6-axis force/torque sensor module in a robot system

In modern industrial and collaborative robots, a 6-axis force/torque (F/T) sensor module acts as the “tactile layer” at the robot wrist. It measures forces in three linear axes (Fx, Fy, Fz) and torques in three rotational axes (Tx, Ty, Tz), so the controller can keep contact force within a narrow window, detect abnormal collisions and monitor process quality instead of relying only on positional feedback.

On collaborative robots, the F/T module is typically mounted between the last joint and the end-effector. Force-controlled polishing, deburring or insertion can run with smoother motion and more consistent surface finish. The same module also supports safety features: when measured forces exceed configured limits, the robot controller can rapidly slow down or stop motion according to project-specific safety rules and standards.

On traditional six-axis industrial robots, the same type of F/T module helps keep contact force constant during grinding, dispensing, screwdriving or press-fit operations. Instead of tuning every step based only on position, the process engineer can use torque and force signatures to detect misalignment, missing parts, abnormal friction or tool wear, which improves throughput and reduces scrap without redesigning the whole mechanical line.

Compared with single-axis load cells or torque sensors, a 6-axis F/T module provides full 3D force and torque information and can distinguish between normal axial loading and harmful side loads or bending. The trade-off is higher cost, more complex calibration and tighter requirements on bandwidth and latency. The module therefore makes the most sense where precise contact control, flexible multi-tool usage or human–robot interaction are important project drivers.

From a system architecture point of view, the F/T module sits between the robot arm and the motion controller. Inside the module, multiple strain-gauge bridges feed low-noise instrumentation amplifiers and high-resolution ADCs. A small MCU or FPGA performs temperature compensation and matrix-based cross-talk correction and then exposes the six calibrated components over SPI or an industrial fieldbus. The controller side may connect to this module through SPI, RS-422/RS-485 or fieldbuses such as EtherCAT or TSN-enabled Ethernet, but the detailed network implementation is handled by higher-level communication or controller pages in the overall system.

From strain-gauge bridges to six-axis forces and torques

Inside a 6-axis F/T sensor, a machined structural body carries multiple strain-gauge bridges. Each bridge measures a particular combination of normal and shear strains at a specific location. Under applied forces and torques, the mechanical structure distributes loads across these locations, and the corresponding bridge outputs change by a few millivolts or less relative to the bridge excitation voltage.

A single bridge channel therefore does not map to only one axis. Instead, every bridge has a sensitivity to all six components (Fx, Fy, Fz, Tx, Ty, Tz) with different weights. During factory calibration, controlled loads are applied along and around each axis and the responses of all bridges are recorded. This data is used to build a sensitivity matrix that links the set of bridge outputs to the six physical components. In operation, the digital part of the module uses the inverted matrix to reconstruct the six-axis force/torque vector from the measured bridge voltages.

Because the bridge signals are small, each bridge or bridge group feeds a precision instrumentation amplifier (INA). The INA provides high gain, low input noise and high CMRR so the full-scale output of the bridge is expanded into a voltage span that matches the input range of the ADC. Depending on the architecture, several bridges may share a multichannel INA, or each bridge can have a dedicated amplifier channel where matching and layout simplicity are more important than saving components.

The amplified signals are then sampled by a high-resolution ADC. For many F/T modules, a sigma-delta ADC with 20- to 24-bit resolution and simultaneous sampling per channel is attractive, because it provides good noise performance at modest bandwidths and keeps all axes time-aligned. Other designs use fast SAR ADCs with carefully managed sampling sequences. Anti-aliasing filters in front of the ADC set the analog bandwidth and protect against high-frequency noise from drives, switching power supplies and communication interfaces in the robot cabinet.

Once digitized, the samples enter a MCU or FPGA that implements offset and gain trimming, temperature compensation, cross-talk correction and basic diagnostics. The sensitivity matrix from factory calibration is applied to convert the set of bridge-based readings into a 6-element vector expressed in physical units. Digital filtering reduces noise, but also adds group delay, so bandwidth and latency targets are typically agreed early in the project to ensure that force-control loops and safety functions still meet their timing budgets.

Top-level performance metrics for the complete chain include full-scale ranges on each axis, sensitivity and minimum resolvable force/torque, noise floor, usable bandwidth and end-to-end latency from a mechanical disturbance at the flange to a stable digital 6-axis reading at the module interface. The following sections drill down into each block of this signal chain, covering practical selection and layout rules for bridge INAs, ADCs, compensation strategies, isolation and EMI-aware PCB design in a real robot environment.

Bridge INAs: lifting microvolt signals into a usable range

At the heart of a 6-axis force/torque sensor module, each strain-gauge bridge only produces a few millivolts of differential output at full-scale load, often specified as 1–3 mV/V of excitation. For a 5 V bridge supply, this corresponds to roughly 5–15 mV of raw signal. Small changes in force or torque become microvolt-level shifts inside this span, so the instrumentation amplifier (INA) must expand this tiny signal into the input range of the ADC without adding excessive noise or drift.

The required gain can be estimated from the bridge full-scale output and the ADC input range. A bridge that produces 10 mV at maximum load and an ADC with a ±2.5 V input span suggest a target gain in the 200× region, leaving some headroom for over-range loads and mechanical tolerances. At the same time, the INA must provide enough bandwidth to pass the desired force/torque bandwidth—often a few hundred hertz—while keeping noise low enough that the smallest meaningful force increment is not buried under amplifier noise.

When reading the datasheet, the noise section deserves special attention. The input noise density in nV/√Hz indicates how much broadband noise is injected per root hertz of bandwidth, and the 0.1–10 Hz noise specification shows how much low-frequency flicker noise will appear as slow drift and offset wander. Force-control applications operate mainly in the low-hertz to a few hundred hertz region, so both noise density and 0.1–10 Hz noise directly shape how stable the zero-force reading will look on the controller side.

Input offset voltage and offset drift translate directly into apparent force error. Offset multiplied by gain becomes a fixed output shift that shows up as a non-zero reading at zero load. Offset drift, expressed in µV/°C, becomes a slow change in that reading as the module warms up or ambient temperature moves. When this error is mapped back through the mechanical sensitivity into Newtons or Newton-meters, it can easily reach the same magnitude as the smallest force or torque that the application is supposed to detect, so offset specifications are just as critical as noise numbers.

The common-mode range of the INA must be compatible with the bridge excitation and wiring scheme. Some bridges use 3-wire connections, others 4-wire Kelvin sensing, and not all amplifier architectures accept inputs that ride close to the supply rails. The datasheet input common-mode versus supply plots should be checked against the actual bridge bias point. In many designs, the ADC reference or the bridge excitation voltage is shared with the INA to form a ratiometric scheme, so that supply fluctuations are removed together with the bridge excitation variation.

In a real robot cabinet, the bridge and its wiring sit in an environment full of fast dv/dt from drives, motor cables and switching power stages. High common-mode rejection ratio (CMRR) at higher frequencies is therefore important, not just the 50/60 Hz CMRR headline. Insufficient high-frequency CMRR allows common-mode disturbances to convert into differential error, which shows up as jitter and apparent force spikes on otherwise constant loads. Designs that share grounds with noisy power electronics need especially strong high-frequency CMRR and careful attention to routing and shielding.

Multi-channel INAs integrate several matched channels in one package, offering good channel-to-channel tracking and a compact layout for multiple bridges. They are attractive for generic 6-axis modules where consistent behavior across axes is important. Single-channel INAs, in contrast, provide maximum freedom to place each amplifier close to its bridge, tune gain per axis and even mix different amplifiers for different ranges, at the expense of PCB area and BOM complexity. A practical selection checklist therefore starts with bridge full-scale and excitation, then walks through required gain and bandwidth, input noise, offset and drift, common-mode range, high-frequency CMRR and finally the choice between multi-channel and single-channel architectures.

Low-noise ADC and dynamic range: tying Nm/N precision to bandwidth

Once the bridge outputs have been amplified into a suitable voltage span, the ADC defines how finely the signal can be digitized and how much bandwidth can be used at a given noise level. In a typical robot application, the motion controller runs a servo loop at 500–1000 Hz, and force-control modes often target 200–300 Hz of usable F/T bandwidth. At the same time, the mechanical design might need ±500 N on a force axis with around 0.5–1 N resolution, and ±50 Nm on a torque axis with 0.05–0.1 Nm resolution. These targets immediately translate into requirements on ADC resolution, effective number of bits (ENOB) and overall noise.

The smallest meaningful force or torque step in the application corresponds to a small change in bridge voltage and, after amplification, to a small step at the ADC input. If that step is much smaller than the ADC’s effective LSB including noise, it will be invisible. The combination of INA gain, ADC reference and ENOB should therefore be chosen so that several ADC codes cover the target force or torque resolution, leaving comfortable headroom for mechanical tolerances, hysteresis and sensor nonlinearity. Nominal resolution alone is not enough; the ADC must deliver enough ENOB at the chosen sampling rate and bandwidth.

Multi-axis force/torque data is most useful when all six components are sampled at the same instant. Simultaneous-sampling ADCs provide a separate sample-and-hold per channel, so Fx, Fy, Fz, Tx, Ty and Tz are time-aligned. This simplifies force-control tuning, collision detection and vibration analysis. Multiplexed ADCs, in contrast, sequentially sample each channel, and the sampling skew between channels grows with channel count and conversion time. For high-bandwidth or dynamic applications, this skew can introduce apparent phase shifts between axes that complicate matrix calculations and control laws.

Digital filtering is the main tool for trading off noise and latency. Simple moving-average filters reduce noise roughly by the square root of the number of samples averaged, but they also introduce group delay proportional to the filter length. More sophisticated FIR and IIR low-pass filters can be tuned for specific bandwidth and ripple targets, yet they still add milliseconds of delay at kHz-level sampling rates. This delay accumulates with other sources in the control loop, so bandwidth and filter settings should be agreed between the force-sensor team and control engineers early in the project.

Oversampling and decimation provide a structured way to improve effective resolution within a given bandwidth. By sampling at a rate several times higher than the required signal bandwidth, then applying a well-designed digital filter and down-sampling, the system can push quantization and some noise energy out of band. However, oversampling does not fix poor analog front-end design: if INA noise or mechanical noise dominates, increasing the ADC sample rate alone will not improve the smallest resolvable Newton or Newton-meter. A practical approach is to view noise, effective resolution and bandwidth as a three-way trade-off and to select ADC resolution, sampling topology and filtering so that the chosen operating point fits comfortably inside the project targets.

Temperature compensation and calibration: from factory to field re-zero

Both the strain gauges inside a 6-axis force/torque sensor and the mechanical structure that carries them have strong temperature dependencies. Gauge resistance, adhesive properties and structural stiffness all vary with temperature, so the apparent sensitivity and cross-talk between axes drift as the module warms up or the environment changes. Without temperature-aware calibration, a robot may see force changes on supposedly idle axes or slow drift on zero-load readings even when the mechanical setup has not changed.

Practical modules therefore include at least one on-board temperature sensor, such as an NTC close to the bridge area or a digital temperature sensor on the PCB. During development, controlled loads are applied at multiple temperatures and the resulting outputs are compared against the expected forces and torques. The mismatch is recorded as temperature-dependent correction terms, and a lookup table or parametric model is derived so the digital core can compensate sensitivity and cross-talk changes as a function of temperature during normal operation.

Factory calibration is usually the most comprehensive step. The sensor is mounted in a fixture, exposed to a set of temperatures and load cases that cover the specified operating range, and all bridge channels are sampled simultaneously. From this dataset, the manufacturer solves for a calibration and compensation matrix that converts raw bridge readings into corrected Fx, Fy, Fz, Tx, Ty and Tz, including temperature terms. The final coefficients are programmed into non-volatile memory so that each module ships with its own calibrated behavior baked in.

In the field, end users rarely have access to a full calibration rig, but they still benefit from simple, repeatable procedures. A common practice is to perform a re-zero at power-up under known no-load conditions, and to schedule periodic no-load recalibration when the robot is in a safe pose. These steps correct for residual offset drift, mounting stress changes and long-term aging effects without touching the full factory compensation matrix. Some systems also allow a limited span or linearity trim against a reference load to fine-tune the response for a specific tool or fixture.

From an integration point of view, it helps to clearly separate the responsibilities of factory and field calibration. The factory owns the heavy lifting: multi-temperature characterization, matrix solving and programming. The user-side procedures focus on zero-offset management and simple checks that the module still behaves within specification. A well-documented calibration flow and a few dedicated commands in the module interface make it much easier for system integrators and maintenance teams to keep long-term force/torque accuracy under control.

Isolated SPI and power: hardware interface to the robot controller

The wrist of a robot is rarely a quiet electrical environment. Motor phases, brake coils, valve drivers and switching regulators share the same harness and mechanical space as the 6-axis force/ torque sensor module. In addition, the potential at the robot arm may not be identical to the cabinet ground, and ESD or surge events can occur at the tool or flange. For these reasons, it is common to isolate the F/T module’s digital interface and power from the robot controller, so that noise and ground shifts do not corrupt the measurements or damage sensitive front-end electronics.

A typical architecture uses a digital isolator for the data interface, such as SPI or UART, combined with an isolated DC-DC converter that provides a local, floating supply on the sensor side. The analog front end and ADC sit on the sensor ground domain, while the controller and motion card use the cabinet ground domain. A clearly defined isolation barrier separates these domains so that high-frequency common-mode noise, ground potential differences and transients see a robust dielectric rather than a direct copper path between them.

When selecting digital isolators, isolation voltage and common-mode transient immunity (CMTI) are key figures of merit. High CMTI ensures that rapid dv/dt events from nearby motor drives or DC-DC converters do not inject spurious edges into the isolated SPI signals. At the same time, the isolator must support the required SPI clock frequency and add acceptably low jitter so that the controller can still meet setup and hold times at the chosen bus speed. Designs that rely on precise timing or synchronous sampling across several isolated modules should pay particular attention to skew and propagation delay specifications.

On the power side, an isolated DC-DC converter—or a digital isolator with integrated isolated power— creates a local supply that follows the module’s load requirements without tying its ground directly to the robot controller. Input filtering, surge and ESD protection and sensible creepage and clearance on the PCB complete the isolation picture. The sensor analog ground can then be laid out for low noise and short return paths, while the controller ground is free to handle higher-current switching paths and other subsystems.

Above this hardware layer, different projects may choose SPI, RS-422/RS-485, CANopen or EtherCAT slave interfaces to integrate the F/T module into the broader control network. The force/torque page focuses on the physical isolation boundary and signal integrity aspects. Detailed protocol mapping, PDO layouts and network timing belong to higher-level communication and robot controller topics that sit on top of the isolated physical link from the sensor module.

EMI and PCB layout tips: survival of precision bridges in a servo cabinet

A 6-axis force and torque sensor module rarely lives on a quiet lab bench. In real robot cells it shares space with servo drives, motor cables, brake coils, contactors and switching power supplies. These sources inject high dv/dt, large current loops and fast transients into the cabinet and harness. The strain-gauge bridges and instrumentation amplifiers inside the module work with microvolt and millivolt signals, so a careless PCB layout can easily destroy most of the precision that was promised by the datasheets.

The most sensitive part of the signal chain is the short stretch between each bridge and its INA inputs. Bridge outputs should be routed as tightly coupled differential pairs, kept as short as possible and placed away from high dv/dt nodes such as gate drivers and switching converter loops. The INAs themselves should sit close to the connector or bridge interface so that unamplified lines do not wander across the board. Underneath the bridge and INA area, a continuous analog ground plane helps keep loop areas small and provides a stable reference for low-noise operation.

Clear separation between analog and digital regions is just as important as careful routing. The strain gauges, INAs, ADC and temperature sensor form an analog island with its own sensor ground. Digital logic, isolators and communication transceivers live on the digital or controller side. The sensor ground should be allowed to form a solid local plane and should converge in a star fashion at the INA or ADC reference point. Only after the isolation boundary does the design connect to the controller ground domain, so that noisy return currents from drivers and I or O modules cannot flow through the precision bridge area.

Reference voltage generation and temperature sensing deserve dedicated placement as well. The ADC reference node should be routed with short, well-decoupled traces and kept away from high-speed digital lines. Any small RC network used to filter the reference must be chosen so that it does not introduce excessive dynamic error. Temperature sensors should be located close to the mechanical path and strain-gauge area rather than near hot digital components, so that compensation tables reflect the real sensor conditions instead of microcontroller self heating.

Additional EMI countermeasures start with modest input RC filters and shielding rather than brute-force bandwidth reduction. Small series resistors and capacitors at the INA inputs can tame high-frequency interference if their cutoff frequency is chosen well above the target force-control bandwidth plus margin. Metal shields over the bridge and INA region, tied to the analog ground, help reject radiated noise. Connector pinouts should keep differential bridge or SPI pairs together with nearby ground or shield pins, rather than interleaving them with high-current or switching signals in the same shell.

For digital interfaces, small series resistors near the driver side of SPI lines help control edges and reduce ringing on long runs, and compact common mode chokes at the board edge can suppress common mode noise on external cables. The final objective is not to remove all noise, but to give high precision bridge and INA circuits a quiet local environment so they can deliver the microvolt level performance the application needs, even inside a busy servo cabinet.

Diagnostics and self test: preventing silent failures

A 6-axis force and torque sensor that fails quietly can be more dangerous than one that stops working outright. The module should therefore treat its own health as a first class output alongside Fx, Fy, Fz, Tx, Ty and Tz. A useful way to think about diagnostics is to divide the design into layers and ask what each layer can do to detect faults: the bridges and excitation, the INAs and analog front end, the ADC and sampling path, the digital processing and calibration, and the interface that reports data to the robot controller.

At the bridge and excitation level, open and short circuits on the strain-gauge networks or wiring can be detected by monitoring excitation current and by checking whether raw readings sit in a plausible range. Supply faults on the bridge excitation rail often disturb all axes at once and can be flagged as a dedicated error. On the INA and analog front end layer, saturation or loss of bias can be detected by watching for persistent rail to rail behavior on one channel while others remain normal, or by comparing readings against expected limits during known no-load and reference conditions.

The ADC and sampling subsystem can contribute its own self test mechanisms. Built in reference channels and over range flags help detect stuck codes, dead channels and clipping. Sampling clock faults or configuration errors often appear as jittery or intermittently frozen outputs, which can be detected by sanity checks on update timing and numerical variance. Above that, the digital processing layer should validate calibration data integrity with checksums, apply range checks after matrix operations and make sure that transformation from bridge units into Newtons and Newton meters does not produce impossible values under nominal poses.

Built in self test can go further by injecting known test stimuli. Some designs provide a small test excitation path that can apply a repeatable offset to the bridge or tie in a reference resistor network while the robot is in a safe pose. The module then verifies that the resulting digital output falls inside a tight acceptance window, which exercises the bridge, INAs, ADC and processing pipeline together. Additional consistency checks can use the physics of the system: at a defined zero pose with no external load, the net force and torque should be close to zero, and in certain motions one or two axes may have well defined relationships that can be monitored over time.

Finally, the interface to the robot controller should carry clear diagnostic information, not just force and torque values. Status bits can indicate which layer raised an issue and whether the condition is a warning or a hard fault. The status word can also suggest safe fallback actions such as disabling force control and reverting to position only control when data integrity is in doubt. With layered diagnostics and explicit health reporting, the F T module becomes a sensor that can speak up when its own readings are no longer trustworthy.

Brand & IC Mapping: Matching components for precision, bandwidth, and cost-sensitivity

Choosing the right combination of components is essential for a robust force/torque measurement system. This section covers three typical application scenarios—high precision, high bandwidth, and cost-sensitive solutions—and explains how to select the best combination of INA, ADC, Isolator, and Isolated DC-DC components. We’ll demonstrate the process of matching datasheet specifications with your application needs in terms of noise, bandwidth, sample rate, temperature range, and isolation voltage.

We’ll explore three examples, each representing one of the key requirements for force/torque modules, and show how datasheet indicators such as noise, bandwidth, and isolation voltage can guide your decisions.

High Precision IC Combination

For high-precision applications such as scientific research and laboratory measurements, the required IC combination must ensure ultra-low noise and high stability.

• INA: Low noise, high-precision INA like INA333 or INA125.
• ADC: 24-bit precision, low-noise ADCs such as ADS1256.
• Isolator: High isolation voltage digital isolators like ISO1540.
• Isolated DC-DC: Precision temperature-compensated isolated DC-DC converters like LTM2885.

Matching datasheet specifications:

• Noise: nV/√Hz.
• Temperature range: -40°C ~ 85°C.
• Isolation voltage: ≥ 3.75kV.

High Bandwidth IC Combination

For high-bandwidth applications, such as robotics or sensor networks, the IC combination must support real-time response and fast sampling rates.

• INA: High-speed INA, such as INA333 or INA128.
• ADC: High bandwidth ADCs like ADS5400 or LTC2325.
• Isolator: High-speed SPI isolators such as ISO7842.
• Isolated DC-DC: High-bandwidth, wide voltage input isolated DC-DC converters such as LTM8066.

Matching datasheet specifications:

• Bandwidth: ≥ 20MHz.
• Sampling rate: ≥ 1 MSPS.
• Isolation voltage: ≥ 3kV.

Cost-Sensitive IC Combination

For cost-sensitive applications, where performance is still important but the cost must be optimized, a more affordable IC combination will suffice while maintaining sufficient accuracy.

• INA: Cost-effective INA like INA333 (low-cost version).
• ADC: 16-bit ADCs such as MCP3208.
• Isolator: Cost-effective isolators like ISO122.
• Isolated DC-DC: Cost-effective Isolated DC-DC converters like LTC3880.

Matching datasheet specifications:

• Noise: Moderate levels.
• Sampling rate: 200kSPS ~ 1MSPS.
• Isolation voltage: ≥ 2kV.

The right IC combination depends on the specific needs of your application, but by focusing on key metrics like noise, bandwidth, and isolation voltage, you can quickly match the right components to achieve optimal performance while meeting your project goals.

FAQs: Answering 12 Key Design Questions for Force/Torque Sensors