China’s Dexterous Hand Industry, 2026 Part 3 — The Players

The five kinds of company building dexterous hands in China, MIR’s composite top ten, eight vendor profiles, six field deployments, and every organisation named in the white paper linked to the RobotToday Supplier Directory.

1. Vendor Landscape

MIR segments the Chinese supply base by origin story rather than by product, which is the more useful cut: where a company came from predicts its cost structure, its natural customer and its likely weakness.

Source: MIR Databank, Table 2.

MIR composite vendor ranking

MIR’s composite list assesses integrated hardware–software delivery capability, breadth of technology-route coverage and progress in field deployment. MIR states explicitly that the list is not ranked in order of merit.

Source: MIR Databank, Figure 6, with archetype and route attributes from Tables 2 and 4 and profile coverage from Chapter 7. MIR states explicitly that the list is not ranked in order of merit — the numbering is the order of publication only. A route marked present appears somewhere in the vendor’s range; MIR does not identify a primary architecture. DexRobot and DH Robotics are on the composite list but get no Chapter 7 profile, which is why they have no vendor profile below. Company names link to the RobotToday Supplier Directory where a profile exists.

Vendor profiles

INSPIRE ROBOTS

Founded 2016, Beijing. Anthropomorphic dexterous hands and the micro servo electric cylinders that drive them, with servo drive units and hand mechanics developed and produced in-house. Hardware is committed to linkage transmission. Product ladder: RH56BFX and RH56DFX (12 DOF / 6 DOA) for speed, precision, grip force and low mass; RH56E2 adding tactile sensing and power-off self-locking; RH56F1 with multi-axis tactile sensing in a full-metal monocoque body; RH5DG2 at 18 DOF (13 DOA) with multi-axis tactile sensing; RH56H1 (11 DOF / 6 DOA) as the compact option. Downstream concentration is humanoid supply and research/education. An early entrant with mature iteration and volume delivery, already shipping in quantity to Agibot, Unitree and Galbot.

BrainCo

Founded 2015, Hangzhou. Originated in non-invasive brain–computer interfaces and multi-axis neural sensing, then extended into bionic rehabilitation limbs and humanoid dexterous hands. Revo 1 (10 DOF / 6 DOA) leads on grip force with high-speed precise control, in base and tactile versions; Revo 2 (11 DOF / 6 DOA) reduces volume and mass across base, Pro and tactile versions; the Revo 3 flagship moves to 21 DOF fully active direct drive, with the high-end variant integrating combined vision and tactile sensing. Customers include Unitree and Leju Robotics, alongside broad supply into university laboratories, robotics competitions and research institutes.

LinkerBot

Beijing; the white paper is internally inconsistent on the founding year, giving 2023 in Chapter 4 and 2018 in Chapter 7, so none is stated here. Operates a standardised multi-model product matrix across all three mainstream architectures with mature volume production and delivery. Linkage is the core line: O6 (11 DOF / 6 DOA) as the value option, L6 Industrial (11 DOF / 6 DOA), L20 Lite (20 DOF / 10 DOA), L20 (21 DOF / 16 DOA) and L25 (22 DOF / 17 DOA). Direct drive is represented by O30 at 20 active DOF; the tendon flagship L30 reaches 25 DOF (7 DOA) with high biomimetic fidelity. Paired with a full embodied-AI software stack — Dex-Serl real-machine reinforcement learning, sub-millimetre fine manipulation from vision and force fusion. Downstream: research and education, humanoid supply, and industrial flexible manufacturing, with industrial variants deployed in 3C, automotive and lithium-battery precision assembly.

Sharpa

Founded 2024. Global headquarters in Singapore, manufacturing and R&D centre in Shanghai, co-founded by core founders of Hesai. SharpaWave is the single product line: 22 active DOF on a fully direct-drive architecture with proprietary Dynamic Tactile Array vision-tactile sensing — a micro camera and over a thousand tactile pixels per fingertip. Downstream is high-end humanoid supply and leading global research institutions; humanoid customers include Galbot and Unitree. Selected by NVIDIA in 2026 as the standard hand for the NVIDIA Isaac GR00T reference humanoid.

CasiaHand

Founded 2024, Nanjing, with a core team from the Chinese Academy of Sciences Institute of Automation. Tendon-led with direct-drive fusion for payload and precision, core hardware self-controlled, paired with in-house embodied-AI algorithms. The graded Casia Hand matrix spans 11 to 25 DOF and 5 kg to 30 kg payload: Casia Hand-M11 (11 DOF / 11 DOA, hybrid, 5 kg active payload, rigid–flexible hybrid control, active cooling); Casia Hand-M6 (11 DOF / 6 DOA, linkage, ultra-compact, multimodal sensing); Casia Hand-X12 (12 DOF / 12 DOA, direct drive, true self-locking, 30 kg active payload, industrial grade); Casia Hand-X25 (25 DOF / 21 DOA, tendon, high biomimetic fidelity). Industrial flexible manufacturing is the growth vertical; research and education is the stable base.

AGILINK

Founded 2026, Shanghai, spun out of Agibot’s internal dexterous-hand division, with in-house micro drive units and multi-DOF hand structures. The OmniHand 3 series spans three routes: Lite (11 DOF / 6 DOA, linkage) as the compact, drop-resistant, cost-effective humanoid fit; Ultra-T (22+3 DOF, hybrid) with a proprietary tendon-tensioning structure for payload and grasp stability; Ultra-M (20 DOF / 20 DOA, direct drive) with full-surface 3D tactile sensing and fast force-control response. Humanoid robotics is the core application, with OmniHand serving as the principal end effector on Agibot humanoids for data collection and general grasping interaction.

WUJI

Founded 2019, Shanghai. Anthropomorphic hands and micro precision actuation, with a team drawn from micro servo and robot force-control work, committed to a fully direct-drive route. Wuji Hand and Wuji Hand 2 both run 20 active DOF: the first generation prioritises payload, low mass and impact resistance; the second improves high-frequency force-control performance, with the high-end variant carrying a full-surface flexible tactile array. Both ship with a complete hardware and software development ecosystem. Downstream is high-end research laboratories and humanoid OEM supply, with open SDKs and simulation models supplied to lower the research development barrier.

Xynova

Founded 2024, Hangzhou, led by CAE academician Xia Changliang. High-DOF anthropomorphic hands, integrated joint modules and micro electric cylinders, with motion-control algorithms and multimodal force sensing developed alongside. Xynova Flex 1 (25 DOF / 20 DOA, pure tendon, 380 g complete hand) balances payload against compliant grasping; Flex 2 (23 DOF / 19 DOA, tendon-plus-direct-drive hybrid) uses an integrated arm–hand design that relocates the main actuators to the forearm — bringing palm mass below 400 g, a figure not comparable with Flex 1’s complete-hand 380 g — while palm-mounted direct-drive joints hold positioning accuracy, with four-in-one vision, tactile, force and proximity sensing. Supplied as a standardised end effector to humanoid OEMs and, through open hardware and software interfaces, to universities and research institutes.

2. Deployment Evidence

MIR documents six field cases. They are the closest thing the report offers to evidence of commercial readiness, and they are worth reading for what they concede as much as for what they claim — Case 2 is described in the source as a PoC pilot. One caution before the table: these are vendor accounts reproduced by MIR, not third-party audited results. No throughput, yield, uptime or cost figure appears in any of them, and none should be cited as evidence of production capacity or quality performance. Read them as existence proofs that a deployment happened, not as performance data.

Source: MIR Databank, Cases 1–6. RobotToday translation. Every claim above is as stated by the vendor in the source document; neither MIR nor RobotToday has verified any of them, and no case carries quantified performance data.

3. Company Index with RobotToday Supplier Links

Every organisation named in the white paper is listed below with its role in the report. Where a profile exists in the RobotToday Supplier Directory, the company name is a live link to it. Entries marked “Not listed” had no matching profile at the time of compilation and are candidates for addition.

3.1 Dexterous hand vendors — MIR composite list

3.2 Other hand and end-effector suppliers

3.3 Humanoid robot platforms

3.4 Industrial embodied-AI robot players

3.5 Embodied-AI software and end customers

3.6 International and historical reference organisations

Academic institutions cited in the historical section — Japan’s Electrotechnical Laboratory, MIT, Stanford University, Gifu University, Beihang University (BH-3, BH-4), Harbin Institute of Technology and Shanghai Jiao Tong University — are not carried in the supplier directory as commercial entities and are omitted from the index above.

4. What This Means

The judgements in this section are RobotToday’s reading of the white paper, not MIR conclusions. Figures remain MIR’s and are attributed where used.

For robot OEMs

Buying rather than building remains the rational default through at least 2027. The specialist supply base now offers a genuine architecture ladder — linkage for cost-sensitive volume, direct drive for precision and force fidelity, tendon for compliance and mass, hybrid for platforms that need all four properties — and the leading suppliers carry more than one route. The differentiating question to put to a supplier is no longer degrees of freedom; it is what their manipulation model can do without re-teaching, and what real-machine interaction data it was trained on.

For investors

The report’s own logic points away from hardware specifications as a screening criterion. If low-level control is commoditising and hardware parameters are converging, the durable asset is the data–model–scene loop: proprietary real-machine interaction data, a manipulation model tuned to isomorphic hardware, and deployed installations generating more of both. On that test, the interesting subset of the field is small. The three companies MIR selects as its software technology cases — Genesis AI, LinkerBot and CasiaHand — are the ones that describe a closed loop of their own, with Sharpa and Xynova strong on the hardware and sensing side of it. MIR presents these as illustrative cases and neither ranks them nor states that any company has the loop running; treating them as the leading set is RobotToday’s inference.

Prepared by RobotToday. Source: MIR Industry Research / MIR Databank, 2026 China Dexterous Hand Industry White Paper. Translation, condensation, structuring and supplier-directory mapping are RobotToday’s. Forecast figures are MIR estimates as published and have not been independently verified. Company names link to profiles in the RobotToday Supplier Directory at robottoday.com/suppliers-discovery.