--- title: "The \"last centimeter\": China's dexterous hands are finally learning to feel" date: 2026-10-01 category: Humanoids & Robotics site: NeuroAI canonical: https://neuroai.site/a/na-dexterous-hands-last-centimeter language: en --- # The "last centimeter": China's dexterous hands are finally learning to feel > As humanoid bodies learn to walk, Chinese makers of dexterous hands (灵巧手) are racing to give robots a sense of touch — with Force Robot shipping past 10,000 units and a new magnetic-tactile hand that can sense the weight of a butterfly. A robot can now cross a room without tripping. Ask it to pick up a soft drink bottle without crushing it, and the illusion collapses. The gap between "can walk" and "can work" lives in the last centimeter — the hand. In China this year, that last centimeter became its own industry. Makers of dexterous hands (灵巧手) are scaling output and, just as importantly, teaching hands to feel. ## The numbers behind the hand The market is small but growing fast. According to GGII (高工产业研究院): - China's dexterous-hand sales reached about 19,200 units in 2025. - They are projected to hit 70,200 in 2026. - By 2030, GGII expects the figure to surpass 430,000. That curve is what happens when humanoid production scales and every robot needs a pair of hands. ## Force Robot: from hundreds to ten thousand Inspire Robots (因时机器人), a Changshu-based maker, delivered more than 10,000 dexterous hands in 2025 — ranked first in China by GGII for the year. Its 2026 target is 30,000 to 50,000 units, and a new Suzhou-area factory is set to lift capacity more than fourfold. The company says it serves over 700 customers. Cost is falling in step with volume. A flagship hand that sold for about ¥50,000 (≈ US$6,940) in 2023 is now closer to ¥20,000 (≈ US$2,780). For a component that can be a fifth to a quarter of a humanoid's total cost, that matters. ## The real upgrade: touch For years, robotic hands could close their fingers but had no idea what they were touching. That is changing. - In 2024, tactile sensors appeared on only about 10% of dexterous hands. - By the end of 2025, attachment had passed 50%, and today more than 60% of hands on the market carry some tactile sensing, per industry reporting. Why it matters: a hand that feels weight, texture, and hardness can gently place a glass instead of shattering it — the difference between a demo and a warehouse shift. ## A hand that senses a butterfly's weight At the June 2026 Zhangjiang Embodied Intelligence Supply Chain Conference in Shanghai, UoU (傲意科技) unveiled the ROH-AP003, a 3D magnetic tactile dexterous hand built with sensor firm Melexis (迈来芯). Its sensors can detect forces as small as 0.1 newtons — about the weight of a butterfly landing in your palm. It is described as China's first mass-produced 3D magnetic tactile hand. The point is not the spec sheet; it is that "feeling" is moving from research lab to product line. A robot hand that knows it is holding something fragile can finally do the boring, delicate jobs humans still do better. ## Why hands decide deployment Industry observers are blunt: a humanoid is only as useful as its ability to manipulate the world. As one executive put it, how many robots a company sells will ultimately depend on "how much work the hand, paired with the intelligence system, can actually do." Tactile data also feeds the brain. Hands fitted with touch sensors are now used to collect the multimodal training data that makes embodied-AI (具身智能) models smarter — turning the hardware into a data source, not just an endpoint. ## Why the cost math matters A dexterous hand can be 20% to 25% of a humanoid's total bill of materials, which makes it one of the few components expensive enough to swing a robot's whole economics. That is why the price drop from about ¥50,000 to ¥20,000 is not a footnote — it is what could make a useful humanoid affordable at all. It also explains the transmission-route fight. Linkage hands are rigid and precise (good for industry); tendon hands are soft and human-like (good for service); direct-drive hands are simple and accurate but costly. No route has won, so makers hedge by building several, which slows standardization. ## Tactile data feeds the brain The hands are not just end effectors; they are data collectors. A hand that feels can generate the touch dimension that vision-only training misses — how hard to grip, whether something is slipping, what a surface is made of. That data is exactly what embodied-AI (具身智能) models need to move from "looks right in simulation" to "works on a messy shelf." For now, high-quality tactile data is scarce, which caps how fast those models improve. The companies shipping sensing hands today are also, quietly, building the training sets for tomorrow's robot brains. ## Honest limitations - Reliability is still the wall. Some hands fail in weeks or months under real factory loads, far short of the durability users need. - Return on investment in industrial settings is unproven; many customers are still in proof-of-concept, not repeat purchasing. - Transmission routes have not converged — linkage, tendon, and direct-drive designs each trade off cost, precision, and lifespan, so there is no standard yet. - High-quality tactile data remains scarce, which limits how fast "feeling" hands can train better models. ## What readers can do now - When judging a humanoid, look past walking videos to the hand's degrees of freedom, tactile sensing, and rated lifespan — that is where real work is won or lost. - Watch dexterous-hand shipment and failure-rate data from GGII and China Electronic News as a leading indicator of embodied-AI maturity. - For investors or buyers, prioritize vendors with named industrial pilots and repeat orders over those showing only prototype dexterity. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-dexterous-hands-last-centimeter Free to quote with attribution and a link to the original.