Robot Vision Implementation Cost and ROI in Thailand — Supply, Accuracy and Lighting Decide It

“Add robot vision and the robot will pick the parts however they happen to be lying.” Most enquiries open with some version of that sentence. And when the cell is commissioned, the thing that stalls it is almost never the camera. How the parts arrive at the cell, whether the lighting holds steady, and how much error the robot and the gripper can absorb between them — none of these appear on a single line of the camera quotation. This article breaks down what moves the cost by an order of magnitude, and what has to be true for the investment to pay back in a Thai factory, including a worked model built entirely on assumed figures.

Robot vision is not a device that looks — it is a mechanism that hands over coordinates

Some vocabulary first. Factory cameras do two fundamentally different jobs, and lumping both under “image processing” leads to a badly written requirement specification from the very first draft.

The first job is the camera that judges pass or fail. Is there a scratch, is the print legible, is the component present. The output collapses to a single value, OK or NG. It replaces an inspection a person used to do by eye, and the work is mostly about building the judgement criteria and agreeing the limit samples. We cover how to pin that side down in how to decide whether AI visual inspection is right for your line.

The second job, and the subject of this article, is the camera that hands over coordinates. The output is not OK or NG but a set of numbers saying “this workpiece exists at this point in the robot coordinate system, in this orientation”. In 2D that means X and Y on a plane plus a rotation angle. In 3D, height and tilt are added, giving up to six numbers — three degrees of freedom in position and three in orientation. The robot receives those numbers and moves.

The difference shows up in how failures present themselves. When a pass-fail camera gets it wrong, a defect escapes or a good part is scrapped. When a coordinate camera gets it wrong, the robot grabs at empty air, fails to close on the workpiece, or in the worst case collides with a fixture or with surrounding equipment. The first is an error in a result. The second is an error in a motion, and the line physically stops. That is why the requirement specification for robot vision must not be written around “what recognition rate will it achieve” but around “will the robot reliably pick the part using the coordinates it is given”.

What sits between “detected” and “picked”

Go to a demonstration and you will see the workpiece outlined in a coloured box on screen, with coordinate axes overlaid to show its orientation. That is detection. Between detection and a successful pick there are at least four more gates.

• Can the position found in the camera coordinate system be converted into the robot coordinate system (hand-eye calibration)
• Can the robot actually reach the converted coordinate (absolute accuracy, working envelope, singularities, whether the orientation is achievable)
• When the gripper goes there, does it clear the surrounding parts and the bin walls (collision-free path planning)
• At the reached position, can the gripper absorb the dimensional variation of the workpiece plus whatever positional error is left (gripper tolerance)

A demonstration shows you the first one at best. The remaining three can only be verified with your own workpiece, your own gripper and your own supply condition. This is precisely why starting a robot vision study from “which camera should we buy” always turns into the long way round.

Where the line falls between processes that need only 2D and processes that need 3D

A 3D camera costs more than a 2D camera, the processing is heavier, and commissioning takes more engineering hours. So the first decision is “do we need 3D at all”, not “which 3D camera”.

The rule for drawing that line is simple, and it comes down to how many numbers the robot has to be given. If parts sit in a single layer on a plane, at a known and constant height, with orientation varying only by rotation within that plane, then three numbers are enough — X, Y and rotation angle. That is obtainable in 2D. As soon as parts overlap, tilt, vary in height from piece to piece, or arrive face-up and face-down mixed together, height and two tilt axes become unknowns as well, and depth information from 3D is required. That is exactly why bin picking needs 3D.

Do not overlook the “2.5D” middle ground

What you find on a real shop floor is rarely either extreme. Parts are laid out in a single layer on a tray, but come in a few different thicknesses. Parts flow along a conveyor, but occasionally two overlap. In cases like these it is often enough to obtain the planar position in 2D and use a displacement sensor or a simple height measurement to tell only which layer a part is on. Before building a full 3D recognition model, always check whether you can eliminate just one height unknown by other means. Skip this step and go straight to 3D, and the cost of the “teach” block described later climbs sharply.

What each 3D method is good and bad at

3D acquisition methods include stereo, structured light, time of flight and laser triangulation, and each has its strengths and weaknesses. Shiny metal parts tend to lose point cloud data to reflections, and black plastics absorb the light and drop out. Transparent and translucent parts are difficult for most methods. In a real evaluation, the right approach is to bring not the representative good sample but the shiniest piece, the blackest piece and the dirtiest piece, and ask to see the point cloud from those. The sample images in a catalogue were taken with the workpiece that method handles best.

Catalogue accuracy is not shop-floor accuracy — the three layers of positioning error

This is the chapter that generates the most misunderstanding, and the point most often missed at the start of a robot vision study.

Whether the robot can actually pick the part is not determined by a single “accuracy” figure. The final deviation is the result of at least three layers stacked on top of each other.

Layer 1 — catalogue accuracy is the accuracy of measuring

The accuracy quoted in a 3D camera catalogue is a measurement accuracy based on the evaluation standards for optical 3D measurement systems (the VDI/VDE family). For example, the DLP structured light camera Orbbec showed at Automate 2026 is quoted at roughly 0.07 mm VDI/VDE accuracy and 0.05 mm Z repeatability at a distance of 0.5 m. That is very high accuracy, but what it means is “a shape at that distance can be measured to that error”. It does not mean the part can be picked and inserted.

There is a further practical caution. Figures of this kind apply at a specific distance and under specific subject conditions. Change the working distance and the accuracy changes; widen the field of view and the resolution drops. When evaluating, do not accept “the catalogue accuracy” — ask for the accuracy at your working distance, over your field of view, with your workpiece.

Layer 2 — repeatability and pose accuracy are entirely different metrics

ISO 9283, the performance test standard for industrial robots, defines pose accuracy and pose repeatability separately. The former is the difference between the commanded coordinate and the position actually reached. The latter is the spread when the robot is sent to the same point repeatedly. Testing assumes the largest cube that fits inside the robot’s working envelope (the test cube) and is carried out repeatedly at several poses within it.

Industrial robots generally have high repeatability but rather less impressive absolute accuracy. And what appears in the catalogue is normally repeatability. For conventional applications where the robot returns to a taught point, that is entirely adequate in practice, because the taught point is itself a position the robot demonstrably reached.

The problem starts the moment vision is introduced. Vision hands over a different coordinate every cycle. The question becomes whether the robot can go to a coordinate it was never taught, exactly as commanded. What matters here is absolute accuracy, not repeatability. The two are sometimes said to differ by an order of magnitude, but that is secondhand information and we will not assert it as fact. The practical response is clear enough — ask the manufacturer or the system integrator for a measured absolute accuracy figure for that specific model, in writing, together with the measurement conditions. If the answer is that they cannot provide one, that itself is design information. If no measured figure is forthcoming, the error has to be absorbed at Layer 3, on the gripper side.

Note also that absolute accuracy varies with the flatness of the mounting, gearbox backlash, arm rigidity and temperature. In Thai factories, buildings without air conditioning and areas taking heat through the roof can produce a large difference in machine temperature between first thing in the morning and mid-afternoon. Coordinate drift over time is a genuinely reported phenomenon, and periodic recalibration has to be built into the operating routine as a countermeasure. That becomes a line item in the annual running cost discussed later.

Layer 3 — designing give into the gripper is cheaper than buying a better camera

Layer 3 is what lets the robot pick successfully even with the errors from Layers 1 and 2 still present, and it consists of the mechanical tolerance between gripper and workpiece. In concrete terms it looks like this.

• Put a chamfer or taper on the insertion feature so that a part slightly out of position is guided in anyway
• Build a floating mechanism into the gripper so that position self-corrects on contact
• Take the grip point on a face where “anywhere along here is fine”, so that positional error matters less
• Pick roughly once, set the part down on a regrip station at an accurate position, and pick it up again

That last one, the regrip, is particularly effective. You reduce what vision has to deliver to “accurate enough to get it out of the bin” and create the precise positioning mechanically, on a regrip station. With that design, the accuracy demanded of the camera drops a level, and so do the cost and the commissioning hours. The principle that accuracy is designed, not purchased is concentrated in exactly this point. We reached the same conclusion about how gripper and peripheral design drive the total in the cost structure of a palletizing robot installation.

What moves the cost by an order of magnitude is the supply condition (loose in a bin, semi-ordered, fully ordered feeding)

Everything so far has been about accuracy. From here on it is about cost. And the single largest driver of cost is not the camera model number but the state the workpiece arrives in.

Before choosing bin picking, ask why the parts are piled loose

Bin picking is the most spectacular technology and the biggest draw at a trade show. Photoneo, for instance, described its AI-driven 3D vision at Automate 2026 as handling delicate part picking, unordered bin picking and pallet dimensioning. As a technology it is unquestionably maturing.

But as a factory decision, the order of reasoning is usually backwards. The study begins with the assumption already in place that “the parts arrive piled loose today, so we will automate them piled loose”. Trace back why they are piled loose in the first place and you often arrive at reasons that could be changed — an upstream process convenience, a supplier’s packaging specification, an internal rule about transport containers.

If the supply side can be moved to semi-ordered, the vision requirement drops from 3D to 2D, the recognition model build disappears, and collision-free path planning becomes unnecessary. As the worked model later in this article shows, that is a difference of several hundred thousand THB in the total. Tidying up the supply is cheaper than strengthening the vision. That is the single most important practical recommendation in this article.

Where to look if you are moving to fully ordered feeding

If you can go all the way to fully ordered feeding, the conversation moves into the territory of fixtures and dedicated machines. Once orientation is fully determined by mechanical elements — part feeders, chutes, locating fixtures, dedicated pallets — vision shrinks to confirming “is the part there, is it the right way round”. For processes where the workpiece orientation can be set by a fixture, such as screw fastening and press fitting, this is the standard answer. We have summarised how to approach design and fabrication in this area in how to decide on jig design and manufacturing in Thailand.

Fully ordered feeding has limits too. In processes with many variants, large shape differences and frequent changeovers, building a fixture per variant means the number of fixtures, the storage space and the changeover time all keep growing. Past a certain number of variants and changeover frequency, switching a vision model becomes cheaper than the sheer physical volume of fixtures. That crossover point differs by process, so there is no universal answer. The practical method is to take the calculation format used in the payback chapter below — annual benefit alongside investment — and apply it to both the keep-building-fixtures option and the vision option.

Semi-ordered is the realistic answer for most factories

In practice, semi-ordered supply is where the cost-benefit lands most easily. Ask the supplier to deliver in compartmented trays, move internal transport between processes onto trays, separate parts into a single layer on the conveyor. None of these is a major capital project; most are reachable through changes to operating practice and containers. Get to that state and vision can be 2D, recognition can be simple shape matching, and commissioning shrinks to a matter of weeks.

Read robot vision cost in four blocks (see, teach, grip, connect)

Line up several quotations side by side and the difference in price looks like a difference in camera model. It is not. What gets compared in a quotation is only Block 1, while what actually moves the money is Block 2 and Block 3 — Block 1 moves too, but only as a consequence of what the supply condition dictated.

Block 1, “see” — insist that lighting and shielding are itemised

Camera body, lens, and then lighting, light shielding and the mounting frame. Be wary if the quotation says nothing more than “lighting set”. Insist on itemisation down to the type of lighting (coaxial, ring, bar, dome, flat area lights), the shielding method (enclosure, blackout curtains, hoods) and whether ambient light countermeasures are included. Lighting gets its own chapter below.

The mounting frame is easy to dismiss too. The frame holding the camera vibrates slightly, or someone knocks into it and the position shifts. When the mounting shifts, the coordinate transform that had been working collapses all at once. Put frame rigidity, and a means of detecting that a shift has occurred (periodic imaging of a reference marker, for example), into the specification.

Block 2, “teach” — hand-eye calibration is what this really is

This is the hardest block to see and the heaviest in engineering hours. At its core sits hand-eye calibration, which is the work of determining the transform between the camera coordinate system and the robot coordinate system. The camera describes the world it sees in its own coordinates. The robot moves in its own coordinates. Until the transform linking the two has been solved, no amount of measurement accuracy will get the robot to that location.

There are two configurations.

• Eye-in-hand — the camera is mounted on the robot wrist. Because the robot can be moved to change the viewpoint, one camera can look down into a deep bin or cover several locations. What you solve for is the pose of the camera relative to the tool coordinate system.
• Eye-to-hand — the camera is fixed on a frame and looks at the robot. Because the camera does not move, imaging conditions are stable and it is easy to run imaging in parallel with robot motion. What you solve for is the pose of the camera relative to the robot base coordinate system.

In either configuration, the transform is solved by collecting data at multiple poses — moving a calibration board with the robot, or touching known points with the robot TCP (tool centre point) using the TCP touch method. The engineering hours land here. Then come recognition model registration (generating a model per variant from CAD or from the physical part), grip point definition (several candidate grip points per variant), teaching and motion verification.

Ask for the Block 2 quotation to be split into an initial portion and an ongoing portion. Sign a contract based only on the figure for the first few variants and an extra quotation will arrive every time a variant is added. How many variants per year are expected, what is the price per variant, is training included so that your own people can register them. Get this on paper.

Block 3, “grip” — the gripper and the supply-side modifications

The gripper (EOAT) is often a bespoke design per process, and the cases where an off-the-shelf unit will do are limited. With vacuum, whether suction works at all depends on the surface condition of the workpiece — holes, surface irregularity, oil, dust. With a parallel chuck, the parallelism of the gripping faces and the dimensional variation of the workpiece are what matter. With magnets, material and residual magnetism become the issue.

This is where the cost of the designed-in give described in the previous chapter belongs. Floating mechanisms, regrip stations, locating fixtures. Modifications on the supply side (moving to trays, adding dividers, a separating mechanism on the conveyor) sit in this block as well. Alongside the safety costs in Block 4, this is the block most often missing from an initial quotation.

Block 4, “connect” — safety and commissioning attendance

Communication specifications with the PLC and higher-level systems, signal handshaking, handling of abnormal conditions. And then safety. For any cell people may approach — safety fencing or a laser scanner or a light curtain, the execution and recording of a risk assessment, emergency stop and recovery procedures. Using a collaborative robot does not remove safety; it only changes what is required. We deal with this in detail in the safety standards and costs overlooked in collaborative robot projects.

Commissioning attendance belongs here too. At an overseas site, that means travel, accommodation and per diem for engineers coming from Japan or a third country, and the cost grows the longer the period runs. An important point — Block 4 is in principle entirely initial investment and must not be double-counted into the annual running cost. Safety fencing, risk assessment and commissioning attendance are all one-off expenditure. In the worked model below, the annual running cost is limited to three items only — consumables, periodic inspection and recalibration, and variant additions.

Design cycle time from the worst case, not the average

The cycle time of a robot vision cell has a property that conventional robot motion does not. It does not repeat the same motion. The workpiece is in a different place every time, so the travel distance differs. Recognition time differs too, depending on the state of the bin. And every so often, the pick fails.

The motion forms a loop — image, recognition processing, path planning, move and grip, transfer and place, image again. With bin picking, the state of the bin changes with every part removed, so in principle a new image is needed each cycle. The fact that this re-imaging returns the loop to its starting point is precisely why the cell must not be designed on averages.

A cycle time model built on assumed figures

Everything below is an assumed figure. It is not data from a real machine; read it as a model showing what to think about at the design stage.

The standard total is 6.0 s and the worst case total is 9.2 s. The quotation states the former. The downstream process waits when it is the latter.

Build the failure rate into the numbers

More important still are failed picks. The gripper closes but the part is not there, the part is dropped in transit, the part is recognised but the required orientation is not achievable. The probability depends on the state of the bin, and it rises as the bin empties. Pieces left against the walls and in the corners face tighter interference and a narrower set of usable grip orientations.

Take assumed figures and work it through. One bin holds 200 pieces, and the loss per failed pick is 3.8 s (imaging 0.4 plus recognition 0.6 plus path 0.3 plus move and grip 2.5). If the first 180 pieces have a 3 percent failure rate and the last 20 pieces a 30 percent failure rate, the time for one bin comes out as follows.

• First 180 pieces — 180 pieces × (6.0 s + 0.03 × 3.8 s) = 180 × 6.114 = 1,100.5 s
• Last 20 pieces — 20 pieces × (6.0 s + 0.30 × 3.8 s) = 20 × 7.14 = 142.8 s
• Subtotal — 1,243.3 s (about 20.7 minutes)

Now add two pieces per bin that cannot be picked at all and require human intervention, at 3 minutes per intervention. That adds 6 minutes, giving about 26.7 minutes (about 1,603 s). Divided by 200 pieces, the effective cycle time is 8.0 s. Against a capacity quoted on the 6.0 s standard, effective capacity is 6.0 ÷ 8.0 = about 75 percent, in other words 25 percent below the quoted capacity.

That 25 percent gets built into the production plan without appearing anywhere in the capacity calculation. So design cycle time from the worst case rather than the average, and put the following three things into the specification.

• The retry limit when picks fail consecutively, and the behaviour on reaching that limit (shake the bin, move to the next bin, call a person)
• A call mechanism that assumes human intervention, and the restart procedure after intervention
• Buffer capacity (a work-in-progress staging area) so the downstream process does not wait, or parallel cells

Solve “the last few pieces” mechanically

The failure rate spikes at the bottom of the bin and against the walls. Trying to solve that with smarter recognition sends the cost through the roof. Mechanical and operational answers are overwhelmingly cheaper — add a mechanism to tilt the bin, change to a container with a drop-away bottom, or make it a rule that a person empties the bin once the remaining count falls below a threshold. Switch the thinking from reducing the number of unpickable pieces to not creating the unpickable situation in the first place.

Lighting is not an extra cost — it is the first specification

The most common cause of commissioning trouble in robot vision is neither the camera nor the recognition algorithm. It is lighting. And lighting is rolled into “one set” at quotation stage and reappears as an extra cost at commissioning. The only way to avoid that is to write lighting into the specification from the start.

Lighting reaches pose accuracy by several routes. Underexposure blurs edges and shifts the position. Specular reflection from shiny metal — a blown-out highlight, or blooming — either punches a hole in the point cloud or, conversely, produces a surface that does not exist. Hard shadows cause the shadow outline to be mistaken for the workpiece outline. If brightness changes through the day, recognition that worked in the morning degrades in the afternoon.

On top of that, lenses have distortion that increases toward the edges, so a workpiece detected at the edge of the field of view carries more coordinate error. Correcting it in calibration is the baseline, but designing so that the robot never has to pick at the edge of the field (narrowing the field, taking multiple images, positioning the supply toward the centre) is more dependable.

Lighting issues that get added in a Thai factory

Thai factories differ from Japanese factories on several counts.

• There is a lot of ambient light — buildings designed for daylight, areas worked with the shutters open, roof lights. The direction and quantity of incoming light change through the day. Leave shielding to be dealt with later and it will become a problem after commissioning without fail.
• Light levels differ between dry season and rainy season — the sky is different at the same hour. It is not unusual for equipment commissioned in the dry season to lose recognition rate once the rainy season arrives. Consider whether acceptance conditions can include a check that spans seasons.
• Differences between shifts — some factories light different parts of the building on day shift and night shift. Equipment that does not perform on night shift loses the unattended running benefit described below outright.
• Dust and oil mist — contamination of the camera window and the lights builds up gradually and is therefore hard to notice. Turn cleaning frequency and procedure into an operating rule, and decide who does it.

The most effective countermeasure is physical shielding. Enclose the cell, fit a hood, run blackout curtains. If you are erecting safety fencing, consider at the design stage whether that fencing can double as light shielding. Integrating shielding into the safety fencing compresses the cost of Block 1 and Block 4 at the same time.

Payback does not come from labour reduction — three pillars and a worked model

From here we pin the cost-benefit down in numbers. As background, Thailand’s minimum wage is 337 to 400 THB per day and has been held unchanged going into 2026. Divided across eight hours, that is an hourly equivalent in the range of 42 to 50 THB. On that basis, the conclusion we have stated consistently holds here as well. Labour reduction on its own does not pay back a robot vision investment.

Payback has to be built from these three pillars.

• Shorter changeover and variant switching time — replacing a fixture change with a model change shortens the changeover itself
• Preventing defect escapes and misassembly — less downstream rework, re-inspection and customer complaint
• More operating hours through unattended running at night and at weekends — more output from the same equipment

Assumptions for the model case (all assumed figures)

Picture a Japanese-owned factory in Thailand with one small-part assembly cell. Every figure below is an assumed figure and not a quotation for a real project. Use it as a skeleton for recalculating with your own numbers.

• 250 operating days per year, contribution margin for the cell of 1,200 THB per hour
• Variant changeovers 6 times per day, 1,500 times per year
• Current rework caused by misassembly at 400 cases per year, average cost of rework, re-inspection and transport of 550 THB per case

Two scenarios are compared. Scenario 1 is “solve it with 3D vision, parts left piled loose”. Scenario 2 is “change the supply to semi-ordered first, then solve it in 2D”.

Investment (initial, four blocks plus the robot itself)

Figures are in THB. Moving the supply to semi-ordered cuts “see” by 450,000 THB and “teach” by 650,000 THB, while “grip” rises by 400,000 THB. Net, that is 700,000 THB, taking the total from 4,000,000 THB down to 3,300,000 THB. The point of the table is that what moved the cost was the supply condition, not the camera model number.

Annual running cost (only what recurs every year, excluding initial investment)

Variant additions are set at 40,000 THB per variant in Scenario 1 (160,000 THB for 4 per year) and 15,000 THB per variant in Scenario 2 (60,000 THB for 4 per year). Creating a 3D recognition model and defining grip points is reliably heavier work than registering a 2D shape.

Safety fencing, risk assessment and commissioning attendance are already booked as initial investment under Block 4 and are not included in this annual running cost. That is stated explicitly to avoid double counting.

Annual benefit (the three pillars)

Pillar 1, shorter changeover and variant switching. A 12-minute fixture change becomes a 2-minute model change plus verification, saving 10 minutes per occurrence. At 1,500 occurrences per year that is 15,000 minutes, or 250 hours. Multiplied by the contribution margin of 1,200 THB per hour, that gives 300,000 THB per year. Treated as common to both scenarios. Note, however, that Scenario 2 has locating fixtures on the supply side, so in processes where trays or dividers have to be swapped per variant the changeover does not disappear entirely and this benefit is not truly common. Confirm before quotation whether the tray geometry can be shared across variants. Further, this benefit only becomes money if the cell is the capacity bottleneck and there is work to run in the time freed up. If all it produces is more idle waiting, the 300,000 THB is an illusion.

Pillar 2, preventing misassembly and defect escapes. Current rework cost is 400 cases × 550 THB = 220,000 THB per year. Assuming vision-based checking of variant, orientation and presence removes 80 percent of that, the benefit is 176,000 THB per year. Treated as common to both scenarios.

Pillar 3, more operating hours through unattended running at night. Take 2 hours per night × 5 days per week × 50 weeks per year = 500 hours as the unattended running window. Here the scenarios diverge. With the cycle time model of the previous chapter left as it is, bin picking requires two human interventions per bin, so unattended night running does not work at all. Assume the mechanical measures mentioned at the end of that chapter (a bin tilting mechanism, a drop-away container, moving to the next bin once the remaining count falls below a threshold) are added so that no one has to be called, and set the proportion of the window that can be run through unattended at 50 percent to reflect the failed picks that remain. That gives 250 hours, which at 1,200 THB is 300,000 THB per year. Semi-ordered supply has fewer failed picks, so at 80 percent it gives 400 hours and 480,000 THB per year. Both of these proportions are assumed figures, and as the sensitivity analysis below shows, they influence the conclusion more strongly than anything else.

Payback period and five-year ROI

Line the definitions up. Investment is the sum of initial expenditure across the four blocks plus the robot itself, and annual net gain is the annual benefit less the annual running cost. Block 4 is counted only on the investment side and is not in the annual running cost. On that definition, simple payback works out as follows.

• Scenario 1 — 4,000,000 ÷ 416,000 = about 9.6 years
• Scenario 2 — 3,300,000 ÷ 756,000 = about 4.4 years

Accumulated over five years, the picture looks like this.

Same process, same benefit framework, and yet changing the supply condition alone can flip the sign of the five-year ROI. That is the core claim of this article expressed in numbers. As the sensitivity analysis below shows, though, the reversal is not driven by the investment difference alone; it assumes that tidying up the supply also extends unattended running.

What happens if you count labour reduction alone

For comparison, calculate the case where the benefit is restricted to labour reduction. Assume one person is replaced, at 337 to 400 THB per day, 250 days per year, with statutory contributions and welfare adding 25 percent (an assumed figure). Annual labour cost then falls in the following range.

• At 337 THB per day — 337 × 250 × 1.25 = about 105,000 THB per year
• At 400 THB per day — 400 × 250 × 1.25 = 125,000 THB per year

Even calculated at the upper bound of 125,000 THB, that is below the 200,000 THB annual running cost of Scenario 2. In other words, if labour reduction is the only benefit counted, the annual net gain is minus 75,000 THB and no payback period can be calculated at all. Scenario 1 is minus 235,000 THB, an even larger annual outflow. Even ignoring running cost and simply dividing the investment, the result is 3,300,000 ÷ 125,000 = 26.4 years and 4,000,000 ÷ 125,000 = 32.0 years.

The moment labour reduction becomes the main justification in the capital request, the project stops paying back. This is the conclusion we have stated consistently, and robot vision does not change it.

Sensitivity analysis — which assumption overturns the conclusion

The assumption doing the most work in the model above is the contribution margin of 1,200 THB per hour. Pillars 1 and 3 together come to 780,000 THB in Scenario 2, roughly 80 percent of the 956,000 THB annual benefit. Halve the margin to 600 THB per hour and Pillar 1 becomes 150,000 THB, Pillar 3 becomes 240,000 THB, Pillar 2 stays at 176,000 THB, the annual benefit becomes 566,000 THB and the annual net gain 366,000 THB. Payback becomes 3,300,000 ÷ 366,000 = about 9.0 years and the five-year ROI -44.5 percent, so even Scenario 2 stops working.

The second sensitive assumption is the frequency of variant additions. If Scenario 1 goes from 4 to 12 variants per year, the variant addition cost becomes 12 × 40,000 = 480,000 THB, the annual running cost swells to 680,000 THB, and the annual net gain falls to 96,000 THB. Payback becomes 4,000,000 ÷ 96,000 = about 41.7 years. The frequency of variant additions kills payback far more effectively than camera performance ever could.

The third is the unattended running success rate assumed in Pillar 3. Break down the 2,400,000 THB difference in the five-year balance and the investment difference is only 700,000 THB; the rest is made by the Pillar 3 difference (900,000 THB) and the annual running cost difference (800,000 THB). Drop the semi-ordered success rate to the same 50 percent as bin picking and Scenario 2’s Pillar 3 becomes 300,000 THB, the annual benefit 776,000 THB and the annual net gain 576,000 THB, giving a five-year balance of 2,880,000 − 3,300,000 = minus 420,000 THB and a five-year ROI of -12.7 percent, so the sign does not flip. Working backwards, Scenario 2’s five-year balance turns positive once the success rate exceeds 64 percent. Tidying up the supply works because fewer failed picks mean more hours running unattended; if that extension cannot be captured, the investment difference alone will not overturn the payback.

Note that “variant additions kill payback” does not contradict the earlier point that “more variants favour vision”. What works in vision’s favour is the number of times an already registered variant is switched to; what kills payback is the number of newly registered variants. Vision is strong in a process with a fixed variant mix and frequent switching, and weak in a process where new variants accumulate every year.

When you recalculate with your own numbers, the first three things to confirm are these — the value of time (contribution margin), the frequency of variant additions, and the unattended running success rate.

Additional issues at a Thai or overseas plant

Installing robots is hardly unusual any more. According to the International Federation of Robotics and its World Robotics 2025 release (25 September 2025), new industrial robot installations worldwide reached 542,000 units in 2024, with an operational stock of 4,664,000 units, up 9 percent year on year. Asia accounted for 74 percent of new installations, Europe 16 percent and the Americas 9 percent, with China at 295,000 units, Japan at 44,500 units and South Korea at 30,600 units by country.

Turning to Thailand, a report from the Tokyo SME Support Center Thailand Office (18 July 2025, sourced from IFR and NECTEC) states that Thailand installed about 3,300 units in 2022, ranking 14th in the world and 2nd in ASEAN behind Singapore. The Thai figure is as of 2022, but even allowing for the difference in year, it is an order of magnitude away from Japan and China. That gap shows up directly as a thin local pool of engineers who can work on these systems.

Who handles variant additions

As the sensitivity analysis showed, the frequency of variant additions drives payback. And at an overseas site, what makes this a problem is not so much the cost as the lead time. You ask the system integrator to register a model for a new variant, arrange a visit, and have the work done. During all of that, the variant cannot run on the robot. What takes a few days in Japan takes several weeks overseas.

There is only one countermeasure. Enable local staff to handle model registration and re-teaching themselves. What that requires you to settle at contract signing is the number of people to be trained, the language of the training material, the language of the operator interface, whether procedures will be provided in Thai, and the support channel for questions after training. Leave this vague in the contract and the capability ends up resting on one individual, and the line stops the day that individual resigns.

The practicalities of language and display

The language of the operator interface is easy to overlook. Vision systems whose configuration screens are English-only, or whose error messages appear only in Japanese, genuinely exist. What the operator on the floor reads is not only the error message. Routine variant changeover means opening configuration screens too, and deciding what to do when something goes wrong also depends on the language on screen. If Thai display is not possible, extend the commissioning scope to include producing a posted chart listing error codes and their responses in Thai.

Drawing the boundaries of responsibility

Robot vision sits astride system boundaries. The camera manufacturer, the robot manufacturer, the gripper fabricator, the system integrator doing the integration, and your own production engineering team. When positioning fails, whose scope it falls under is not immediately obvious. The state that burns the most time is “the camera is delivering its specified accuracy”, “the robot meets its repeatability”, “and it still cannot pick the part”.

To avoid that, write the final acceptance criterion in terms of successful picks before signing. Not recognition rate, not measurement accuracy, but how many times out of a specified number of attempts the specified workpiece was picked from the specified supply condition and placed at the specified location. Whether that one sentence exists changes how commissioning stalls. We have set out how to divide responsibility in five boundaries to examine when selecting a robot system integrator. Vision straddles several of the boundaries listed there, so it needs an especially careful line.

Spare parts and recovery time

Cameras, lights, cables and the bespoke gripper. Recovery time when one of these fails is decided by whether the part is in the country. Decide whether to hold spares or accept the downtime, by comparing the money. If the loss from one day of stoppage exceeds the price of one spare unit, hold the spare. Judge this with your own numbers, not assumed ones.

Ten items to settle on paper before you order

Everything above condenses into a pre-order checklist. Put all of it in writing and agree it with the other party. Anything confirmed verbally will diverge at commissioning without exception.

Of these, the ones most likely to differ between quotations are 1, 4, 5 and 7. Take competing quotations on conditions where those four are aligned and the price difference becomes almost purely a difference in capability. Take them unaligned and the cheap quotation is, in most cases, simply the one that leaves things out.

Frequently asked questions

How much does robot vision cost

In the model case in this article (assumed figures), initial investment including the robot itself came to 4,000,000 THB for the bin picking plus 3D scenario and 3,300,000 THB for the semi-ordered plus 2D scenario. This varies enormously with the process, the workpiece and the supply condition. What decides the money is not the camera model number but the supply condition and the volume of modification on the gripper and supply side. When taking quotations, require a breakdown across the four blocks of see, teach, grip and connect, and compare the “teach” and “grip” lines in particular.

Is 3D bin picking genuinely production-ready

As a technology, yes. Acquiring depth information to determine the pose of stacked workpieces and extract them without collisions is productised by several vendors. The question is not feasibility but cost-benefit. In the worked model here, the five-year ROI of solving it with the parts left piled loose was -48.0 percent, against +14.5 percent for the scenario that moved the supply to semi-ordered. If you are going to choose bin picking, first confirm whether the reason the supply cannot be changed is genuinely unchangeable.

Can a piece picking robot handle a high variant mix

It can, but the cost of variant additions decides whether it pays back. In the sensitivity analysis here, Scenario 1 had a payback period of about 9.6 years assuming 4 variant additions per year; at 12 per year, the annual running cost rose to 680,000 THB and payback stretched to about 41.7 years (both being calculations on assumed figures). To use it across many variants, the prerequisites are lowering the registration cost per variant and building the local capability to do the registration.

Can gripper selection wait until after the camera is chosen

The opposite. Think about the gripper first. Only once you know how much positional error the gripper can absorb can you determine the accuracy demanded of the camera. Chamfers, a floating mechanism, or a configuration that picks roughly and re-places the part on a regrip station will all drop the camera requirement by a level, along with the cost and the commissioning hours. Designing give into the gripper is cheaper than upgrading the camera — that is the practical principle.

Will we be able to do robot teaching ourselves

Where vision is involved, three tasks are added on top of conventional teaching — registering recognition models, defining grip points, and re-acquiring the calibration. Of these, model registration and grip point definition are usually within reach of local staff after training. Calibration is the same, provided the procedure has been standardised. Specify at contract signing the number of people to be trained, the language of the training material, the language of the procedures, and the support channel for questions after training.

Should we choose screw fastening automation with ordered feeding, or robot vision

First, processes where the workpiece can be located by a fixture, such as screw fastening and press fitting, belong with fully ordered feeding as a matter of course. Vision in that case shrinks to confirming “is the part present” and “is it the right way round”, and robot vision in the coordinate-handing sense is unnecessary. Coordinate output is only needed where the position of the workpiece changes every time.

Beyond that, whether to lean toward ordered feeding equipment or toward vision is decided by variant count and changeover frequency. With few variants and rare changeovers, part feeders and dedicated fixtures are cheaper, faster and more robust. As variants multiply and changeovers become frequent, the number of fixtures, the storage space and the changeover time all grow, and at some point switching a vision model becomes cheaper. To decide, line up the fabrication cost and storage cost of one fixture set, the time per changeover, and the annual number of changeovers, and calculate annual benefit against investment in the same format used in this article.

Where is the value in adding robot vision to assembly automation

The greatest value comes from shorter changeovers and the prevention of misassembly. Replacing a per-variant fixture change with a model change shortens changeover time and reduces the physical volume of fixtures. Using vision to confirm variant, orientation and presence stops misassembly from reaching the downstream process. Conversely, if the only benefit counted is removing one person, it will not pay back at Thai wage levels (337 to 400 THB per day).

Summary

What decides whether robot vision succeeds is neither camera resolution nor an accuracy specification. It is these three — how the workpiece is supplied, whether the lighting is stable, and how much error the robot and the gripper can tolerate between them — and none of them appears on the camera quotation.

Positioning error is the sum of three layers — camera measurement accuracy, robot absolute accuracy, and the mechanical tolerance between gripper and workpiece. What appears in a catalogue is Layer 1 and, from Layer 2, only repeatability. In an application where vision hands over a different coordinate every cycle, absolute accuracy is what matters, and that figure is usually not published. The design answer, therefore, is not to buy accuracy but to build give into the gripper side.

Read the cost in four blocks — see, teach, grip and connect. What gets compared in a quotation is only “see”, while what moves the total is “teach” and “grip”. And the largest single influence is the supply condition. In the worked model on assumed figures in this article, solving it in 3D with the parts left piled loose gave a five-year ROI of -48.0 percent, while moving the supply to semi-ordered and solving it in 2D gave +14.5 percent — a reversal of sign. Design cycle time from the worst case rather than the average, and build failed picks and human intervention into the numbers. Write lighting as a first specification, not an extra cost. And payback does not come from labour reduction; build it from the three pillars of shorter changeovers, prevented defect escapes, and more operating hours from unattended running.

TOMAS TECH is based in Bangkok and provides production management, IoT and automation to Japanese manufacturers across Thailand and ASEAN. On robot vision projects, working through “can the supply be changed” and “can we design give into the gripper” before choosing a camera almost always results in a faster and cheaper start-up. It is fine if the process is not yet fixed, or if you are simply stuck comparing quotations whose breakdowns do not line up. Bring us your current supply condition and photographs of the workpiece, and we can help from the initial assessment of whether 2D will do or 3D is required. Get in touch through our contact page to start the conversation.

References

• International Federation of Robotics, “World Robotics 2025” press release (25 September 2025) — 542,000 new industrial robot installations in 2024, operational stock of 4,664,000 units, breakdown by region and country
• Tokyo SME Support Center Thailand Office, “Market trends for industrial and service robots in Thailand” (18 July 2025) — Thailand installed about 3,300 units in 2022, ranking 14th in the world and 2nd in ASEAN
• RoboDK, “Robot Validation ISO 9283” — definitions of pose accuracy, repeatability and path accuracy under ISO 9283, and the test cube conditions
• Mech-Mind, “Principles of hand-eye calibration” — the eye-in-hand and eye-to-hand configurations, the transform to be solved, and the TCP touch method
• Keyence, “Positioning capability and calibration are the keys to industrial robot productivity” — coordinate drift over time, unstable lighting, lens distortion, mounting shift
• Keyence, “3D robot vision for bin picking” — the definition of bin picking and why depth information makes 3D necessary
• Orbbec, “Automate 2026” news (22 June 2026) — roughly 0.07 mm VDI/VDE accuracy and 0.05 mm Z repeatability at 0.5 m for the DLP structured light camera
• Photoneo, “Automate 2026” — AI-driven 3D vision for delicate part picking, unordered bin picking and pallet dimensioning