When a plant starts looking at screw fastening automation, the first requirement is almost always the same. Set a target torque on the electric driver, pick up the completion signal in the PLC, stop the line on NG. That much always appears in the quotation. Then the cell goes live, and parts that the equipment judged OK at the specified torque turn up downstream, or at the customer, with screws that never seated. This article breaks down why that happens on electronics and automotive component assembly lines, and how to specify an automatic screwdriving cell so the defect is caught inside the process rather than after it.
Screw fastening automation does not end with torque control
When an operator drives a screw by hand, they process several signals at once without thinking about it. The feel of the bit reaching the seating face, the moment the torque suddenly builds, a sound that is not the usual one, the catch of a screw starting to go in crooked. None of this is written in the work instruction, yet it is what makes the judgement “this screw is wrong” possible.
When the operation is automated, only a fraction of that is handed over to the machine. A general purpose automatic screwdriving unit typically judges two things. Did torque reach the set value, and did it get there within the allowed time. The shape of the torque build-up and the position of the screw head simply do not exist on the equipment side unless someone puts them in the specification.
Automation removes the path by which a human notices something is wrong
Defects do not usually escape because equipment breaks. They escape because the equipment judges them normal and lets them through. With a human at the station, the one screw in a hundred that feels wrong gets someone to stop and look. On an automated line there is nobody to catch that feeling. Any abnormality that is not in the judgement criteria flows straight to the next process.
That is the starting point for why torque control alone is not enough. Torque control moves exactly one of the signals the operator had into the machine and leaves the rest behind. The wider question of automating a whole assembly sequence is covered in why part tolerance, not robot accuracy, decides assembly automation. Screwdriving on its own has pitfalls that the assembly-wide view does not surface.
There are only three signals available for judgement
In practice a screwdriving station can acquire three pieces of information.
- Torque. The rotational force the driver delivered, with completion declared when the set value is reached
- Angle. How far the screw turned after seating was detected, which estimates clamp load more reliably than torque alone
- Height, meaning the Z axis coordinate. Where the screw head finally came to rest, read from the position feedback of the robot or the electric slide
Most quotations include only the first of these. Whether the second and third go into the specification is what decides whether the screw float described below can be detected at all. The difference is not in the parts or the drawing. It is in the requirement written before the order is placed.
How screw float slips past torque control
Screw float means the head never made contact with the seating face
Screw float is a defect in which the screw head does not sit flush against the seating face of the workpiece and stops with a small gap remaining. It can look tightened to the eye, and a finger may not feel it. Yet the intended clamp load was never generated, so it shows up later as loosening under vibration, a waterproof seal that does not seal, or rising electrical contact resistance.
What makes it awkward on an automated line is detection. Japanese factory automation guidance treats screw float as a defect that goes unnoticed on automated lines precisely because there is no operator to sense it. On a manual line somebody eventually says the screw looks slightly proud. An unattended station has nobody in that role.
An electric driver cannot separate the force that advances the screw from the force that seats it
The torque generated during fastening splits into two components. The force consumed driving the screw through the female thread, which is running friction, and the force applied after the head contacts the seating face, which is the actual clamping. The designer wants to control the second one. The torque sensor in the driver measures the sum of both.
So if running friction becomes abnormally high for any reason, the total torque reaches the set value even though the head has not touched the seating face. The driver treats that as normal completion and issues the OK signal. That is what screw float is. The sensor is not broken and the setting is not wrong. The quantity being measured is simply not the quantity being managed.
Three causes of screw float
The causes fall into three groups. None of them is exotic, and all of them occur routinely in volume production.
| Cause | How it appears on the floor | Behaviour of the torque value | Effective countermeasure |
|---|---|---|---|
| Cross threading | Robot positioning error or workpiece tolerance shift makes the screw enter the hole at an angle and lock | Torque spikes partway through and reaches the set value after a short rotation angle | Workpiece locating jig, compliance mechanism at insertion, Z axis end position monitoring |
| Trapped debris | Chips or dirt from an upstream process sit in the tapped hole and stop the screw partway | Resistance rises before seating and torque climbs gradually | Air blow of tapped holes, chip control upstream, Z axis end position monitoring |
| Part to part variation | Screw dimensional tolerance or insufficient tap depth means the screw never reaches full depth | Hard to tell from a good part, sometimes indistinguishable from the waveform alone | Incoming inspection, in-process assurance of tap depth, angle method in combination |
What all three share is that torque reaches a normal value while the screw head position does not. As long as only torque is being watched, none of them can be detected regardless of the underlying cause. Before choosing a countermeasure for each cause, the process has to be made capable of detecting the condition at all.
The countermeasure is a double check of Z axis coordinate and torque
The effective answer is to put both torque and height into the judgement criteria. If the same screw goes into the same part number, the height of the screw head at correct seating falls within a fixed band even allowing for scatter. That gives three rules.
- Torque reached the set value and the Z axis stopped inside the allowed band, so the result is OK
- Torque reached the set value but the Z axis stopped higher than allowed, so the result is NG for suspected screw float
- The Z axis is inside the band but torque never reached the set value, so the result is NG for under tightening
Adding this alone stops most cross threading and trapped debris inside the process. The important point is that the Z axis value usually already exists on the equipment. If the thrust axis is built from a robot or an electric slide, position feedback is available and can often be folded into the judgement with no additional hardware. Conversely, failing to state the requirement at quotation time and then starting with a control software modification after go live is far more expensive.
Note that the allowed Z band cannot be derived from the drawing alone. Plate thickness tolerance, the presence of a washer, and paint or plating thickness all affect the head height, so the band has to be drawn from measurements taken under near production conditions after looking at the distribution of good parts. Setting the limits from a desk calculation produces over detection, good parts get stopped, and the floor ends up disabling the judgement.
Three ways a screwdriving robot project fails
When an automated screwdriving cell ends up unused, the reason usually falls into one of three groups. Cases where the task was technically impossible are rare. Cases where a decision that belonged in the design phase was never made are the norm.
Failure 1, losing the cycle time race against a person
“A person is faster.” It is the most common verdict after installation. Guidance on collaborative robot screwdriving puts cycle time delay at the top of the list of failure factors.
The reason is that fastening a screw is not a single motion. Every screw requires picking the screw onto the bit, confirming the pick, moving to the fastening position, locating, driving, and confirming completion, in that order. A person overlaps several of these and prepares the next screw while moving. A robot generally executes them in series.
Three time components are routinely missed at design time.
- Communication time with upper level equipment. Signal exchange between the PLC and the driver controller accumulates once per screw
- Screw feeder cycle time. The robot can arrive to pick and still have to wait for the next screw to be presented
- Recovery time from a safety stop. When a collaborative robot stops because a person came close, restarting and returning to the home posture takes time
Calculate with the catalogue speed of the robot alone and the measured result can be 1.5 to 2 times the plan. On a small assembly with four or six screws, that gap becomes exactly the conclusion that a person is faster. Build the per screw time from communication and waiting at the quotation stage, and write the on-machine verification condition into the contract.
Failure 2, minor stoppages in screw supply
The second is the screw feeder. Screws that do not come out, or two that come out together and jam, several times a day, put a person back at the station to clear them, and the purpose of the automation is lost.
The cause is less often a failure of the feeder itself than a mismatch between screw geometry and feeder, or unmanaged variation in air pressure and other peripheral conditions. Even within M3, head shape, shank length, whether a washer is pre-assembled, and surface finish all change the behaviour. Equipment can be delivered exactly to specification and still be unproven on your own screws until it is tried with the real parts.
Stoppages of this kind last tens of seconds each, so they barely register in the daily downtime figure and are invisible if only the availability number is watched. How to capture them is covered in where OEE actually hides minor stoppages. At a screwdriving station in particular, counting supply related stoppages separately from other stoppages is what makes the effect of a countermeasure measurable.
Failure 3, quality defects
The third is quality. Cross threading, screw float, and cam out, meaning the driver bit slipping out of the screw head. Cam out is caused by bit wear, insufficient thrust force, or an inclined axis, and it rounds out the recess in the screw head, which also makes later disassembly and rework difficult.
All three come back to the absence of Z axis monitoring and to positional error. Put differently, if locating design and height monitoring are decided first, most of this can be eliminated on paper.
| Failure pattern | Typical symptom | Decision required at design stage |
|---|---|---|
| Cycle time delay | Measured time runs 1.5 times the plan or worse and the work goes back to manual | Per screw time build-up including communication, feeder cycle and safety stop recovery |
| Screw supply stoppages | Jams and empty strokes several times a day with a person stationed to clear them | Feed trial with the real screws, allowed air pressure range, re-validation procedure when a screw part number is added |
| Quality defects | Screw float, cross threading and cam out found downstream or at the customer | Whether Z axis monitoring is included, thrust force control, bit replacement interval and trigger |
Specify the cell in four layers
Quotations diverge mostly because nobody has decided which layers belong to whom. A screwdriving cell separates into four layers, and looking at it that way exposes the missing requirements.
Layer 1, supply
Everything from taking one screw out of the feeder to handing it to the bit or chuck. What has to be decided here is the list of screw part numbers in scope and the changeover method. If one line handles several screws, decide whether there will be several feeders or a changeover mechanism, and how many minutes the changeover takes.
Layer 2, locating
Everything that establishes the relative position of workpiece and bit. What matters is not only robot repeatability but how much tolerance the workpiece jig absorbs. A screw hole is a single point on the drawing, but the real part carries scatter from moulding or machining. Decide from measured data whether that is absorbed by a compliance mechanism, a floating joint, or vision correction. Where the gripping method is part of the question, three gripper types and how to calculate gripping force is a useful reference.
Layer 3, fastening
The electric driver, torque sensor, thrust axis and bit. Robot mounted electric drivers are available with torque accuracy of roughly ±2% to ±5%, and some record torque and angle curves at 1,000 samples per second or more. Robot mounted screwdrivers are on the market covering torque ranges such as 0.15 Nm to 5 Nm. Decide the torque range and accuracy demanded by the clamp load you need first, then select a model that meets it.
Layer 4, judgement and recording
The judgement logic and the record. This is the layer most often dropped from a quotation and the most expensive to add later. Write the judgement criteria, the fields recorded, the storage location, the retention period and the retrieval method into the specification.
| Layer | Main components | Requirement most often missing from the quote |
|---|---|---|
| Layer 1 supply | Screw feeder, transfer tube, vacuum bit or chuck | Feed trial across every screw part number, changeover procedure and time |
| Layer 2 locating | Robot, electric slide, workpiece jig, vision sensor | Where workpiece tolerance is absorbed, measured scatter of hole position |
| Layer 3 fastening | Electric driver, torque sensor, thrust axis, bit | Required torque range and accuracy, need for angle measurement, bit life |
| Layer 4 judgement and recording | Z axis acquisition, judgement logic, storage destination | Definition of the criteria, retention period and search method |
Specifications that reach layer 4 are not common. Yet for products such as automotive components or medical devices, where fastening records may be requested afterwards, layer 4 can be the actual purpose of automating at all. Designing the record at unit level is covered in designing automotive parts traceability for IATF 16949 and customer audits.
Estimating cost and payback
The robot price and the cell price are different numbers
On industrial robot cost, Japanese system integrator Sanmei Kiko states that robot and cobot bodies are available from around several million yen. In a real automation project, however, peripherals such as hands, sensors, control panels and conveyors are required, along with safety fencing and the engineering work of design, build, installation, teaching and commissioning. The same guidance gives figures by scale of automation, roughly 5 to 15 million yen for single process automation such as transfer, feeding and unloading, roughly 10 to 30 million yen for a mid-sized system covering assembly, inspection and process support, and 30 million yen and above for full line automation linking several machines.
A screwdriving cell usually sits between single process automation and a mid-sized system, so in practice it belongs in the 5 to 30 million yen band. Put another way, the catalogue price of the robot and the amount actually paid differ by a factor of several. For a screwdriving cell the gap breaks down roughly as follows.
| Cost item | What it covers | Easy to overlook |
|---|---|---|
| Robot body | Arm, controller, teach pendant | The catalogue price is the body only and will not run on its own |
| Fastening unit | Electric driver, driver controller, thrust axis | Torque range, accuracy and angle capability move the price band |
| Screw feeder | Feeder, transfer tube, pick up section | Several screw part numbers mean several feeders or a changeover mechanism |
| Workpiece locating | Jig, pallet, stopper, clamp | The structure that absorbs tolerance. Cutting it increases float and cross threading |
| Judgement and recording | Z axis monitoring, torque and angle logging, upload to the host system | Frequently dropped from the quote and costly to retrofit as a control change |
| Safety | Fencing, light curtain, emergency stop, risk assessment | Contact force assessment is still required with a collaborative robot |
| Engineering | Design, build, installation, teaching, commissioning, witness testing | A large share of the total cell cost |
How much of this breakdown sits inside a given integrator’s quotation varies by company. That is why the same requirement can produce quotations differing by more than a factor of two, and how to level the comparison is set out in the five scope boundaries behind a 2-3x quote gap.
Payback cannot be explained by labour substitution alone
Build the payback case for screwdriving automation purely on headcount and the numbers usually fall short, because a screwdriving station occupies few people and the labour content removed is small. At least three further effects belong in the case.
- Bringing missed and under tightened screws close to zero, which removes rework and customer complaint handling
- Holding a fastening record for every unit, which narrows the scope of investigation when a problem does occur
- Removing dependence on operator skill, which shortens ramp-up when headcount is added or a model changes
The second and third are often struck from the capital request because they are hard to monetise. Yet the sorting and reshipment cost of a single customer complaint frequently exceeds the cost of one cell. Even if they are excluded from the payback calculation, they belong on the table as decision criteria.
Collaborative robots can sometimes dispense with fencing, compressing floor area and part of the engineering, so total system cost tends to be lower than an industrial robot equivalent. Removing the fence still requires contact force assessment and a risk assessment, and that step cannot be skipped. The preconditions and cost structure are set out in collaborative robot implementation cost and approach.
For Japanese owned plants in Thailand, labour cost is not the whole argument
In Thailand as of 2026 the statutory minimum wage sits between 337 and 400 baht per day depending on the province. Using that figure directly as the justification for automation is premature. The effective labour cost a production engineer should work with is not the minimum wage itself but a number that includes statutory benefits, shift allowances, recruitment and training cost, and the frequency of retraining caused by turnover.
Thailand is also seeing labour shortage and reduced utilisation in some sectors at the same time. A plain argument that people cannot be hired so the work must be automated does not hold on a process whose loading is uncertain. On screwdriving, the argument that lands is quality and traceability rather than labour. Identifying which processes can pay back from the labour side is handled separately in spotting the processes automation can pay back as Thai labour costs rise.
Five steps to implementation
Step 1, take inventory of candidate processes
Start by choosing which of your fastening operations suits automation. The criteria are screws per unit, the number of screw part numbers, changeover frequency, and the consequence of a fastening defect escaping. A process with few screws and many part numbers ranks low because feeder changeover dominates. A process driving many identical screws where a defect is costly is the first candidate.
Step 2, prove the screw and feeder combination
Prepare the real screws and run a feed trial. Being inside the catalogue range does not guarantee that they will present reliably one at a time. Make this trial an explicit evaluation item before the order and keep the result in writing, and responsibility for minor stoppages after go live becomes far clearer.
Step 3, define the judgement criteria
Decide which of torque, angle and Z axis will be used, and draw the upper and lower limits from measurement. What matters here is less the numbers themselves than the operating rule of who looks at which data and how it gets revised. Criteria need review after volume production starts, so fix the change procedure and the authority for it in advance.
Step 4, decide where the record goes
Decide where fastening data is stored. Internal memory in the driver controller has a capacity limit and will overwrite the oldest data. Decide, from customer requirements and internal rules, whether the data goes to a host system keyed to a unit serial number or lands on a server as daily files, and how many years it must be kept.
Step 5, plan maintenance and bit management
Bits are consumables. As they wear, cam out increases and screw heads get rounded. Set the replacement interval by screw count, and decide who replaces them and where the replacement record is kept. Add feeder cleaning intervals, air filter changes and periodic torque calibration to the maintenance plan on the same basis. Automation does not reduce maintenance work so much as change what kind of work it is.
Frequently asked questions
How much does screw fastening automation cost?
A robot body on its own can be had from around several million yen, but configured as a usable screwdriving cell, including peripherals, safety and engineering, the guide range is roughly 5 to 15 million yen at single process level and 10 to 30 million yen for a mid-sized system covering assembly and inspection. The range is wide because the number of screw part numbers, the difficulty of workpiece locating and the recording requirement all change the configuration. When comparing quotations, always check how far the judgement and recording layer is included.
Why is torque control on its own not enough?
Because the torque an electric driver measures is the sum of the force used to advance the screw and the force used to clamp after seating. If friction rises partway through because of cross threading or trapped debris, the set torque is reached even though the head never touched the seating face, and the equipment declares normal completion. Preventing that requires monitoring the Z axis, meaning the height at which the screw head stopped, alongside torque.
Should we choose a collaborative robot or an industrial robot?
If cycle time demands are tight and no human needs to enter the area, an industrial robot has the speed advantage. If the cell is being retrofitted to an existing line and has to run in the same space as operators, a collaborative robot becomes the candidate. Even then, contact force assessment and a risk assessment cannot be skipped. Whether the fence can be removed is decided by the risk assessment result, not by the model selection.
How many screws per unit justify automation?
No single screw count draws the line. More screws mean more labour removed, but many screw part numbers introduce feeder changeover that cancels the gain. In practice the strongest candidates drive four or more screws of the same part number, change model no more than once a day, and carry a high escape cost for a fastening defect. A process with two screws per unit across five part numbers is better served by improving jigs and procedures first.
How do we reduce minor stoppages in screw supply?
Running a feed trial with the real screws before ordering is the single most effective action. Beyond that, state the allowed air pressure variation in the specification and fit covers that keep debris out of the feeder area. Operationally, log supply related stoppages separately from other stoppages so that frequency can be seen per screw part number, which leads upstream to consolidating part numbers or revisiting surface finish.
Can we reuse our existing handheld electric drivers?
Usually not as they are. Handheld drivers are designed on the assumption that the operator supplies the thrust force, and they typically lack the thrust axis control needed when robot mounted, together with external start signals and judgement signal input and output. This is even more true where torque and angle records are required, so the practical answer is a driver and controller designed for robot mounting.
Summary
What derails screw fastening automation is neither robot accuracy nor driver performance. It is the design of the information used for judgement. Equipment that watches only torque passes screw float caused by cross threading or trapped debris as normal completion. Put a double check of torque and Z axis coordinate into the specification, prove the feeder against the real screws, and decide where the record ends up. Fix those three before placing the order and you avoid starting the post-installation phase with a software modification.
On cost, compare the breakdown of the whole cell rather than the price of the robot. And evaluate payback on more than labour substitution, including escape prevention and the narrowing of investigation scope that a complete record provides, which fits the reality of a fastening process far better.
TOMAS TECH supports Japanese owned plants in Thailand on single process automation such as screwdriving, from organising the requirement specification through designing the judgement criteria to connecting fastening data into the production management system. It is perfectly fine to talk at the stage where you are still deciding which process to tackle first, and if you can describe the current process and the problem, we can help you organise the decision criteria. Contact us here.
References
- Causes of screw float and detection on automated lines, Kojo Automation Navi, in Japanese: kojo-automation-navi.com
- Common failures in collaborative robot screwdriving automation, Robot SIer, in Japanese: robot-sier.com
- Guide figures for industrial robot implementation cost, Sanmei Kiko, in Japanese: sanmei-kikou.co.jp
