Perception and Sensors
Force Torque Sensors: Which Numbers Actually Matter
Rated range is the least useful number on a force torque datasheet. Resolution, crosstalk, drift and overload capacity decide whether the sensor works in your task.
Size on resolution and noise, not on rated range. A sensor rated to 500 N with 0.5 N of noise is worse for a 20 N assembly task than one rated to 100 N with 0.05 N of noise, because you need to resolve the signal, not survive it. The three numbers that decide a project are noise floor, crosstalk and overload capacity, and only the last is usually printed prominently.
The specifications that matter
| Specification | Why it matters | Typical good value |
|---|---|---|
| Noise floor | Sets the smallest force you can actually detect | 0.01 to 0.1 N |
| Resolution | Quantisation of the reading | 0.005 to 0.05 N |
| Crosstalk | A force on one axis appearing on another | under 2 % full scale |
| Overload capacity | Survival of a crash | 5 to 20x rated |
| Thermal drift | Zero shift as the sensor warms | under 1 % full scale per shift |
| Output rate | Whether it fits the control loop | 1,000 Hz or more |
| Latency | Delay from event to reading | under 1 ms |
| Stiffness | Deflection under load, affects position accuracy | as high as the task allows |
| Rated range | Where the reading saturates | Choose 2 to 3x expected peak |
Matching sensor to task
| Task | Force range | Resolution needed | Rate |
|---|---|---|---|
| Peg in hole assembly, small parts | 1 to 30 N | 0.02 N | 1,000 Hz |
| Connector insertion | 5 to 60 N | 0.05 N | 1,000 Hz |
| Polishing and deburring | 5 to 150 N | 0.1 N | 500 to 2,000 Hz |
| Press fitting | 100 to 2,000 N | 1 N | 500 Hz |
| Collision detection on a cobot | 10 to 200 N | 0.5 N | 1,000 Hz |
| Payload identification | 1 to 300 N | 0.05 N | 100 Hz |
| Surface following | 2 to 50 N | 0.05 N | 1,000 Hz |
Do you need an external sensor at all
Three ways exist to sense force on a robot, and the external six-axis sensor is the most accurate and the most expensive.
- Motor current estimation. Free, already present, and crude. Friction and inertia dominate the signal, so resolution is on the order of several newtons and direction information is poor. Adequate for crash detection, useless for assembly.
- Joint torque sensors. Built into many collaborative arms. Good sensitivity, and the estimate at the tool degrades with distance from the sensed joints and with model error.
- External six-axis sensor at the flange. Measures directly where the work happens. Best signal, adds 0.3 to 1.5 kg of mass, 20 to 60 mm of length, a cable, and a real cost.
The length penalty deserves attention. Adding 40 mm between flange and tool moves the load centre of gravity outward, which reduces usable payload on a torque-limited wrist by a noticeable margin.
Practical notes
- Bias the sensor with the tool attached, in the pose where work happens. Gravity compensation for the tool mass has to be applied continuously as the orientation changes.
- Let it warm up. Thermal drift over the first 20 to 40 minutes is normal and larger than the signal in delicate tasks.
- Filter carefully. Aggressive low-pass filtering removes the transient that indicates contact, which is exactly the event of interest.
- Protect against overload. A crash at full speed can exceed rated capacity by an order of magnitude. Mechanical stops or a breakaway coupler cost less than the sensor.
- Check crosstalk in your real load case. A long tool converts a small lateral force into a large torque, and crosstalk error scales with it.
Frequently asked questions
How do I choose a force torque sensor range?
Two to three times the expected peak force. Larger ranges waste resolution, because the noise floor generally scales with the rated range, and detecting a 5 N contact with a 2,000 N sensor is not practical.
Are joint torque sensors enough?
For collision detection and coarse force control, usually yes. For assembly tasks needing sub-newton resolution at the tool, an external six-axis sensor at the flange is substantially better, because the estimate does not depend on a robot model.
What is crosstalk and why does it matter?
A force on one axis producing a reading on another, typically 1 % to 3 % of full scale. With a long tool a small lateral force becomes a large torque, so crosstalk error scales with tool length and can exceed the signal you are looking for.
Why does my force reading drift?
Thermal drift, usually 0.5 % to 3 % of full scale over a shift, plus uncompensated tool weight as the orientation changes. Warm the sensor up, re-bias in the working pose, and apply continuous gravity compensation.
Can motor current replace a force sensor?
Only for crash detection. Friction and inertia dominate the current signal, giving resolution on the order of several newtons with poor directional information, which is far short of what assembly tasks require.
Sources
- ISO 10218-1:2025, Robotics, Safety requirements, Part 1International Organization for Standardization, force limiting and sensing requirements
- ISO/TS 15066:2016, Annex AForce limits that measurement equipment must resolve
- Robotics at NISTNational Institute of Standards and Technology, robot performance measurement and test methods