To pick up an egg without crushing it, a robot needs to know how hard it is squeezing and whether the egg is starting to slip. Two kinds of sensors provide that sense: 6-axis force-torque sensors that measure overall force at the wrist or joints, and tactile sensors that read contact at the fingertips. This article confirms how each works from public sources, calculates actual signal sizes, and offers selection criteria by task.
Two layers of force sensing: net force and contact
| Type | What it measures | Usually mounted | Typical uses |
|---|---|---|---|
| 6-axis force-torque (F/T) sensor | Three forces and three moments (net load) | Wrist, ankle, joints | Force control in assembly and polishing, contact detection, ground reaction in walking |
| Tactile sensor | Contact location, contact shape, pressure distribution, slip | Fingertips, palm, skin | Delicate grasping, in-hand manipulation, slip prevention |
As ATI describes it, a multi-axis F/T sensor measures all six components of force and torque and is used for product testing, robotic assembly, grinding and polishing, as well as research in surgical robotics and haptics. What the net force cannot tell you is where on the fingertip an object is touching, or whether the contact patch is creeping — that is what tactile sensors are for.
How 6-axis F/T sensors work: strain gauges and bridges
Many F/T sensors bond strain gauges to a metal structure designed to flex slightly under load. A strain gauge becomes narrower and longer when stretched, raising its resistance, and the reverse when compressed. The ratio of fractional resistance change to strain is the gauge factor, a little over 2 for common metallic foil gauges. Semiconductor gauges have larger gauge factors but are more expensive, more temperature-sensitive and more fragile.
The resistance change is tiny, so it is read with a Wheatstone bridge. For a bridge with one active gauge, the output is approximately:
where is the excitation voltage, the gauge factor and the strain. Working through real numbers shows why precision amplification is unavoidable:
def bridge_output_mv(excitation_v, gauge_factor, strain):
return excitation_v * gauge_factor * strain / 4 * 1000 # mV
# 5 V excitation, gauge factor 2, 500 microstrain (assumed values)
print(bridge_output_mv(5.0, 2.0, 500e-6)) # 2.5 mV
With outputs of a few millivolts, an F/T sensor also contains amplifiers, an analog-to-digital converter and a calibration matrix that maps many gauge signals to the six force and torque components. Temperature shifts the zero point, so zeroing before a task and temperature compensation are part of normal use.
Camera-based tactile sensors: photographing how a membrane deforms
Fingertip tactile sensors include arrays whose resistance or capacitance changes with pressure, but much recent research uses camera-based designs. A soft elastomer membrane is lit from inside and photographed; when an object presses into it, the imprint and its changes show up in the image. From that image the sensor estimates contact shape, pressure distribution and slip.
GelSlim builds on GelSight techniques and redesigns the optical path from light source to camera to shrink the sensor to finger size. The authors improved durability with tougher gel materials, a textured fabric skin, a rigid body and a calibration process, and evaluated it by tracking signal-quality metrics over more than 3,000 grasping experiments. Wear of the contact surface from repeated friction is the first limit teams hit when putting camera-based sensors into real work.
DIGIT focuses on making the same class of sensor cheap and compact. Its paper shows neural-network controllers trained on DIGIT data manipulating glass marbles in-hand with a multi-finger robot hand. Because the tactile output is an image, it slots into learning pipelines the same way camera data does.
How products use them
Humanoid makers now treat fingertip sensing as a selling point. Figure’s Figure 03 announcement says its fingertip sensors detect forces as small as 3 grams, and palm cameras keep close-range objects in view when the head cameras are blocked. Sensor configurations of other platforms are summarized in Humanoid platforms compared.
From signals to control
What matters is less the sensor itself than how the controller or learned model uses it. A wrist F/T sensor usually feeds a force-control loop: if the arm presses harder than the target force it backs off, and if it presses too lightly it pushes further, keeping contact force steady. Detecting the sideways force when a part catches on the edge of a hole during assembly, and correcting the position, works the same way.
Tactile signals are widely used for slip detection: when the contact imprint starts creeping to one side, the gripper increases its grip force. More recent work feeds tactile images into learned models alongside camera frames, letting the model see contact states that vision alone misses — useful for rotating objects in-hand or handling transparent items. How such models generate actions is covered in How diffusion policies work.
In walking robots, ankle F/T sensors measure ground reaction forces to judge when a foot touches down and where the center of mass is shifting. The connection to whole-body balance is covered in Bipedal locomotion and whole-body control.
Choosing by task
- Holding a steady force, as in assembly or polishing — control the net force with a wrist 6-axis F/T sensor.
- Grasping fragile objects or preventing slip — detect contact changes with fingertip tactile sensors.
- Detecting ground contact in walking robots — consider ankle F/T sensors alongside current-based estimation in the joints.
- Values to check before choosing a sensor — maximum range, overload limit, update rate, zero drift and how often the contact surface must be replaced.
An accurate sensor is useless if the control loop is slow. Check that the sensor’s update rate is comfortably faster than the force-control loop and how much latency the signal path adds.
Calibration and maintenance in daily use
Both sensor families need routine care that is easy to overlook in early prototypes. F/T sensors should be zeroed with no load before each task, because temperature changes and mounting stress shift the baseline; if the tool mounted on the sensor changes, its weight and center of mass must be re-entered so the controller does not mistake gravity for contact. Camera-based tactile sensors need their elastomer surfaces inspected and replaced on a schedule, and a short recalibration after replacement keeps lighting and contrast consistent — the same concern the GelSlim authors addressed with their calibration process. Logging these steps alongside task data makes it much easier to tell a real change in contact behavior from sensor drift when a grasp suddenly starts failing.