Tactile & Force Sensors from China: Supplier Landscape

Why this category is harder to shop than actuators or arms

Actuators have a torque number and a bus you can sanity-check. Tactile and force sensors don't compress that cleanly. A "40,000-units/cm²" density figure and a "6-axis" label tell you almost nothing about whether the sensor will survive your gripper's duty cycle, stream fast enough for your control loop, or ship with a driver your stack can actually read.

The segment is also genuinely new. Most of these products are one or two hardware generations old, datasheets are terse, and cross-vendor benchmarks barely exist. Treat every spec below as a starting point for an RFQ conversation, not a settled fact — and see how to buy robots from China for the PI and inspection discipline that matters doubly here.

The three sensor types you'll actually evaluate

Fingertip tactile arrays. A grid of sensing elements (taxels) on a fingertip or gripper pad, reporting normal force and often shear and contact location. The two common transduction methods are capacitive (dense, thin, mature, but prone to drift and crosstalk) and integrated/piezoresistive tactile units such as PaXini's ITPU approach, which packs many taxels with fine force resolution. These live on dexterous hands — PaXini's own DexH13 GEN2 hand carries ≈978 ITPU taxels at 0.01 N resolution, and Unitree's Dex5-1 lists 94 touch points. If you're specifying a hand, the tactile skin usually comes bundled; if you're retrofitting, you're buying the array separately.

Vision-based (optical) tactile. A small camera inside a soft, deformable gel dome watches the surface deform on contact; software reconstructs a high-resolution contact map, and often slip and texture, from the image. The GelSight lineage popularized this; Daimon Robotics' DM-Tac W is a Chinese production example, quoting an extremely high effective sensing density on a millimetre-thin surface. Strengths: spatial resolution no capacitive grid matches, and rich shear/slip data. Costs: the gel is a wear item, latency is bounded by the camera frame rate, and you inherit a vision-processing load.

6-axis force/torque (F/T) sensors. A single sensor — typically strain-gauge — mounted at the wrist that resolves the full wrench: three forces (Fx, Fy, Fz) and three torques (Mx, My, Mz). This is the workhorse for robotic-arm force control, assembly, polishing and safety limiting. Kunwei's KWR75 series is a representative 75 mm-flange-class part aimed at cobot wrists. Unlike tactile skin, F/T sensors are a comparatively mature, well-understood format — the buying question is range/overload and drift, not novelty.

Where they go: manipulation and safety

• Manipulation. Fingertip arrays and vision-based skin close the grasp loop: detect contact, estimate grip force, catch incipient slip before the part drops, and localize edges for insertion. A wrist F/T sensor handles the coarser task-level wrench (peg-in-hole, wiping, force-limited placement).
• Safety and compliance. A wrist F/T sensor lets a cobot detect a collision and stop, or hold a commanded contact force against a surface. This is force limiting and compliant control, not a substitute for a rated safety-certified stop — verify certification claims independently and design your safety function to your local machinery regulation.

The two layers are complementary: F/T at the wrist for the arm's control loop, tactile at the fingers for the grasp. Many manipulation stacks use both.

Selection criteria that actually decide it

1. Range and overload. For F/T, match rated force/torque to your task and check the overload rating — a wrist sensor that's fine at rated load can be destroyed by a single hard collision. For tactile arrays, ask the per-taxel force range, not just resolution.
2. Resolution. Force resolution (e.g. 0.01 N per taxel) and spatial resolution (taxel pitch, or camera pixels for vision-based) are separate numbers. High spatial density with coarse force resolution suits texture/slip; the reverse suits grip-force control.
3. Sampling rate / latency. This is the spec most often buried. A tactile sensor feeding a real-time grasp loop needs low, deterministic latency — PaXini quotes <10 ms response for its multi-dim array. Vision-based sensors are frame-rate bound; confirm the actual streamed rate, not the camera's raw capability.
4. Interface. Confirm the electrical and data interface end to end: analog vs digital, the bus (RS-485, CAN, EtherCAT, USB, Ethernet), and — the real gate — whether a documented SDK/ROS driver exists in a language your team reads. A sensor with no usable driver is a research project, not a component.
5. Form factor and durability. Flange size and bolt pattern for F/T; pad geometry and cabling for fingertip arrays; gel replacement interval and cost for vision-based. Thin, conformable skin that fits your existing fingers is worth more than a denser sensor that doesn't.
6. Calibration and drift. Ask how units are calibrated, whether calibration is per-unit or per-batch, temperature sensitivity, and long-term zero drift. This is where cheap sensors quietly disappoint.

Reference makers (catalog)

All three are quoted per-unit or per-array via RFQ — there is no reliable public street price for this category as of 2026, so we don't publish price bands we can't stand behind. MOQ is 1 unit across the board.

For context on bundled fingertip sensing, PaXini's DexH13 GEN2 hand (≈978 ITPU taxels · 0.01 N) and Unitree's Dex5-1 (94 touch points) are covered under dexterous hands.

The wider landscape, qualitatively

Beyond these three, China's tactile/F/T scene includes a growing set of university spin-offs and sensor startups riding the humanoid and cobot wave, plus established industrial F/T suppliers competing on price against Western references. Chinese 6-axis F/T sensors are typically quoted well below Western incumbents — the gap is real, but so is the variance in calibration quality and documentation. We're deliberately not ranking the long tail: in a segment this young, a maker's standing can shift in a single product generation, and per-company announcements are the main public data source, which we don't treat as verified.

Honest caveats before you buy

• It's early. Datasheets are thin, third-party benchmarks are scarce, and generation-to-generation changes are large. Expect to do your own characterization.
• Headline numbers oversell. A huge taxel or density figure is a marketing anchor; latency, drift, overload and driver quality decide whether the sensor works in your loop.
• The SDK is the product. As with actuators, the single best predictor of success is buying one unit and integrating it before ordering a fleet. If the driver, protocol doc and example code are complete, the rest follows.
• Safety claims need independent verification. Force limiting is not a certified safety stop. Verify any certification against your destination's machinery regulation, and see how to buy robots from China for putting acceptance tests on the proforma invoice.

When you're ready to compare live quotes across these makers, send an RFQ with your force range, target sampling rate, interface and form factor — those four lines get you a usable quote faster than a model name alone.