Perception and Sensors
Tactile Sensing: The Sense Robots Still Lack
Humans detect slip within about 100 milliseconds and adjust grip. Robots mostly cannot. What tactile technologies exist, what they measure, and why almost none reach production.
Humans regulate grip force from tactile feedback within roughly 100 ms of a slip beginning, using around 17,000 mechanoreceptors per hand. Robots overwhelmingly do not have this. Without it, a gripper must either squeeze hard enough to guarantee no slip, which damages fragile items, or model the object perfectly in advance, which fails on variation. That is the practical cost of the missing sense.
What tactile sensing actually measures
The word covers several distinct quantities, and a sensor that provides one may provide none of the others.
- Normal force distribution. How hard, and where across the contact patch. The most commonly implemented quantity.
- Shear force. Lateral load, which is what indicates incipient slip. Much harder to measure and far more useful.
- Contact geometry. The shape of what is being touched, which enables in-hand pose estimation.
- Vibration. High-frequency signal from texture and from micro-slip, detectable before gross slip occurs.
- Temperature. Rarely useful in industry, occasionally used for material discrimination.
| Technology | Measures | Spatial resolution | Rate | Weakness |
|---|---|---|---|---|
| Resistive array | Normal force | 2 to 10 mm | 100 to 1,000 Hz | Drift, hysteresis, wear |
| Capacitive array | Normal force, some shear | 1 to 5 mm | 100 to 1,000 Hz | Sensitive to humidity and nearby metal |
| Piezoresistive | Normal force | 1 to 5 mm | 500 to 5,000 Hz | Temperature dependence |
| Optical, camera behind a gel | Geometry, shear, force | 0.05 to 0.5 mm | 30 to 100 Hz | Bulky, gel wears out, slower |
| Barometric cells | Normal force | 5 to 15 mm | 100 to 500 Hz | Coarse, but robust and cheap |
| Acoustic and vibration | Micro-slip, texture | whole finger | 1 to 10 kHz | No spatial information |
Why so little reaches production
- Durability. A fingertip is the part of a robot that collides most. A sensing surface that degrades after a few hundred thousand contacts is a consumable in a machine designed to run for years.
- Wiring. Hundreds of sensing elements in a finger means a bundle of conductors through the most articulated joints on the machine, which is the same problem that makes dress packs fail.
- Calibration drift. Resistive and capacitive elements drift with temperature, humidity and age, and recalibrating a hundred elements in a fingertip is not a field procedure.
- Software. Raw pressure maps are not directly actionable. Turning them into a grip force decision requires a learned or hand-designed policy that most integrators do not have.
- Cheaper alternatives exist. A form-fit gripper that captures a feature removes the need to detect slip at all, and it costs a fraction of a sensorised hand.
What is used instead
Because tactile sensing rarely reaches production, cells solve the same problems by other means, and the alternatives are worth knowing before specifying a sensorised hand.
| Approach | Replaces | Cost | Limitation |
|---|---|---|---|
| Form-fit gripper capturing a feature | Slip detection | low | Needs a consistent part geometry |
| Vacuum pressure monitoring | Grip confirmation | very low | Only for vacuum grips |
| Gripper position feedback | Object presence and size check | none, built in | Cannot detect a slow slip |
| Flange force torque sensor | Contact detection | medium | No spatial distribution |
| Camera check after lift | Grasp verification | low to medium | Adds 150 to 600 ms per cycle |
| Compliant fingertips | Force limiting | very low | Passive, no measurement at all |
The first and last rows carry most of the load in real cells. Mechanical design that removes the need to sense is almost always cheaper than sensing, and that remains true even as tactile technology improves.
Where it does pay today
| Application | Quantity needed | Why nothing else works |
|---|---|---|
| Handling fragile food items | Normal force, low threshold | Minimum viable grip force varies per item |
| Deformable object handling | Contact geometry | Shape changes between perception and contact |
| In-hand reorientation | Geometry and shear | Vision is occluded by the hand itself |
| Cable and connector assembly | Shear and vibration | Seating is felt, not seen |
| Unknown object grasping | Slip detection | Friction coefficient is unknown in advance |
The pattern across all five rows is that vision cannot supply the signal. Where a camera can see what is happening, vision remains cheaper and more robust, which is why tactile sensing has stayed in the laboratory for tasks where sight suffices.
Frequently asked questions
Why do robots need tactile sensing?
To regulate grip force from contact rather than from assumption. Without it a gripper either squeezes hard enough to guarantee no slip, damaging fragile items, or relies on a friction model that fails on variation.
Which technology is most capable?
Optical tactile sensing, where a camera images a deformable gel from inside, giving geometry and shear at 0.05 mm to 0.5 mm resolution. It is also the bulkiest, runs at camera rates and treats the gel as a consumable.
Why is tactile sensing rare in industry?
Durability, wiring through the most articulated joints, calibration drift, the software needed to act on raw pressure maps, and the fact that a form-fit gripper often removes the need entirely at a fraction of the cost.
Can force torque sensing replace tactile sensing?
Partly. A flange-mounted sensor gives the resultant force and torque but no spatial distribution, so it detects that contact occurred without indicating where or whether one finger is slipping.
Where is tactile sensing worth the cost today?
Where vision cannot supply the signal: fragile food handling, deformable objects, in-hand reorientation, connector seating and grasping objects whose friction is unknown in advance.
Sources
- arXiv robotics preprints, tactile sensing and manipulationPrimary literature on sensor technologies and slip detection
- Robotics at NISTNational Institute of Standards and Technology, robot performance measurement and test methods
- ROBOTS guideIEEE Spectrum, hardware context for sensorised hands