Chapter 8: Hands, Grippers, and Touch — The Contact-Intelligence Supply Chain

Overview

The question of this chapter is: When does a dexterous hand or tactile sensor become a dependable contact-intelligence system rather than an impressive component? A gripper can close, a five-finger hand can reproduce a human pose, and a tactile array can stream thousands of values. None of those facts alone establishes that a robot inserts a connector more reliably, drops fewer objects, damages fewer surfaces, needs fewer resets, or earns back its lifecycle cost.

The thesis is that hands and tactile systems matter only when sensing, durability, interfaces, and policy benefit are jointly evidenced. Degrees of freedom determine a motion space, not useful dexterity. Sensor resolution describes a measurement channel, not calibrated force accuracy across temperature and wear. Coverage describes where contact may be observed, not whether the controller uses that observation. A product demo is evidence of possibility under selected conditions; a deployment requires repeatable task completion, replacement procedures, versioned software, safe failure behavior, and known cost.

LinkerBot, PaXini, and Tashan Technology are the three locked deep profiles in this chapter ^{[3] [7] [12]}. They occupy different layers. LinkerBot centers on dexterous hands, teleoperation, and manipulation data. PaXini combines Hall-effect tactile sensing, hands, robots, and data services. Tashan centers on tactile chips, fingertip and gripper sensors, electronic skin, and visuo-tactile training platforms. The selection is an evidence-controlled comparison, not a declaration that they are the only relevant suppliers or a cross-company performance ranking.

After reading this chapter... - You can separate degrees of freedom, actuation, tactile modality, spatial coverage, calibration, and policy benefit. - You can read issuer specifications without converting them into independent durability, shipment, or task-success evidence. - You can compare LinkerBot, PaXini, and Tashan across identity, products, technology, traction, maturity, interfaces, ecosystem visibility, WRC status, strengths, limits, and confidence. - You can estimate lifecycle cost from hand price, skins, fingers, cables, calibration, downtime, integration, and human intervention. - You can design a connector-insertion pilot whose logs reveal whether touch actually changes action.

8.1 The Contact-Intelligence Stack

Contact intelligence is a chain, not a sensor. The mechanical layer creates contact through fingers, jaws, tendons, linkages, gears, or direct-drive joints. The sensing layer converts deformation, force, proximity, vibration, temperature, or an optical image into signals. The estimation layer turns signals into contact location, normal and shear force, slip, texture, or object state. The policy chooses an action. The low-level controller executes it under force, torque, position, speed, and safety limits. The operating system records success, wear, reset, and replacement.

An error at any layer can resemble an error at another. A loose fingertip changes calibration. A worn elastomer looks like distribution shift. A cable delay resembles an unstable force loop. A kinematic mismatch makes the policy demand impossible fingertip poses. A controller that receives tactile data but ignores it can produce the same failures as a sensor that is absent. Procurement must therefore test the chain end to end.

A parallel gripper is often the correct baseline. It has a smaller action space, simpler collision geometry, fewer cables, and easier cleaning. A multifinger hand earns its complexity only when the task needs object reorientation, tool use, shape adaptation, narrow access, handover, or multiple contact modes. Similarly, tactile sensing earns its place when vision is occluded, tolerances are uncertain, friction changes, objects deform, or safe force matters. The right question is not “how humanlike is it?” but “which uncertainty does it remove at acceptable lifecycle cost?”

8.2 Degrees of Freedom, Actuation, and Task Fit

A hand may advertise total DoF, active DoF, or independently controlled DoF. These are not interchangeable. Passive joints can conform to an object without adding independent commands. Coupled tendons or linkages can give several moving joints but fewer control dimensions. Underactuation is not inferior by definition: it can reduce motors and allow robust enveloping grasps. It is a limitation when a task needs independent fingertip placement or in-hand reorientation.

Actuation determines more than force. Tendon routing can move motors away from fingers and reduce distal mass, but tendons stretch, wear, and require tension management. Linkages can be compact and self-locking, but introduce coupled motion and backlash. Direct drive can expose each joint to control and simplify transmission models, while increasing packaging, cable, thermal, and mass demands. Servo-based research hands lower price and improve repairability, but continuous load can produce heat ^[21].

Payload is especially easy to misuse. A hand holding a static mass with an enveloping grasp does not prove fingertip force, dynamic payload, impact tolerance, precision, or safe mounting on an arm. A claimed 20 kg load must retain the issuer, hand orientation, grasp type, duration, and laboratory qualification. The arm must also carry the hand and object without losing acceleration or safety margin.

Task fit is better expressed as a matrix. A suction or parallel gripper may dominate for planar boxes. A compliant gripper can handle shape variation. A five-finger hand may justify itself for tools, plugs, knobs, folded material, and tasks designed for human hands. Touch adds the most where final state is visually hidden: connector seating, clip engagement, slip onset, screw start, or fragile compression.

8.3 Deep Profile I — LinkerBot

The Korean rendering is 링커봇, the English brand is LinkerBot, and the Chinese brand is 灵心巧手. The official legal name shown by the 2026 World Robot Conference is 灵心巧手（北京）科技股份有限公司. LinkerBot's official site identifies its headquarters in Haidian District, Beijing ^[1]. A foundation year is not stated in the captured official company sources; it remains undisclosed here rather than being inferred from registration aggregators.

Flagships include the Linker Hand family, Open TeleDex teleoperation and data collection, LinkerSkillStore, and the Linker Genesis model ^[2]. The portfolio spans tendon, linkage, and direct-drive routes. The official WRC exhibitor page describes O30 as a high-DoF direct-drive hand with independently moving joints, L30 as tendon-driven, O6 as a compact linkage hand, and the data system as synchronizing vision, touch, and proprioception ^[3]. Those descriptions establish architecture and intended use, not matched task success.

The principal technical strength is breadth across hand mechanism, controller, teleoperation, and data. It allows a buyer to ask whether one data interface can span several hands rather than treating every mechanism as an isolated device. Open TeleDex is also described in a research manuscript, but the paper's device matrix and tests must be read under their stated embodiments and cannot be promoted into commercial uptime ^[6].

The developer surface is comparatively visible. The official developer center lists manuals, SDK tools, CAN protocol documentation, demonstrations, connectors, and 3D models ^[4]. An official L20 manual links Python, ROS, and ROS 2 SDK repositories ^[5]. This verifies the existence of those interfaces, not complete feature parity, deterministic timing, security support, long-term API stability, or compatibility with every ROS distribution. A validated NVIDIA Isaac Sim or Isaac Lab package was not established in the captured official evidence and is therefore unverified.

Tactile modality, coverage, and calibration vary by hand and option. The portfolio claim of “touch” cannot be treated as one uniform sensor specification. For the O30 highlighted by WRC, the organizer page verifies direct drive and 20 fully independent active degrees of freedom, but it does not disclose tactile modality, coverage map, calibration procedure, bandwidth, drift, or ingress rating. For procurement, those fields remain model-specific and undisclosed until a dated data sheet and acceptance test are supplied.

Finance is not publicly resolved by the captured primary evidence. Funding total, revenue, gross margin, cash flow, and valuation are undisclosed or unverified here. The company website makes issuer claims about market share and thousand-unit production; those claims lack a matched independent denominator and audit in the evidence packet, so this chapter does not use them as comparative facts. WRC organizer evidence does verify 2026 exhibitor status, legal name, booth B205, and displayed products including O30 ^[3]. It does not verify shipment, policy performance, or factory uptime.

Commercial maturity is best assessed as a marketed product family with public developer resources and organizer-verified exhibition, while long-duty industrial evidence remains incomplete. Price is quote-based or undisclosed in the captured official pages. Warranty duration, rated lifecycle by joint, tendon or fingertip replacement interval, mean time between failures, repair turnaround, and consumable pricing are also undisclosed. Those omissions are material because the inexpensive hand that stops a cell can be costlier than a higher-priced device.

Media-corpus visibility is unmeasured: no complete eligible-news universe, date window, language rule, or duplicate policy was available. WRC visibility is verified only for the official 2026 listing. Strengths are architecture breadth, an accessible developer surface, teleoperation/data integration, and multiple task-fit options. Limits are uneven public model specifications, no matched independent endurance or production benchmark, and unknown lifecycle cost. Confidence is medium for identity, headquarters, portfolio, developer resources, and WRC status; low for finance, price, field durability, and policy benefit.

The official product view below identifies the articulated fingers and palm form of the Linker Hand family. Because it is a staged product view, it should not be read as evidence of sensor coverage, durability, or field performance.

8.4 Deep Profile II — PaXini

The Korean rendering is 파시니, the English brand is PaXini Technology, and the Chinese name is 帕西尼感知科技. The official site identifies the legal headquarters entity as 帕西尼感知科技（深圳）有限公司, reports a 2021 foundation, and locates headquarters in Shenzhen, with additional Shanghai, Tianjin, Suzhou, and Suqian sites ^{[8] [9]}. The functions of the sites differ, so “presence” is not counted as an equivalent factory or customer deployment.

Flagships span PX tactile and force/torque sensors, the DexH and GMH hand lines, TORA robots, data-collection systems, OmniSharing DB, and the OmniVTLA model. This vertical scope is central to PaXini's proposition: the same issuer offers sensing elements, an instrumented hand, data infrastructure, and a policy narrative ^{[7] [8]}. Vertical scope can reduce integration boundaries, but it can also create proprietary coupling. The buyer must inspect raw data access, time synchronization, schema export, model portability, and replacement compatibility.

GMH18 Gen3 provides the clearest product record. The official page discloses five fingers; 18 DoF, comprising 11 active and seven passive; a five-DoF thumb; coreless motors with a multi-linkage mechanism; a self-locking design; built-in ARM processing; and USB, EtherCAT, Ethernet, Modbus, and CAN-FD ^[10]. Power is specified as 24–48 V DC or USB PD 3.1. The 20 kg payload is marked as PaXini laboratory data, so it is not a field payload or a comparison against another hand.

PaXini reports that GMH18 Gen3 contains more than 1,000 ITPU multidimensional tactile sensing units and can perceive 15 tactile types ^[10]. The page describes coverage on fingernails, fingertips, finger pads, and palm. These are useful disclosures, but they remain issuer statements. Independent durability, cross-unit calibration, sensor bandwidth, confusion among the 15 types, task-level ablation, and performance after contamination or skin replacement are still necessary.

The official page calls the sensing “full coverage,” but procurement should translate that phrase into a geometrical map. What percentage of usable surface is instrumented? Are sidewalls and distal edges covered? Does a fingernail sensor measure proximity, force, or both? What dead zones arise at seams? How do wires cross joints? Can individual modules be replaced without changing the learned distribution? A thousand sensing units are only valuable if timestamps, frames, saturation, noise, and health state are accessible.

The disclosed protocols are a strong integration signal, yet a protocol is not an SDK. The captured official evidence does not establish public ROS or ROS 2 packages, a stable Python or C++ API, a semantic tactile message definition, a calibrated URDF, or validated Isaac assets for GMH18. These fields remain unverified. The buyer should request example rates under simultaneous joint, tactile, and camera streaming; clock synchronization; packet-loss behavior; firmware update and rollback; and whether internal force control remains available through every protocol.

Traction is also issuer-qualified. The official company page describes supply relationships and scaled use in manufacturing, automotive production, and healthcare, while giving no matched count of active hands, sites, paid renewals, or independently audited production output ^[8]. Funding amount, revenue, profitability, and unit economics are undisclosed in the captured sources. The official store marks GMH18 as coming soon or sold out with a zero-yuan placeholder, which is not a price ^[11]. Delivered price, sensor replacement cost, software fee, warranty, and volume discount therefore remain undisclosed.

Product maturity is commercially marketed with detailed issuer specifications, not independently established as continuous-duty industrial equipment. Lifecycle questions include linkage backlash, motor heat, joint and skin life, drift, nail and pad replacement, sealing, cleaning chemicals, cable strain, and field recalibration. No complete cycle protocol or independent long-horizon task report was found. The organizer directory verifies PaXini at B203; it does not establish program participation or performance. Media-corpus visibility is likewise unmeasured.

Strengths are integrated sensing coverage, detailed hand architecture, broad industrial protocols, and a sensor-to-data portfolio. Limits are issuer-only payload and traction, unclear public SDK semantics, undisclosed price and lifecycle cost, and no matched evidence that the policy uses all advertised touch types. Confidence is high for the official foundation year, Shenzhen headquarters, and GMH18 page fields; medium for maturity and deployment interpretation; low for finance, independent durability, and policy benefit.

The official GMH18 Gen3 product view below shows the palm, finger pads, fingertips, and nail regions together. It is used only for visual identification and does not verify force accuracy or a full-surface coverage percentage.

8.5 Deep Profile III — Tashan Technology

The Korean rendering is 타산 테크놀로지 or 타산과기, the English brand is Tashan Technology, and the Chinese name is 他山科技. The official legal entity is 北京他山科技有限公司. Its official about page reports foundation in December 2017 in Beijing and describes a team formed from researchers associated with Tsinghua University, the University of Manchester, and other institutions ^[13]. The captured official contact source gives a Beijing address; a distinct “global headquarters” designation is not separately disclosed.

Tashan is not primarily a complete dexterous-hand vendor. Its flagship robot-facing products include TS-F-A/B/C/L fingertip sensors, TS-E-A/B gripper sensors, TS-ES electronic skin, TS-V visuo-tactile training, TS-VT data collection, tactile simulation, and the TS-R service-robot collaboration ^{[12] [14]}. That component position is strategically different from LinkerBot's hand and PaXini's vertical stack. It can fit multiple hands and grippers, but integration responsibility shifts to the hand maker, controller supplier, policy team, and system integrator.

The technical proposition combines a mixed-signal AI tactile chip, an R-SpiNNaker-inspired architecture, tactile algorithms, true three-dimensional force sensing, material recognition, and proximity. The TS-F-A page states normal-force resolution of 0.01 N, tangential-force resolution of 0.25 N, proximity of at least 1 cm, support for three-dimensional force touch, and IIC communication ^[14]. These fields apply to the named configuration; they must not be copied to every TS-F or TS-E variant.

Tashan reports 0.01 N force resolution, 1,000 N safe overload, and a three-million-cycle life test ^[12]. The issuer does not publish in the captured page a complete test protocol, load waveform, contactor geometry, environmental condition, sample count, failure criterion, drift after cycling, or independent reproduction. The three numbers therefore remain qualified product claims, not proof of field life or superiority.

Coverage is explicitly a system-design choice. TS-F-A is an independent fingertip unit; TS-F-B uses two coordinated units; TS-F-C uses multiple coordinated units; TS-F-L can add a tactile biomimetic nail; TS-E-B is an array for a gripper ^[12]. This modularity may let an integrator place sensing where the task needs it. It does not by itself yield whole-hand coverage, because seams, finger sides, palm, and joints can remain blind.

Calibration evidence is incomplete. Resolution is not accuracy, and three-dimensional-force estimation can change with mounting preload, cover material, adhesive, temperature, hysteresis, crosstalk, and aging. The product page verifies IIC for TS-F-A, and the current official product API directly links TactiSim and a Tashan-Isaac-Sim GitHub repository. The supported-model and version matrix, maintenance commitment, simulation fidelity, and portfolio-wide ROS, ROS 2, Python/C++, EtherCAT, and CAN-FD scope still require separate verification. TS-VT promises recording, labeling, cleaning, and export of three-dimensional force, proximity, and contact changes, while the simulation platform references MuJoCo and NVIDIA platforms ^[14]. These statements establish intended integration scope, not cross-simulator fidelity.

Traction is issuer-qualified. The robot product page says the sensing products have enabled more than 200 partners, but it does not define paid customer, shipped unit, active deployment, period, or renewal ^[14]. Revenue, funding total, valuation, production capacity, shipment, gross margin, and recurring software share are undisclosed in the captured primary sources. Price is quote-based or undisclosed; so are replacement modules, calibration service, software licenses, warranty, delivery time, and lifecycle service levels.

Commercial maturity is best described as a marketed tactile component and integration portfolio with some event and partner evidence. Tashan's archived site reports participation in WRC 2024, while the current organizer directory independently verifies its 2026 booth B218 ^[15]. Neither record establishes program participation or performance. Media-corpus visibility is unmeasured under a complete eligibility rule.

Strengths are component modularity, multiple tactile modalities, product-specific force-resolution fields, and training/data tools. Limits are incomplete calibration and cycle-test protocols, unclear cross-hand integration effort, undisclosed pricing and field replacement economics, and limited independent policy ablation. Confidence is high for legal identity, 2017 foundation, Beijing location, product families, and named TS-F-A fields; medium for product maturity; low for finance, lifecycle, broad compatibility, and downstream policy benefit.

The official TS-F-A product view below shows the tactile module itself, separated from any host hand or gripper. It therefore should not be read as evidence that Tashan manufactures a complete hand or that a particular host is mechanically compatible.

8.6 Same-Scale Comparison

The three profiles should not be ranked by a single “technology” score. Their saleable units differ: a hand and data stack, an integrated tactile hand and ecosystem, and tactile components and platforms. A fair comparison asks which responsibilities are included and which remain with the integrator.

This table is deliberately asymmetric. “Not applicable” is different from “undisclosed,” and both differ from “unverified.” Tashan does not need a hand DoF because it sells sensing components. PaXini gives a detailed GMH18 number, but that number should not be applied to every DexH product. LinkerBot covers three actuation families, so the correct comparison is at named-model level.

8.7 Touch Is Modality, Coverage, Placement, and Representation

Tactile sensors use different physical channels. Resistive or capacitive arrays can be thin and direct but may drift and couple to cover construction. Magnetic elastomers can separate replaceable skin from electronics, though nearby ferromagnetic material matters. Optical sensors offer rich contact images but add cameras, illumination, gel wear, and calibration. Force/torque sensors measure aggregate wrench well but may not localize multiple contacts. Proximity can prepare a hand before collision but does not replace contact force.

Coverage, placement, and cross-sensor representation are distinct tactile design variables ^{[16] [17] [18]}. F-TAC Hand reports high-resolution sensing over a large portion of the hand under its research protocol. Tactile Information Distribution studies where sensing contributes across manipulation stages. AnyTouch asks whether representations transfer across different visuo-tactile sensors. None authorizes the conclusion that more taxels always yield a better commercial hand.

Coverage is geometrical: which surfaces can sense. Placement is informational: which locations reveal the variables needed by the task. Representation is computational: whether different sensor outputs become a stable state for learning. A connector task may benefit more from two well-calibrated fingertip shear sensors than from a large palm array. An enveloping grasp may need palm and side coverage. In-hand rotation needs sequential contact across several fingers.

Cross-sensor representation matters because factories replace components. A policy that silently binds to the noise pattern of sensor serial number A can fail after sensor B is installed. AnySkin explicitly treats replaceability and cross-instance transfer as a design target ^[19]. ReSkin separated a low-cost replaceable magnetic skin from persistent electronics and reported a cycle and recalibration process in a research setting ^[20]. These papers do not prove vendor lifecycle; they show what an acceptance test should measure.

8.8 Calibration, Drift, Durability, and Lifecycle

Calibration is not one event. Factory calibration maps raw signal to a physical estimate. Installation adds mounting preload and coordinate frames. Operational calibration compensates temperature, zero drift, or cover change. Replacement calibration asks whether a new sensor, gel, skin, finger, or hand can enter service without collecting a new policy dataset.

A practical calibration record includes serial number, firmware, calibration file hash, fixture, applied loads, temperature, mounting torque, residual error, saturation, hysteresis, crosstalk, and date. During a task, the system should log zeroing and health. If the controller automatically normalizes values, the raw stream and normalization version should remain traceable.

Durability should match the failure physics. Tendons fatigue and stretch; gears and linkages wear and gain backlash; motors and drivers heat; optical gels scratch and cloud; elastomers tear; adhesives creep; exposed wiring snags; seals admit dust or fluid. One cycle count under one normal load says little about combined shear, contamination, temperature, impacts, and edge contacts. Cycle count also needs an endpoint: electrical survival is weaker than calibration retention and successful task execution.

Lifecycle cost can be written as:

\[

C_{task}=\frac{C_{hand}+C_{integration}+C_{spares}+C_{service}+C_{downtime}+C_{human}}{N_{accepted\ tasks}}

\]

The denominator is accepted tasks, not commanded closures. Integration includes mounts, cables, drivers, safety validation, policy adaptation, and logging. Spares include skins, fingertips, tendons, motors, and whole-hand swaps. Human cost includes cleaning, recalibration, reset, annotation, and vendor escalation. A low list price can lose if replacement requires hours and policy retraining.

Price evidence needs five layers: public list price, actual quote, delivered system price, recurring licenses and consumables, and downtime risk. Research baselines help reveal what is possible: LEAP Hand reported a roughly $2,000 build, DIGIT a low bill of materials at scale, and ReSkin a very low-cost replaceable skin ^{[21] [22] [20]}. Prototype BOM is not a commercial delivered price with quality control, warranty, support, and inventory.

8.9 SDK and Data Contracts

An “open” hand must expose more than a position command. A useful developer contract covers joint targets and states, current or torque, velocity, tactile raw and calibrated signals, timestamp and clock, temperature, fault, calibration identity, firmware, and emergency behavior. It should say which loops run on-device and which may run over ROS, Ethernet, CAN, or EtherCAT.

ROS package availability is only the first rung. Reproducible integration needs a URDF with collision and inertial properties, transmissions, limits, message semantics, example bags, controller rates, tests, supported distributions, releases, and issue handling. Isaac or MuJoCo support needs matching geometry, joint dynamics, actuator limits, tactile approximation, and a versioned sim-to-real procedure. TACTO demonstrates fast simulation for optical tactile sensors but also approximates gel mechanics and shading ^[23]. A simulator logo does not establish tactile fidelity.

Data contracts are equally important. Vision, touch, joint, wrist-force, and action streams need a shared clock and execution ID. Dropped packets and interpolation should be explicit. Units, coordinate frames, saturation, calibration, and contact masks need versioned schemas. Training data must distinguish autonomous action, operator correction, safety override, and reset. Otherwise the policy can learn recovery artifacts or misread operator force as autonomous competence.

Teleoperation interfaces often capture pose better than force. DexPilot and AnyTeleop demonstrated general vision-based retargeting, but no haptic feedback and occlusion remain limitations ^{[24] [25]}. DexUMI reduces specialized capture hardware by using the human hand as an interface, while contact is inferred rather than directly measured on the human side ^[26]. A hand-data platform should state what touch exists on operator and robot sides and how correspondence is established.

8.10 From Tactile Stream to Policy Benefit

The availability of tactile hardware does not prove that a policy uses touch effectively ^{[27] [28] [29]}. A model can receive tactile tokens while relying almost entirely on vision or proprioception. Demonstrating benefit requires an ablation under the same objects, seeds, controller, resets, and intervention rule: vision only, touch only where appropriate, vision plus touch, and corrupted or delayed touch.

Policy benefit should appear in task outcomes. Useful measures include first-attempt grasp, slip recovery before drop, regrasp count, insertion success, false seating, peak force, damage, final pose, completion time, and human intervention. Report distributions and hard cases, not only a highlight reel. A tactile policy that succeeds more often but doubles cycle time may or may not create value.

TLA connects tactile input, language, and action for a bounded fingertip peg-in-hole setting and reports a dataset of 24,000 tactile-action-instruction pairs and more than 85% success on specified unseen clearances and shapes ^[28]. Its authors also identify limited treatment of temporal touch. The result motivates careful contact-state learning; it does not prove general factory insertion or any vendor hand.

Tactile-VLA and force-informed learning explore ways to inject touch or human contact information into policy learning ^{[27] [29]}. The transfer lesson is architectural, not a borrowed headline. A buyer should demand a vendor-specific ablation on the delivered sensor, hand, firmware, task, and duty cycle. Long-horizon dexterity also exposes reset and infrastructure costs: the Rubik's Cube work showed extraordinary capability alongside extensive task-specific training and hardware intervention ^[30].

8.11 Manufacturing Walkthrough: Connector Seating

Consider a cell that picks a flexible cable connector, aligns it to a board socket, inserts it, verifies seating, and places the assembly in an output tray. Vision can estimate pose before contact, but the final millimeters contain occlusion, compliance, manufacturing tolerance, and friction. This is where a hand and touch either reduce uncertainty or add another failure mode.

Step 1: Define the accepted outcome. Success is the correct connector fully seated with latch engagement, no bent pin, no housing damage, cable within routing limits, and traceability to board and connector lot. “The robot finished the trajectory” is not success.

Step 2: Establish the simplest baseline. Test a parallel gripper with wrist force/torque and compliance before a multifinger hand. Record success, peak force, damage, cycle time, reset, and human minutes. The dexterous system must beat a real baseline on a needed dimension.

Step 3: Select task-fit contacts. Decide whether two opposing fingertips are enough or whether a third finger stabilizes cable orientation. Map the expected contact path to sensor placement. Do not purchase palm coverage if only fingertip shear changes the decision.

Step 4: Qualify mechanics. Measure grasp repeatability, jaw or fingertip pose, backlash, cable load, thermal behavior, and arm payload. Run the intended insertion angle and speed. Verify collision geometry and pinch hazards with tooling and people.

Step 5: Calibrate the sensing chain. Calibrate raw touch, wrist force, joint state, and camera into named frames. Test zero drift, preload, temperature, and mounting variation. Save calibration IDs with every execution.

Step 6: Synchronize and observe. Log camera, tactile field, hand and arm state, commanded action, controller mode, force limit, fault, and safety event under a common clock. Inject packet delay and sensor dropout to verify bounded fallback.

Step 7: Design the contact state machine. Use coarse visual approach, guarded contact search, alignment correction, insertion, seating verification, and withdrawal or escalation. Tactile input should change a named transition or command. If deleting tactile data does not change action, the sensor is not providing operational intelligence.

Step 8: Define recovery. Limit search distance, force, time, and attempts. On misalignment, withdraw along a safe vector and reobserve. After a second failure, quarantine the part or request a person. A reset must not conceal bent pins by silently loading a fresh sample.

Step 9: Run ablations and perturbations. Compare vision-only and vision-touch policies under socket pose error, connector lot, cable stiffness, lighting, contamination, and sensor replacement. Report confidence intervals, false-seat rate, damage, and interventions.

Step 10: Stress lifecycle. Cycle the cell across shifts. Replace a fingertip or hand with a spare and measure physical swap, calibration, software recognition, and policy recovery time. Inspect drift before failure, not only catastrophic breakage.

Step 11: Price the accepted task. Include hand, sensors, spare modules, mounts, cables, compute, license, integration, validation, cleaning, calibration, technician time, scrap, and downtime. Compare cost per accepted connector and avoided damage.

Step 12: Gate production. Promotion requires a fixed success and damage threshold across lots and shifts, an intervention ceiling, documented safety response, reproducible replacement, version rollback, and named ownership between production, quality, safety, IT, integrator, and supplier.

Manufacturing Cell Checkpoint

The task schema should include assembly ID, connector and board lot, target socket, grasp mode, approach pose, insertion direction, expected seating signature, maximum normal and shear force, allowed search volume, retry count, and recovery owner.

The unified log should contain hand and sensor serials, firmware, SDK, model and calibration versions; arm and hand commands and states; raw and calibrated touch; wrist force; video and depth; contact-state transitions; safety stops; resets; replacements; and final inspection under one execution ID.

KPIs should include first-pass yield, intervention-free yield, false seat, bent pin and housing damage, peak force, insertion time, regrasp and search count, reset and quarantine, uninterrupted runtime, drift, replacement time, MTTR, scrap, and total cost per accepted assembly.

Ownership must be explicit. Production owns takt and material flow. Quality owns acceptance and damage. Safety owns force, speed, guarding, and stop authority. IT owns networks, clocks, versions, and data retention. The integrator owns cell logic and traceability. The hand or sensor supplier owns contracted specifications, spares, firmware, calibration support, and response time.

8.12 Evidence Ladder: Demo, Benchmark, Pilot, Deployment

A demo answers “can this configuration do the task at least once?” It can be useful, but object selection, prior attempts, edits, teleoperation, and resets may be hidden. A research benchmark adds a protocol, denominator, and baselines. It is stronger for mechanism evidence, but usually uses a bounded task and prototype hardware. A customer pilot adds real operators, variance, and existing systems. A deployment adds sustained duty, maintenance, costs, and responsibility.

Issuer evidence is valuable for identity, interfaces, product availability, and disclosed specifications. It is weak for superiority unless methods and denominators match. Independent papers are valuable for mechanisms and ablations. They are not automatic evidence for a vendor unless the tested commercial configuration is named. The responsible synthesis keeps both without blending their scopes.

8.13 Limitations and Open Questions

First, the public evidence is asymmetric. PaXini publishes detailed GMH18 fields. LinkerBot publishes broader mechanism and developer coverage. Tashan publishes sensor-level fields. Missing values are recorded as undisclosed or unverified, not zero and not inferiority.

Second, company specifications are primarily issuer evidence. Payload, resolution, coverage language, cycle life, partner counts, and deployment descriptions are not matched across definitions. This chapter does not calculate a composite score from them.

Third, price and lifecycle evidence are sparse. Quotes may differ by hand side, sensor option, quantity, controller, software, and support. Public research BOMs exclude commercial quality, warranty, and integration. No matched field study reports cost per accepted task for the three profiles.

Fourth, policy evidence is indirect. The research papers show why touch, placement, representations, temporal signals, and force-aware learning can matter. They do not establish that a named company policy obtains the same benefit. Vendor-specific ablation, replacement, and long-duty tests remain open.

Fifth, safety needs system assessment. A compliant or tactile hand is not automatically a safe collaborative system. Pinch points, stored tendon energy, sharp tools, payload, arm speed, sensor faults, and unexpected policy actions need risk assessment, independent stops, and validated limits.

Sixth, media-corpus visibility is not measured, and WRC status is company- and year-specific. The organizer directory verifies LinkerBot B205, PaXini B203, and Tashan B218 for 2026; Tashan's official archive separately establishes a 2024 appearance. These records verify exhibitor booths, not program participation or product performance.

Open questions are operational. Can a tactile representation survive replacement across batches without retraining? Which modalities retain calibration under oil, dust, cleaning, and temperature? Can a high-DoF hand expose enough low-level authority while preserving safe onboard control? Will hand suppliers earn recurring revenue from software and service, or will consumables and field repairs dominate? The answers require cohort and lifecycle data, not a launch video.

Relation to Prior Surveys

This chapter carries forward the prior chapters' distinction between platform capability and deployed service. Teleoperation research explains how dexterous demonstrations are collected, while tactile research explains how hidden contact can alter action. Neither resolves procurement. The contribution here is the evidence bridge: named product, interface, calibration, replacement, task ablation, operating log, and cost must connect before “contact intelligence” becomes an industrial claim.

What to Learn Next

Hands and tactile sensors produce observations and actions at the physical boundary. Chapter 9 moves upstream to the data engines, teleoperation systems, foundation models, and VLA stacks that organize those observations into reusable policies. The central question changes from “does the hand feel?” to “how is contact data collected, aligned, governed, trained, evaluated, and improved?”

Carry one discipline into Chapter 9: data volume is no substitute for denominators and lineage. A million tactile frames may contain only a few task transitions. A teleoperated episode may hide multiple corrections. A model checkpoint may depend on one hand, sensor batch, calibration, and reset rule. The next chapter follows that chain from embodiment-specific contact to generalizable robot intelligence.

References

1. LinkerBot (2026a). LinkerBot Official Website. Official headquarters and portfolio source.
2. LinkerBot (2026b). About LinkerBot. Official company and technology description.
3. LinkerBot (2026c). LinkerBot Exhibitor Profile, 2026 World Robot Conference. Organizer source for legal name, booth, and named products.
4. LinkerBot (2026d). Linker Hand Developer Center Overview. Official manuals, SDK, CAN, demo, and model-resource description.
5. LinkerBot (2026e). Linker Hand L20 Product Manual. Official manual linking Python, ROS, and ROS 2 SDKs.
6. LinkerBot (2025). Open TeleDex. arXiv:2510.14771.
7. PaXini (2026a). PaXini Dexterity Ecosystem. Official ecosystem source.
8. PaXini (2026b). About PaXini Technology. Official foundation, technology, and issuer deployment source.
9. PaXini (2026c). Contact PaXini. Official legal entity, headquarters, and site source.
10. PaXini (2026d). GMH18 Gen3 Multidimensional Tactile Dexterous Hand. Official product specification; laboratory and issuer claims retained.
11. PaXini (2026e). GMH18 Official Store Listing. Official availability page; zero-yuan placeholder is not treated as price.
12. Tashan Technology (2026a). Tashan Tactile Product Portal. Official product-family and issuer metric source.
13. Tashan Technology (2026b). About Tashan Technology. Official foundation and company source.
14. Tashan Technology (2026c). Tashan Robot and Tactile Products. Official TS-F, TS-E, TS-ES, TS-V, and TS-VT source.
15. Tashan Technology (2025). Tashan at the 2024 World Robot Conference. Issuer event archive.
16. F-TAC Hand Authors (2024). F-TAC Hand: A Tactile-Sensing Dexterous Robot Hand. arXiv:2412.14482. Terry's reading note #39.
17. Tactile Distribution Authors (2025). Tactile Information Distribution for Robot Manipulation. arXiv:2509.14984. Terry's reading note #41.
18. AnyTouch Authors (2025). AnyTouch: Learning Unified Static-Dynamic Representation across Multiple Visuo-Tactile Sensors. arXiv:2502.12191.
19. Bhirangi, R. et al. (2024). AnySkin: Plug-and-play Skin Sensing for Robotic Touch. arXiv:2409.08276.
20. Bhirangi, R. et al. (2021). ReSkin: Versatile, Replaceable, Lasting Tactile Skins. arXiv:2111.00071.
21. Shaw, K., Agarwal, A., and Pathak, D. (2023). LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning. RSS 2023.
22. Lambeta, M. et al. (2020). DIGIT: A Novel Design for a Low-Cost Compact High-Resolution Tactile Sensor with Application to In-Hand Manipulation. IEEE RA-L.
23. Wang, S. et al. (2022). TACTO: A Fast, Flexible, and Open-Source Simulator for High-Resolution Vision-Based Tactile Sensors. IEEE RA-L.
24. Handa, A. et al. (2020). DexPilot: Vision-Based Teleoperation of Dexterous Robotic Hand-Arm System. ICRA 2020.
25. Qin, Y. et al. (2023). AnyTeleop: A General Vision-Based Dexterous Robot Arm-Hand Teleoperation System. RSS 2023.
26. Xu, M. et al. (2025). DexUMI: Using Human Hand as the Universal Manipulation Interface for Dexterous Manipulation. CoRL 2025. Terry's reading note #8.
27. Huang, J. et al. (2025). Tactile-VLA: Unlocking Vision-Language-Action Model's Physical Knowledge for Tactile Generalization. arXiv:2507.09160.
28. Hao, P. et al. (2025). TLA: Tactile-Language-Action Model for Contact-Rich Manipulation. arXiv:2503.08548.
29. Adeniji, A. et al. (2025). Feel the Force: Contact-Driven Learning from Humans. arXiv:2506.01944.
30. Akkaya, I. et al. (2019). Solving Rubik's Cube with a Robot Hand. arXiv:1910.07113.
31. Antonio Bicchi (2000). Hands for Dexterous Manipulation and Robust Grasping: A Difficult Road Toward Simplicity. DOI:10.1109/70.897777.
32. Antonio Bicchi et al. (2000). Robotic Grasping and Contact: A Review. DOI:10.1109/robot.2000.844081.
33. Mark R. Cutkosky (1989). The Grasping Hand. DOI:10.1007/978-1-4613-8974-3.
34. David J. Montana (1988). Kinematics of Contact and Grasp. DOI:10.1177/027836498800700302.
35. J. Kenneth Salisbury et al. (1982). The Mechanics of Fine Manipulation by Finger Tips. DOI:10.1177/027836498200100201.