A robot gripper closing on a wet mug generates normal force, shear, and micro-slip events at the contact interface. A camera aimed at that same gripper records none of them. That is the core measurement gap behind the current push toward tactile sensing in robotics: grip force, slip, compliance, and the shear that precedes rotational slip are physical quantities that vision cannot measure directly, only infer — poorly — from pixel motion.

The distinction matters for any cell doing grasping, insertion, press-fitting, sanding, or polishing. Those processes succeed or fail on forces at the point of contact, and those forces demand sensors, not inference.

What touch actually measures

Human skin offers a useful reference architecture. It contains several mechanoreceptor types, each tuned to a distinct signal: sustained pressure, vibration, skin stretch, or dynamic response. The brain fuses these channels into grip tightening on slip detection, regrasping on instability, and texture recognition without visual input. A recent review by Yao, Wu, Gao, et al. (Nano-Micro Letters, vol. 18, p. 37, 2026) maps these receptor classes and their static and dynamic response patterns — a useful baseline for specifying what an artificial system should capture.

In practical terms, the measurable tactile modalities include contact size and location, contact intensity, normal and shear force, vibration, temperature, and proprioception — the orientation data that reveals how an underactuated gripper has actually closed around a part. Downstream, signal processing or machine learning converts these into grasp-stability estimates, slip detection, in-hand manipulation control, object and texture recognition, and orientation estimation.

Six transduction principles, six trade-offs

No single sensing mechanism dominates. Each offers a different balance of sensitivity, spatial resolution, bandwidth, durability, and force range:

• Piezoresistive: simple construction, high sensitivity, but drift over time degrades stability.
• Capacitive: low power, fast response, limited spatial resolution.
• Piezoelectric and triboelectric: self-powered and fast, but blind to static, sustained forces.
• Magnetic: clean decoupling of three-dimensional force, but requires a controlled magnetic environment.
• Optical (vision-based tactile): the highest spatial resolution available, paid for with a more complex optical system and constrained form factor.

The selection question is therefore application-driven, not technology-driven. Object geometry, force range, the need for low-force sensitivity, and whether the task demands a simple stable grasp or full in-hand dexterity together narrow the viable modalities — and reveal early what data processing the system will require.

Why deployment lags the lab

Robotic skins and optical fingertip sensors have circulated in academia for years. Production adoption is only now beginning, and the obstacles are instructive for anyone planning a deployment.

First, the hardware: sensors can be fragile, expensive, and noise-sensitive. Second, the signals: tactile data is high-dimensional and tightly coupled to contact mechanics, which makes real-time modeling and control genuinely hard. Third, interpretability: given a contact map from a single grasp, few human experts could identify the object or judge grip quality by eye — which complicates the labeling and annotation workflow that vision pipelines take for granted.

Then there is the data problem, and it is the deepest one. Vision enjoys internet-scale datasets; touch has nothing comparable. Large, diverse, high-quality tactile data remains difficult to capture, and the absence of standard interfaces, data formats, and benchmarks has slowed adoption across the board. Transferring a learned model from one sensor type to another is still an open research problem. Even Meta's Sparsh project, which targets general touch representations at scale, reports that generalizing across dissimilar sensor types remains unsolved. Failure cases compound this: grasps that look correct but fail are precisely the examples hardest to collect and most underrepresented in datasets, capping how well learned models generalize.

Buyers should read published results with that in mind. Many rest on small, unbalanced datasets, so first-pass real-world performance can disappoint relative to benchmarks — a calibration of expectations rather than a disqualification.

Specifying from the task backward

The practical path is to start with the application, not the transducer. The questions that determine the specification: What is being manipulated — simple surfaces or awkward geometries? How many object types, and how heavy or unbalanced? Is the task resolving grasp-pose uncertainty, recognizing objects, or performing genuine in-hand manipulation? What force range applies, and does the low-force end need extra sensitivity? Answering these narrows the sensor field and surfaces the data-processing burden before commitments are made.

None of this positions touch as a replacement for vision. The two modalities answer different questions. Vision is unmatched for geometry, location, and recognition at a distance; tactile and force sensing captures the contact quantities vision cannot see at all. The discipline that separates a reliable deployment from a fragile demo is deciding, up front, what the task critically depends on measuring — then choosing sensors that measure it directly.

The convergence of capable hardware, foundation models, better simulation, and genuine industry interest points to rapid progress from here. The open question for integrators is standards: without common interfaces, data formats, and benchmarks for tactile data, every deployment will keep reinventing its own measurement chain — and cross-sensor model portability, the thing that made vision cheap, will stay out of reach.