Cycle life, not fingers

Sharpa's Wave product page publishes three endurance figures, and they are three different quantities. In the reliability block: "Press Test — Validated through 2,500,000 press cycles". In the hardware table: "Rated Lifetime — Over 1,000,000 cycles". In the tactile block: "Tactile Sensor Lifetime — Over 100,000 press cycles". All three are on sharpa.com/pages/wave, read on 11 September 2026 and again on 12 September.

A rating, a test result, and the life of one component. The rating sits two and a half times below the test result and ten times above the sensor's. Nothing on the page merges them, and nothing says at what force, against what object, at what duty, or what ends a life. The figure that travels — into the coverage, and into our own research files — is the middle one.

We have been wrong about that page twice, in opposite directions, in two days. An earlier version of our Body page credited Sharpa with 2,500,000 cycles as the hand's cycle life. On 11 September 2026 we corrected the page to say Sharpa publishes no cycle life at all. That correction was wrong too, and it was live in two locales. Which page the earlier pass read we cannot establish from our own files. The page was corrected again the same day, and carries all three figures, separately, as of 12 September 2026.

One thing we can establish is that the table is recent. The Internet Archive's last two captures before we read it, both timestamped 5 September 2026, carry a complete specification sheet — headed "Wave Specification Sheet", sectioned Mechanical, Sensing, Durability and Software — whose Durability section lists a press-cycle validation and a friction test and no rated lifetime. The 2 June 2026 capture carries the reliability block and no specification sheet at all. What was live on 11 and 12 September is a different table: differently headed, differently sectioned, carrying a rated-lifetime row and a tactile-sensor-lifetime row the 5 September sheet does not have. That is a capture's presence rather than a capture's absence, which is the stronger direction, and it is still two crawls [single source — Internet Archive, captures of 2 June and 5 September 2026].

One cycle, three definitions

We had it that exactly one vendor in this comparison says what a cycle is. That was our finding on 11 September, and it was false by 12 September, because nobody had asked the question of the right document.

Schunk sells a small parallel gripper, the EGK 25. Its product page carries an ordinary warranty row, and inside the row is a parenthesis:

"24 months or 5 million cycles (one cycle consists of a complete gripping process: "Open gripper" and "Close gripper")"

No object. No force. No success criterion. A Schunk cycle is a movement of the mechanism, and a gripper opening and closing on empty air in a test rig accumulates them at full rate.

Robotiq defines a cycle too. Not on the product page — in the instruction manual, a 145-page PDF on Robotiq's own asset CDN, fetched 12 September 2026. It defines a cycle twice, and the two definitions do not count the same event.

In chapter 7, Maintenance, page 129, in a box directly beneath the maintenance-interval table:

"A cycle is defined as a go to requested position command that results in grasp force being applied (picking an object while opening or closing or closing the fingers on themselves)."

In section 10, Warranty and Patent, footnote 1:

"Cycle count: One (1) cycle is defined as an object picking attempt, successful or not (open or closing onto an object, or closing on itself). It is calculated in the internal memory of the 2-Finger Adaptive Gripper and can been seen with the Robotiq User Interface."

Under the first, a grasp that closes on nothing and applies no force is not a cycle. Under the second it is: the attempt counts, "successful or not". Same product, same document, seventy pages apart. Not a transcription error in one file: Robotiq's Hand-E manual carries both sentences too, one word changed, "grip force" for "grasp force".

Three published definitions. Two vendors. No two count the same event, and one vendor disagrees with itself. A gripper that misses a part is running down its warranty under Robotiq's section 10 and standing still under Robotiq's chapter 7. A gripper cycling on air in a burn-in rig accumulates Schunk cycles and no Robotiq cycles of either kind. So the number is not comparable even between the two vendors who said what they meant. That is a harder problem than silence. Silence you can at least see.

One of the two counts itself. The Robotiq footnote says the cycle count "is calculated in the internal memory" of the gripper and can be read out, with the Robotiq User Interface or on a Universal Robots teach pendant. There is a counter, it sits in the tool, and vendor and customer read the same register when they argue about a warranty claim. Schunk's five million is not counted by the gripper. It is a term. Neither product page mentions any of this.

A second quantity hides in the same cell. A rated life is an engineering claim about when a thing wears out; a warranty term is a commercial claim about when a company stops paying. We counted seven cycle-denominated endurance rows across seven products on their vendors' own pages. Five are labelled Warranty, one Durability, one Rated Lifetime. No page says these are three different kinds of claim. Our own first pass at this article merged them into a single count, which is how we know the merge is easy to make.

The documents disagree about the number too. Robotiq's Hand-E product page publishes "Warranty 5 000 000 cycles", read 12 September 2026. The Hand-E instruction manual's warranty conditions say the warranty applies "Until a 2 000 000 cycle count1 has been reached." Same vendor, same product, a factor of two and a half — and the smaller number sits in the only document that defines the unit. The likely explanation is dull: the warranty was extended after 2019 and the manual was never revised [inference]. We are not saying the page is wrong. We are saying the definition of the unit and the current value of the quantity live in two documents that disagree, which is the same failure as having no definition, reached by a different road.

Both Robotiq manuals are 2019 documents. The Hand-E file's latest revision entry is 13 March 2019, the 2F file's 7 November 2018, and Robotiq's own CDN served both when we fetched them on 12 September 2026. They are what a buyer is pointed at. They are not new.

The last thing in the 2F manual is the one a specifier needs most. Table 7-1's four columns are headed "Daily", "Weekly", "Semiannually (or 1 M cycles)" and "Annually (or 2 M cycles)", and the warranty conditions state the duty that makes the equivalence true: "Usage under normal one-shift operation (40h a week)", with "Or until a 2 000 000 cycle count1 has been reached" as the alternative limit. A year of one-shift work is two million cycles. Divide it out: 40 hours a week for 52 weeks is 2,080 hours, so roughly 960 cycles an hour, one every 3.7 seconds [inference — our arithmetic, not Robotiq's]. A duty is what converts a cycle count into a service life. This one is published in exactly one place, and it is not the place anyone looks.

Who publishes what, and where

Dexterous and humanoid hands.

Industrial end-of-arm tooling.

Read the last two columns together; the gap is the finding. The cell values differ and the difference is the point. In the middle columns, not published means we read that vendor's own page and it carries nothing. In the eighth column, no means a cycle-denominated figure is published there with no definition of a cycle beside it; — means no cycle-denominated figure at all, so there is nothing to define; and the single yes is Schunk's. Across the comparison, three vendors publish a cycle-denominated figure across seven products, five of the seven Robotiq's. Our research file first said five vendors; this article's first draft corrected that to six products and four; both were wrong, and the fourth Robotiq gripper turned up when we opened the same page again. This article's failure mode, caught twice inside the article. Every cell comes from that vendor's own page, datasheet, manual or specification PDF, read on the date in the first column; Schunk's PGN-plus-P datasheet reaches us through a distributor-hosted copy. Sharpa's table compares two Wave variants side by side, and its column headings are images rather than text, so a paired value in that row — 20 N / 12 N, 240×240 / 60×60 — is the two columns and not a range; the rated lifetime and the tactile-sensor lifetime are single values spanning both. not checked means we did not open those documents, not that they do not exist. There is no Tidewell row: no hand of ours exists and we have measured nothing. No price appears in this figure.

Note the fifth row of the lower block. Robotiq's adaptive-gripper page carries five specification blocks; four publish a cycle warranty and an ingress rating. The flagship 3-Finger gripper, same page, same navigation, publishes neither. None of this accuses anyone. Every vendor named may hold excellent internal endurance data, and we would not know. Absence of a published figure is absence of a published figure, and that is all this article claims about any of them.

What happens to a number with no protocol

On 2 June 2026 CNX Software published a specification list for the Wave hand containing the line "Friction test of fingertips – Verified over 4,000 km" [secondary — CNX Software, 2 June 2026]. Sharpa's own page says "Verified over 4,000m travel distance". Four thousand metres, not four thousand kilometres: a factor of one thousand.

The obvious defence is that the vendor changed that line in the three months between. It did not. The Internet Archive holds a capture of sharpa.com/pages/wave timestamped 2 June 2026, the day the article ran, carrying the friction line at 4,000 m. The unit changed in one hop, page to coverage. What the author worked from we cannot establish; we say only what the vendor's page said that day. Any reader can run that check in ten minutes. We ran it late — the kilometre figure had been sitting in our own research corpus for a week.

The same coverage gets the press-test and shock figures right and loses one condition. "1,000+ hours of continuous operation in a waving finger type of test" attaches the page's 1,000-hour figure to the wrong test; the page files those hours under "Extreme Temperature Cycling". The number survives the hop; the condition does not. A protocol has to travel attached to its figure, not one paragraph away from it.

One cycle-life figure attributed to a humanoid vendor has no vendor page under it. UBTech's Walker S2 hand is quoted at about 80,000 cycles on at least six aggregator and reseller pages we located on 11 September 2026 [secondary]. That day, and again on 12 September, we fetched ubtrobot.com's own English Walker S2 page and searched it for "80,000", for "80000", for "cycle", for 寿命 and for 灵巧手. None is there. Half a dozen sightings is not a citation count, and we do not claim this is the figure the field quotes most — only that every instance we found traces to aggregation and the primary does not carry it. That is a statement about where a number lives, not about how long the hand lasts, which we have no way to know.

Sharpa's own small print is the most candid line in the comparison: "Specifications may vary between products", and "Products are continuously iterated and improved; minor differences may occur." That is honest. It is also why a rated life with no protocol is hard to hold anyone to.

The page is not the vendor

Every negative in this article was, at some point, an absence read off a product page. Three of them turned out to be answered somewhere else on the same vendor's own site.

Start with the one that holds. Ingress protection shows the divide plainly, because the industrial side treats it as a spec row. Robotiq publishes IP67 on the Hand-E and the Hand-E C10, IP40 on the 2F-85 and 2F-140. OnRobot's 2FG7 offers "IP67 rating for harsh environments and ISO Class 5 certification for cleanroom use". Schunk publishes two ingress classes for two parts of one gripper — IP67 electronics, IP20 guide and base jaws — plus an ISO 14644-1 class 5 cleanroom rating. Six machines on the other side published none when we fetched them on 11 September 2026: Galbot G1, Pudu D9, Hello Robot Stretch 4, Primech Hytron, Keenon KLEENBOT and the Sharpa Wave. Somebody in the same industry, selling into the same factories, publishes all of it. That turns an absence into a choice.

Now the one that broke. What we read on 11 September 2026: Sharpa's Wave page, Shadow's Dexterous Hand Series page, Robotiq's gripper and tactile-fingertip pages, OnRobot's 2FG7 page, Schunk's EGK 25 page with its technical-data table, and Schunk's PGN-plus-P datasheet. None publishes a preventive-maintenance interval or a service schedule. We added a hypothetical caveat: an operating manual we had not opened is exactly where an industrial vendor would put one. It is not hypothetical. Robotiq's 2F manual publishes four intervals, in four sections, each in its own cell: gripper cleaning "Weekly or daily in dirty operating conditions"; periodic inspection "Monthly"; fingertip replacement at "1 Million cycles or when wear is visible"; and overhaul at "2 Million cycles or at warranty expiration", with the prose underneath reading "Gripper overhaul is necessary when the Gripper reaches 2 Million cycles or when warranty expires." The Hand-E manual carries a three-interval version of the same schedule. A product page that publishes no maintenance interval is not a vendor that publishes none.

The broad form of that sentence — that no humanoid vendor publishes a preventive-maintenance schedule — stood on our own Body page until 12 September 2026, when it was replaced by the bounded version with the operating-manual caveat named. We asserted a negative wider than our search and then stopped searching, which is the same error as reading a figure without its protocol.

And then the one that cost us something. The strongest negative we had was that nobody on either side of the divide publishes calibration stability, drift compensation or a self-diagnosis path. On the product pages that is still true: we re-ran it on 12 September 2026 across the Sharpa Wave page, Robotiq's adaptive-gripper and tactile-fingertip pages, OnRobot's 2FG7, Shadow's series page and Schunk's EGK 25 page, plus the PGN-plus-P datasheet, and the strings "calibrat", "drift" and "self-diagnos" return nothing on any of them.

Then we opened the manual, which is what this article had just taught us to do. Robotiq's TSF-85 user guide — version 1.0.0, © 2026, and unlike the gripper manuals a current document — publishes a per-taxel calibration routine under the heading "Periodic Taxel Sensitivity Calibration", and a per-cycle drift procedure under "Taxel Signal Compensation During a Robot Cycle", whose first step is to "store the current "no load" signal as a baseline for the current cycle. This compensates for short-term drift and hysteresis." Two of the three things we said nobody publishes are published, as procedures, with worked steps.

What survives is the third, for an exact reason. "Periodically calibrate" carries no interval. "Perform regular inspections", in the maintenance section, carries none either. And the guide states, twice, that "Tactile Sensor Fingertip sensitivity will decrease over time at a rate dependent on the application", which is a declared degradation with no rate attached — then puts it out of scope: "This sensitivity decrease is considered normal wear (not covered by warranty)." The only diagnostic in the document is a table of LED states that reports whether a fingertip is connected and powered, not whether it is still telling the truth.

Hold that against the product page for the same part, which publishes "Durability — > 2 000 000 cycles". The user guide does not contain the string "2 000 000" anywhere, warrants the part for "one year from the date of reception", and excludes "all consumable parts, such as fingertips and their normal wear." A two-million-cycle durability figure on the page; a one-year warranty in the manual that explicitly does not cover the degradation the figure is about. Both documents are the vendor's. Neither hides anything. They answer different questions for different readers, and only one is the document anyone reads before buying.

Shadow's technical specifications tell the same story from the other end. Its two hand specification PDFs, thirteen and twelve pages, hosted on shadowrobot.com, contain no occurrence of "cycle", "maintenance" or "lifetime". What they contain instead is an architecture statement: joint data is "provided to the communication bus in raw form and calibrated at the host", the tendon force sensors are "zeroed but not calibrated", the tactile data "uncalibrated". More honest than a number. Also not a stability figure, and there is no stability figure anywhere in this comparison.

So the line where buying stops and building starts has moved, in a direction that costs us an argument we liked. The procedures are published. The element is a part you buy, sold as a drop-in replacement fingertip for the vendor's own grippers, with its taxel count, frequency and force range in the table above. What nobody publishes is how long any of it holds — the interval, the rate, the drift over months, the point at which a fingertip should be believed less. A narrower gap than we had been claiming, and still the right one to build in.

One comparison we will not make. Sharpa's tactile sensor lifetime reads "Over 100,000 press cycles"; Robotiq's tactile fingertip durability reads "> 2 000 000 cycles". Twenty times, on the face of it. Neither publishes the conditions, a press cycle on a capacitive taxel array and a press cycle on whatever Sharpa tested are not established to be the same event, and by now we know two documents from one vendor need not agree on the unit either. It appears here only as an example of a comparison nobody can make.

What we owe, on our own axis

One artefact in this field published what ended its test. The open-source ORCA hand — Christoph and ten colleagues, arXiv 2504.04259, v2 dated 17 September 2025, accepted to IROS 2025 — is a 17-DoF tendon-driven hand with integrated tactile sensors, and its abstract says this:

"Furthermore, we demonstrate its durability, withstanding more than 10,000 continuous operation cycles - equivalent to approximately 20 hours - without hardware failure, the only constraint being the duration of the experiment itself."

The final clause is why the sentence has to travel whole. It turns a ceiling into a floor: the hand did not stop, the experiment did. ORCA is not the only checkable endurance figure here, and after this week not the only published protocol either — Robotiq's manuals define their unit and give their intervals, and one of them states its duty. What ORCA has that no vendor document has is the limit said out loud, in the same sentence as the number. Small thing to publish. Nobody selling a gripper does it. Its diagnosis is our own buy-versus-build argument in someone else's words, framed as a response rather than a complaint — "What if the research community could get started with reliable dexterous hands within a day?" — before it states that "the primary bottleneck in dexterous manipulation lies not only in software but arguably even more in hardware. Robotic hands that approach human capabilities are often prohibitively expensive, bulky, or require enterprise-level maintenance, limiting their accessibility for broader research and practical applications."

Our own rule is published and short: buy what two suppliers can make and whose cost curve falls faster than we could bend it; build what fails in a hospital or a port and nobody is fixing. Audited against this disclosure record it leaves hands, fingertip calibration and the hand bus on the build side — and the audit narrowed the third of those rather than confirming it. We are not building a hand with more fingers. The Body page publishes what we are building instead: actuators moved back into the forearm so the fingers stay cool and slim, sealed finger modules replaceable on site without recalibrating the arm, a bought sensing element with the calibration and drift layer above it built here, and a rated life published rather than a peak figure demonstrated. That axis follows from where we intend to put a machine, a ward and a wharf, not from a universal truth about hands.

And our own rated life is a promissory note. No hand of ours exists. No test has finished. We have measured nothing on a machine, and nothing here should be read as a claim that we have. When the number arrives it arrives with the cycle definition, the duty it assumes, the load and the criterion that ends the test, in the same place as the number — or it does not arrive. Whether a published performance figure has the statistical power to mean anything is a separate argument, made separately; this one is about what a specification discloses and where.

There is a loop behind all of it, and it needs no statistic. A hand that fails in month four returns no interaction data from month five onward, and interaction data is the input to better hands. Tactile is a durability problem before it is a learning problem [inference].

The practical version is two questions, and a buyer can put both in one email. What is one cycle, in your definition, in writing. Which document is that definition in, and does the number on your product page come from the same document? We have watched a vendor answer the first question twice, differently, in one PDF. We have watched ourselves get a durability claim about a named company wrong in both directions inside forty-eight hours. The discipline is not to distrust the numbers. It is to refuse to compare two of them until both definitions are in front of you — and to say which pages you read before asserting that something is not there.