotary tools for threaded fasteners — Performance test method
• Bangkok Minimum Wage Raised to 400 Baht per Day (JETRO)
• Thailand Board of Investment — Investment Application Statistics for Q1 2026