When selecting a robot force/torque sensor, simply comparing catalog specs such as “6-axis” or “high precision” does not guarantee success in assembly or surface finishing processes. It is essential first to define what needs to be detected upon contact, which axis forces to control at what cycle, and how to stop equipment in case of abnormalities. This article organizes the approach for both new installations and retrofits on existing robots in Thai factories, covering candidate methods, RFPs, FAT/SAT, and 90-day proof-of-concept (PoC) procedures.
A force/torque sensor is a device that measures the forces and torques applied near the robot’s wrist or tool. Typical 6-axis F/T sensors measure force in Fx, Fy, Fz directions and moment in Mx, My, Mz directions. However, the “6-axis” label does not guarantee simultaneous control at any desired precision in all six directions. The overall process outcome depends on the sensor’s measurement range, noise, temperature drift, tool mass, robot controller response, and the rigidity of the contact object. Our previous assembly automation robot article explained part tolerances, and the deburring automation article discussed machining methods. Here, we focus on how to select measurement devices common to both processes.
First, Define Process Failures
Before purchasing a force/torque sensor, categorize current defects into at least three types. The first is misalignment—when the center of a hole and pin do not match, causing lateral force to rise at the start of insertion. The second is contact variation—where pressing force is inconsistent due to workpiece dimensions, jig wear, or temperature changes. The third is delayed abnormality detection—where jamming or tool breakage is only discovered after the robot stops or during inspection. This classification changes the required measurement axes and control actions.
For example, in connector insertion, not only axial pressing force but also the rise of lateral force or bending moment can be used to detect misalignment. In polishing, the design aims to stabilize the pressing force in the surface normal direction while following the path. If tool wear or workpiece shape changes are significant, combine force feedback with vision, compliance mechanisms, or jig improvements. The force/torque sensor is not a universal component that automatically corrects defects; it is one element for measurement and decision-making.
In the requirements document, avoid only abstract goals like “reduce defect rate.” Instead, specify representative workpieces, allowable scratches, permissible insertion force, success criteria, processing time, retry count, and recovery after abnormal stops. Determine tolerances from product drawings and actual measurements, not by adopting figures from other companies’ cases. If only average pressing force is used as a pass criterion, you may overlook cracks from momentary overloads or re-acceleration after catching. Also evaluate whether waveforms, peaks, duration, and inter-axis relationships can be recorded.
Evaluate Built-in Force Detection and External 6-Axis Sensors Separately
Even if an existing robot has force control functions, additional sensors may still be necessary. If joint torque estimation or built-in wrist sensors on the robot can achieve the required process accuracy and contact response, there is little reason to add external devices, wiring, or maintenance points. On the other hand, if you need to detect small contact changes at the tool side, acquire waveforms with the existing controller, or require independent measurement logs, consider an external 6-axis sensor. Do not decide between the two from the start—compare them using the same workpiece and evaluation criteria.
Universal Robots’ Force Mode documentation shows that axes maintaining strict position and axes moving according to specified force are separated, with adjustable force/velocity limits, gain, damping, and zero point. This is a concrete example where the term “force control” alone does not define behavior. Available functions differ by robot model and software version, so always check the actual version and license at the quotation stage. When using external sensors, confirm at what cycle the controller reads measurements, how coordinates are transformed, and what the data can be used for.
FANUC describes its own force/torque sensor as detecting six degrees of freedom, listing applications such as assembly, fitting, deburring, and polishing. However, published application examples and load values by model do not guarantee the same yield rate for your workpieces. Robotiq’s FT 300-S lists 300 N, ±30 Nm, 100 Hz, IP65, and RS-485 Modbus RTU on its public page. These are published specs for that model and should not be generalized to other models or actual closed-loop responses. In comparison tables, always specify the product name, publication date, and model.
Do Not Select Measurement Range Based Only on Normal Loads
When choosing sensor ratings, it is insufficient to simply add a margin to the steady pressing force. Separately identify the weight of the tool and adapter, load when lifting the workpiece, inertial force from robot acceleration/deceleration, moment generated by the distance from the tool tip to the sensor, and impact from contact failures. Especially with long tools, even small forces can generate large moments. Since a 6-axis sensor must satisfy the range for each axis simultaneously, you cannot compare by a single maximum force value.
ATI’s Axia80 model page shows measurement range and overload for each axis, mass, and dimensions, with a note stating “applied loads must be within the range for each of the six axes.” Apply this concept to your RFP, requesting load envelopes for normal operation, startup/shutdown, posture changes, tool changes, and workpiece jams. Overload resistance does not mean correct values are obtained beyond the normal measurement range. Always confirm recalibration after collision, mechanical stoppers, and retreat conditions individually.
Resolution and noise should be evaluated against the minimum contact difference required. For example, if you want to detect small pressing differences and zero drift after temperature rise approaches the detection threshold, catalog values at room temperature are insufficient. A high sampling rate alone does not have standalone value. Evaluate together with communication cycle, filter delay, PLC or robot controller program cycle, stop command, and the time until the machine actually stops. It is important to measure the time from contact occurrence to actual action.
For Retrofits, the Product Includes Everything from Flange to Cable
When installing an external sensor between the wrist flange and the tool, you cannot judge mountability by diameter on the drawing alone. Check the robot-side mounting pattern, tool-side plate, centering, tightening torque, redefinition of tool center point (TCP), payload and center of gravity, interference by posture, and cable bending/tension. If the cable wraps around the flange, it can cause not only communication errors but also hinder tool operation itself. Move the robot’s entire range at low speed to check, and leave access space for replacement work.
Robotiq’s installation documentation for the older FT 300 covers not only the sensor body but also couplings, adapters, cables, and power supply. It does not define the current FT 300-S kit. When procuring an FT 300-S, check the latest model-specific documentation for included and separately sourced parts, and list them in the quotation. In Thai factories, waiting for imported replacements directly affects downtime, so clarify supply responsibility, stock location, and zero-point adjustment procedures for sensors, cables, connectors, plates, and spare tools. Comparing only “sensor price” underestimates total cost.
Judge IP rating and temperature range by installation location. If there is polishing dust, coolant, cleaning water, or oil mist, check not only the model’s protection rating but also whether connectors, cables, and junction boxes can withstand the same environment. Do not interpret published protection ratings as guarantees against all chemicals, spray angles, or long-term immersion on site. Route wiring away from operator hands and tool change paths, and ensure inspection and cleaning are possible.
Make Coordinate Systems, Zero Point, and Tool Mass Acceptance Criteria
Even with the same 6-axis data, mixing sensor, tool, and workpiece coordinate systems can cause control to act in the wrong direction. If you do not update mass and center of gravity when changing tools, you may misread gravity-compensated values as contact force. Decide procedures for checking the zero point at startup, after tool change, after temperature changes, and after collisions. Zeroing should be done in a posture and condition where the tool is not touching the workpiece, recording time, operator, tool ID, and result.
In FAT, check the direction of coordinate transformation with a simple known load. When applying a unidirectional force to the tool, verify that the sign and magnitude of the specified axis are as expected. Next, change the robot posture and confirm gravity compensation is working. Having a calibration certificate and correct system measurement after cell integration are separate matters. After transport to the Thai factory, tool mounting, and cable replacement, repeat the same checks and align FAT and SAT measurement conditions.
Abnormal waveforms should be saved with the same importance as good product waveforms. Use test cases such as good product, misalignment, missing workpiece, double feed, tool wear, sensor disconnection, communication delay, and emergency stop, and cross-check machine behavior and logs for each case. Logs should include not only 6-axis force/torque but also robot posture, tool ID, workpiece ID, judgment thresholds, and time synchronization status. However, limit the scope of collection to what is used in the process, and do not estimate based on indefinitely storing massive time series data.
Do Not Confuse Force Control with Safety Functions
Force control for quality purposes does not automatically substitute for safety stops. Even if you create logic to stop the robot program upon detecting abnormal contact force, unless that system is designed and verified as a safety function to protect people, responsibilities for safety light curtains, guards, emergency stops, and speed limits remain. ISO 10218-2:2025 covers requirements for industrial robot applications, cell integration, commissioning, and operation, while ISO 13849-1:2023 covers design methods for safety-related control parts. Apply these to your own cell based on risk assessment.
In the RFP, request diagrams that separate sensor signals for quality and safety. Clearly specify which detections are for good product judgment, which require protective stops, and what happens in case of power failure, disconnection, or communication loss. Even with collaborative robots, neither the sensor model nor the robot’s nominal safety functions alone guarantee cell-wide safety. Verification must include sharp workpiece edges, tools, jigs, operator access points, stop distances, and recovery actions.
12 Items to Include in the RFP
A quotation request for “one set of force/torque sensors” is insufficient. If you have each supplier answer the following twelve items in the same table, you can compare not only initial price but also the scope of implementation.
- Material, dimensions, tolerances, weight, surface condition, and representative samples of the target workpiece.
- Good product criteria, defect modes, allowable scratches, target force/torque values, and verification methods.
- 6-axis loads during normal, abnormal, and collision states; required measurement range; and overload handling.
- Robot/controller model, software version, and available force control functions.
- Sensor model, axis-specific specs, temperature/IP/communication/response conditions, and calibration records.
- Flange, adapter, tool plate, TCP, payload, and cable routing.
- Coordinate transformation, gravity compensation, zero point, and reset procedures after tool change.
- Waveforms, alarms, workpiece ID, time synchronization, retention period, and data output format.
- Separation of safety circuits and quality control, and behavior during disconnection, communication loss, or power failure.
- FAT/SAT test workpieces, defect samples, tolerances, and retest conditions.
- Local installation, training, warranty, spare parts, recovery targets, and language support in Thailand.
- Cost breakdown including licenses, control modifications, jigs, calibration, and maintenance beyond the main unit.
For items with unspecified requirements, do not leave blanks—state “to be decided in PoC,” and set a decision deadline and responsible person. In quotation comparisons, do not just ask “which is cheaper for the same sensor,” but compare assembly success rate, tool life, stop frequency, waveform traceability, and setup man-days under the same conditions. If you only adopt samples that succeeded in the vendor’s demo, you cannot reflect your own tolerances and wear.
The 90-Day PoC Should Be Judged in Three Stages
In the first 30 days, fix the process and workpiece and measure the baseline. Record the current success rate, cycle time, scratches/damage, rework, stops, and operator intervention time for manual or position-controlled processes. Prepare not only good products but also dimensionally marginal and intentionally misaligned samples. At this stage, have designers and quality departments approve the safe upper limit of allowable force and begin actual measurements.
In the next 30 days, install the candidate built-in function or external sensor. Start with low-force, low-speed contact tests, then check axis signs, zero point, filters, speed, and posture before expanding the range. Record each threshold adjustment version and compare with previous waveforms. If only the person who set it can recover, it cannot be accepted for mass production. Also test whether local maintenance staff can reproduce tool changes and zero point adjustments according to the manual.
In the final 30 days, verify continuous operation across shifts and abnormal recovery. Measure not only average cycle time but also worst-case cycle, first-pass yield, retry ratio, unintended stops, rework, replacement work time, and record loss. Since PoC results do not become guaranteed values for mass production, separate thresholds for FAT/SAT and those to be reviewed after mass production. Before expanding to multiple workpieces, solidify maintenance and safety procedures for representative workpieces.
What to Confirm in FAT and SAT
FAT is conducted before shipment with the planned model, firmware, wiring, and tool configuration. Confirm waveforms for good and defective samples, continuous operation, tool change, zero point reset, power cycling, communication disconnection, sensor saturation, robot stop, and manual recovery. Record results linked to process conditions, including marginal workpiece tolerances. “It worked in a demo” videos are not acceptance records. Record test date, equipment number, version, operator, judgment, deviations, and retest after corrections.
In SAT, reconfirm on actual equipment at the Thai factory, including vibration, temperature, dust, cable routing, actual air/power supply, operator flow, and shift changes. If jig or tool positions change during relocation, reconfirm TCP, gravity compensation, and zero point. Also verify whether local parts can be used for replacement, spare part numbers match, and maintenance staff can reproduce and recover from abnormalities. After equipment handover, define authority for threshold changes, log retention, backup, and manufacturer support contacts to complete the process.
Calculate Cost Effectiveness Separately for Defects and Downtime
Include not only the sensor body but also tool adapters, robot options, wiring, control modifications, jig adjustments, calibration, training, spare parts, and startup downtime in total investment. Calculate monthly net savings from separately measured benefits such as fewer defects, less rework, shorter cycle time, longer tool life, and less downtime, then subtract added operating costs. Estimate simple payback as total investment divided by monthly net savings; if net savings are zero or negative, there is no payback. Do not double-count savings for a single defect as both “defect reduction” and “rework reduction.” If production volume changes, normalize to the same workpiece configuration before and after improvement.
Decisions on implementation should compare not only annual savings but also alternatives without additional sensors. Consider whether jig alignment improvements are sufficient, passive compliance in the tool is cheaper, or built-in force functions alone are adequate. If the reason for adopting a force/torque sensor is “to measure process variation, visualize it, and correct it reproducibly,” make that value visible in acceptance testing. If you look for applications after purchase, specs tend to be excessive.
Evaluation Items to Include in Comparison Tables
When comparing candidate products side by side, fix the comparison conditions as well as product names. First, provide all companies with representative workpieces and defect samples from the same process, and compare waveforms obtained with the same robot speed, tool, and jig. If Product A’s demo uses light plastic parts and Product B’s uses heavy metal parts, you cannot compare catalog ranges or setup man-hours. If quotation conditions differ, clearly indicate the differences in separate rows before evaluating the total.
Columns in the evaluation table should include measurement range, noise or resolution, overload, temperature range, protection rating, mass and dimensions, compatible robots, communication method, sampling and control reflection cycle, and calibration method. Next, list adapters, cables, control options, software licenses, local installation, training, spare parts, and warranty coverage. At the end, include “evidence of yield rate,” “stop and recovery in case of abnormality,” and “maintenance staff reproducibility” to move from spec sheet comparison to process adoption decisions.
Note the term “accuracy.” Depending on the manufacturer, it may refer to linearity, repeatability, resolution, noise, or calibration accuracy, and depends on measurement direction, temperature, and filter conditions. Do not put numbers with different definitions in the same column—retain the spec name and source. Decide the required discrimination width first, and verify in FAT/SAT whether actual misjudgment occurs under site temperature and posture conditions.
Decide Thai Factory Procurement and Maintenance Structure in Advance
Even if Japanese head office engineers can handle equipment at installation, local teams will be responsible for recovery during night shift abnormal stops in Thai factories. Design alarm messages, zero point procedures, post-tool-change checks, and waveform storage locations according to actual language and job roles. Do not just hand over Japanese manuals—verify on site whether maintenance staff can distinguish sensor abnormalities from workpiece abnormalities. Include conditions for contacting the manufacturer or SIer, contact details, and response times in the equipment handover.
Lead time for replacement parts may depend not only on the sensor body but also on dedicated cables, converters, connectors, and tool plates. Even if the body is in stock, you cannot recover without connection parts. In the RFP, confirm recommended spare parts and models, compatibility, local inventory, and whether recalibration is required for substitutions. For long-term use, check whether communication is maintained with software and controller updates, and store configuration files and test workpieces.
In contracts, do not equate “FAT passed = ready for mass production on site”—set separate SAT completion criteria. Separate responsibilities for air/power/workpieces supplied by the factory, jigs and software brought by the SIer, robot manufacturer’s warranty scope, and quality department’s approval criteria. Decide in advance on costs and schedules for spec changes, preparation of new test workpieces, and retest responsibilities to avoid blame-shifting after implementation.
The Decision Not to Implement Is Also an Acceptance Result
The purpose of validation is not to justify sensor purchase, but to select the best process improvement method. If jig rigidity and tolerances are the main issue, jig modification may be the priority. If workpiece supply posture variation is the main issue, improvements to feeders or vision systems may be more effective. If contact points vary but load waveforms are stable and distinguishable, the effect of a force/torque sensor is visible. Compare all three alternatives for yield rate, cycle time, downtime, and maintenance burden under the same conditions.
If the PoC result is “not implementing this time,” record the measured waveforms and reasons. For example, if good and defective waveforms overlap and cannot be distinguished, required response is faster than the existing controller cycle, or tool shape changes eliminated the problem, these findings can be used for future equipment procurement. Conversely, if adopting, retain evidence that thresholds worked not only for successful samples but also for failures, that night-shift maintenance staff could recover, and that safety circuits were separately verified.
Designing Change Management After Acceptance
Sensor settings are not always fixed once determined. Changes such as variations in workpiece material or surface treatment, wear of supply jigs, replacement of tool tips, or updates to robot software can all affect signal waveforms. If threshold adjustments are to be made on-site, it is necessary to define who can make changes, with which test workpieces, and within what range. To prevent unapproved settings from being applied to the mass production line, maintain version numbers and change histories, and prepare a method to revert to previous versions.
In monthly reviews, do not only check the “number of anomalies,” but also verify whether the sensor is missing the same defect modes or, conversely, excessively stopping good products. Since the number of anomalies increases with production volume, compare results per number of workpieces. Also record the time required for zero-point checks during maintenance, the number of recalibrations, and occurrences of communication interruptions. These are indicators not only of the sensor’s performance but also of the overall health of the process, including tools, control, and maintenance.
When deploying to another line, do not simply copy the same thresholds if the robot model, tool length, workpiece tolerances, or ambient temperature differ. First, recalculate the coordinate system and load envelope, and repeat FAT-equivalent tests with the smallest representative workpiece and an abnormal workpiece. Rather than copying successful configuration files as-is, documenting what needs to be re-verified as standard procedures will improve reproducibility. Each time a new workpiece is added, separately review the impact on quality judgment and safety.
Three Common Design Mistakes on the Shop Floor
The first mistake is failing to save the reference value before contact, even though waveform data is being acquired. If you judge that “press-in force has increased” from a waveform affected by tool weight and temperature, you cannot tell whether the process has changed or the zero point has shifted. Always save the reference in the non-contact state, the time of initial contact, and the posture. If you also record the date when the inspected sample changes, you can later distinguish between process and measurement differences.
The second mistake is believing that simply lowering the threshold will strengthen defect detection. Overly sensitive settings tend to increase false stops of good products and are often disabled on-site. When the waveforms of good and defective products overlap, evaluate not just the force at a single point, but also the elapsed time from contact, combinations of multiple axes, and correspondence with position. If differentiation is still not possible, review the jig, workpiece supply, camera, or tool shape. Having measurement data does not necessarily mean you can make accurate distinctions.
The third mistake is determining mass production speed based solely on low-speed demonstrations in the test lab. As speed increases, transient loads during collisions, control delays, and tool deflection all change. Even if sampling is fast, if the decision to stop and the actual stop are delayed, the workpiece cannot be protected. Gradually increase speed, repeat tests of good and defective products at each stage, and record stop positions, scratches, and recovery times. It is important not to adopt speeds that do not meet acceptance criteria, even if they are within the catalog’s maximum specifications.
Frequently Asked Questions
Can robot force/torque sensors be retrofitted?
There are possible configurations, but check compatibility with flange, tool, payload, cable, communication, controller software, and existing safety design. Mechanical mountability and availability of required force control are separate issues. For existing cells, request quotations based on on-site surveys and test workpieces.
Can a 6-axis sensor be used for any task?
No. “6-axis” describes measurement directions. Check if required load range, axis-specific noise, dynamic response, temperature, environment, tool, and workpiece conditions are suitable. The appropriate model differs between processes requiring detection of small contact differences and those with large polishing reaction forces.
Can force control eliminate safety fences?
Quality signal from force control alone does not justify elimination. Assess cell risk including hazards to people, tools/workpieces, access, and stop performance, and design/verify safety-related controls. Distinguish between sensor specs and the scope of safety function certification.
How much does a force/torque sensor cost?
It varies by model and implementation scope. Even if a unit price is published, total cost changes when adding adapters, robot options, control modifications, calibration, FAT/SAT, and local maintenance in Thailand. Show the same workpiece and acceptance conditions, and compare multiple quotations with breakdowns for practical evaluation.
Conclusion
Selecting a robot force/torque sensor is not just about comparing 6-axis numbers, but about designing how to detect contact failures in waveforms and how to recover. First, define process defect modes, and compare built-in functions and external sensors under the same conditions. Reflect axis-specific loads, tool moments, installation environment, coordinate systems, zero point, and safety separation in the RFP, and retain evidence reproducible in FAT and local SAT in Thailand.
At ordering, bundling marginal and defective workpieces, current tools, robot software version, measured work time, and stop history will facilitate discussions. From there, clarify “which axis detects what,” “what action follows detection,” and “who resets the zero point.” Even if candidates use different communication methods, if these three points are validated under the same conditions, you can decide on adoption. After acceptance, use the first month’s waveforms and defects as a baseline, and regularly review threshold shifts due to tool wear or workpiece changes to ensure benefits do not end at implementation.
For internal decision-making materials, include both with- and without-sensor test conditions, mass production constraints, and additional costs on the same page. If you also document re-evaluation procedures for unexpected workpieces, you can avoid copying settings to new equipment or lines without review. For every process change, decide which department—quality, maintenance, safety, procurement—must re-approve, ensuring sustained value after implementation.
If you are considering retrofitting assembly, polishing, or insertion processes in a Thai factory, please contact us once you have workpiece drawings and current defect/stop records. We can help organize measurement conditions and PoC acceptance criteria for your target process.
References
- FANUC America: Force Sensor — Public information on 6-DOF detection, applications, and model-specific data.
- Robotiq: FT 300-S Force Torque Sensor — Published specs for measurement range, output cycle, communication, and protection rating.
- Universal Robots: Force Mode — Product documentation on force control axes, speed, gain, and zero point.
- ATI Industrial Automation: Axia80-M20 — Model information on axis-specific ratings, overload, mass, and calibration.
- Robotiq: older FT 300 installation manual — Examples of parts, power, and wiring; separate from the current FT 300-S configuration.
- ISO 10218-2:2025 — Safety requirements for industrial robot applications and cells.
- ISO 13849-1:2023 — Design principles for safety-related control parts.
