--- title: "A Pair of Hands, Building the Next" date: 2026-11-05 category: Humanoids & Robotics site: NeuroAI canonical: https://neuroai.site/a/na-hand-line-15-1105b language: en --- # A Pair of Hands, Building the Next > What a factory has most of is repetition; what it lacks most is a hand that can handle surprises. Driving the same screw ten thousand times proves little; getting it in when the hole is off, the What a factory has most of is repetition; what it lacks most is a hand that can handle surprises. Driving the same screw ten thousand times proves little; getting it in when the hole is off, the light changes or the tool gets knocked askew is what really covers a shift. This article skips the spec sheet and looks only at how one hand runs a production line. According to media reports, at a test station of a joint manufacturing base in Shaoxing, one dexterous hand was running a whole-hand performance test on another hand awaiting shipment. The demo was carried by a live cross-city link into an industry sub-forum in Shanghai. The scene sounds simple, but taken apart it is a long chain of actions. ## One Hand Tests Another First, what this station does. Before any dexterous hand ships, it must pass a whole-hand performance test to check that joints, force control and sensing are all sound. According to media reports, that step is now handed to another dexterous hand, which carries loading through to return on its own. Handing inspection to a hand carries its own meaning: being able to build a hand and being able to steadily build a consistent batch of hands are two different things. The former shows up in a prototype, the latter on a line, and the line's most honest examiner is precisely the next hand due to ship. According to media reports, the company's executive called this scene in a speech "one pair of hands assembling the next pair on the line," describing it as the picture he has in mind. This article borrows the phrase only to set the theme and does not expand on his view of the industry's stage. One thing needs stating: this is a demonstration at a single test station, not a whole factory already running without people. The report says this sequence is completed autonomously by the machine; as for other loading steps or how other stations connect, public information gives no full picture, and it should not be filled in on the reader's behalf. ## Seventy-Two Steps in a Row According to media reports, the full sequence is loading, scanning, clamping, picking up a screw gun and tightening, starting the test equipment, removing screws, and unloading and returning — 72 operational steps in all, completed autonomously with no human intervention. That string of verbs is worth reading slowly. It is not one grasp and done, but a whole chain of picking up a tool, aligning to a hole, tightening, releasing, setting down and returning. If any step drifts, everything after it skews, so for a hand the hard part is not how pretty one motion looks but getting seventy-odd steps in a row right. According to media reports, the company's director of modeling and algorithms told the media the team had tried a conventional gripper approach, but recoil from the work easily made the tool slip; fixing the tool to the arm's end, meanwhile, could not also handle varied steps such as grasping irregular parts or pressing dashboard buttons. In his view, the value of a dexterous hand is covering many kinds of steps with one general end effector, saving the cost of retooling a dedicated machine for each step. That is the company's own trade-off as described to the media, not a verdict on every technical route. ## Deliberately Throwing the Station Off The more interesting part comes after. According to media reports, a stress test was run on site: staff deliberately pushed the arm so the gun shifted, adjusted the fixture base so the screw holes misaligned, and changed the lighting, to see whether the sequence would fall apart. The hardest part is driving the screw. According to media reports, the reusable captive screws at the station carry millimeter-scale play, and a slight shift in the hand's grip can put the gun tip a few millimeters out of alignment — a tiny error that stops the screw going in. According to media reports, against those disturbances the robot corrected and re-seated itself through real-time multimodal feedback from vision and touch; when a screw failed to bite, it automatically flagged the anomaly, retraced the step and re-aligned. The change in lighting did not break the final alignment either. A caution is due here: the shift, the misalignment and the lighting change were all test conditions applied on purpose, not a daily line already in chaos. The demo shows there is a way to correct under disturbance; it does not mean every working condition has been run through steadily. ## A Set of Figures With Caveats According to media reports, what was released on site was first-phase proof-of-concept data disclosed by the company: a 98% task success rate, 8.5 minutes per cycle, about 8 minutes for a skilled worker, 8 hours of continuous operation, and "initially qualified to join the production line." All the figures carry the same string of qualifiers, and not one word of it should be skipped. Each figure has a clear prefix: first-phase proof of concept, published by the company itself, not an industry-wide level and not a long third-party measurement. 8.5 minutes against 8 minutes is half a minute slower, which says the takt keeps up but says nothing about running for a full year. The phrase "initially qualified to join the production line" especially cannot be read quickly. Joining the line means being folded into the formal production line; the original says the conditions are initially met, still a step away from being scheduled into a shift, and far from handing the whole line over. Drop the word "conditions" and the meaning changes entirely. What a set of proof-of-concept data can show is that it works at this station, in this sequence, over this stretch of time. It cannot answer whether other models can copy it, whether accuracy holds with a new batch of parts, or whether it drifts over months of running. Those answers have to be earned with longer hours. ## The Failure Rate Comes Down According to media reports, the company's executive told the media that when the hands were first fitted to full machines the failure rate was close to 10%, and has since fallen below 0.5%; he credits an automotive-grade R&D system, saying much of the core team came from the auto industry and brought in reliability engineering and volume-production control. Keep the qualifier here too: this is a change the executive describes himself, with no third-party verification in public information, and it does not mean other makers reached the same level. From close to 10% to below 0.5% sounds like a large drop, but the starting point was its own early days on full machines. According to media reports, the executive also noted that dexterous hands used to last only two or three weeks, the root cause being long stays in labs where wear, disturbance and fatigue under real conditions rarely surfaced, so hardware iteration was slow. On that account, entering the line is itself a reliability test. Between a demo and a full shift still lie the sums for consistency, durability and maintenance. A hand finishing seventy-odd steps in a row is a good start; whether it can do so every day and every batch is the answer factories are actually waiting for. One pair of hands building the next is a striking image; the hard part is repeating the sequence steadily, day after day. Even the company's own wording says only that conditions are initially met. That restraint is closer to where this line actually stands than any claim of having already entered the factory. **Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.** --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-hand-line-15-1105b Free to quote with attribution and a link to the original.