Humanoid robot hands
Last reviewed
Sources
13 citations
Review status
Source-backed
Revision
v5 ยท 4,618 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
13 citations
Review status
Source-backed
Revision
v5 ยท 4,618 words
Add missing citations, update stale details, or suggest a clearer explanation.
Humanoid robot hands are dexterous end effectors that replicate the grasping, manipulation, and tactile perception of the human hand, which itself has roughly 27 degrees of freedom.[8] They are widely regarded as the hardest subsystem in humanoid robot design, because human-like manipulation demands packing many actuators, joints, and touch sensors into a hand-sized volume while controlling them in real time. As of 2026, AI Wiki documents 120 humanoid robots with hand or gripper systems, ranging from simple two-finger grippers to fully dexterous five-finger hands with more than 20 degrees of freedom (DOF) per hand.[1]
The most capable designs now approach human anatomy: Sanctuary AI's Phoenix hand has 21 DOF and tactile sensors that detect forces as low as 5 millinewtons, close to the roughly 3-millinewton sensitivity of a human fingertip,[2] while the research-grade Shadow Dexterous Hand provides 24 degrees of freedom and over 100 sensors sampled at up to 1 kHz.[3] The field has advanced rapidly, with several robots now manipulating small objects, operating tools, and performing assembly tasks in real factory settings.
A humanoid robot hand is an anthropomorphic end effector mounted on a robot arm to perform the grasping and fine manipulation a human hand would. Its capability is usually summarized by three numbers: the count of degrees of freedom (independently controllable joint motions), the number of fingers, and the density and sensitivity of its tactile and force sensing. A higher DOF count allows more grasp types and in-hand reorientation; richer touch sensing lets the hand modulate grip force and detect slip without vision.
Humanoid robot hands generally fall into several categories based on their design philosophy and intended use.
Dexterous hands attempt to replicate the full range of human hand motion, typically featuring five fingers with 10 or more degrees of freedom per hand. These hands can perform precision grasps (picking up small objects between fingertips) and power grasps (wrapping around larger objects). Examples include the Figure 02 (16 DOF per hand), XPENG IRON (22 DOF per hand), and Phoenix (21 DOF per hand).[1]
The main challenge with dexterous hands is controlling all the joints simultaneously while maintaining stable grasps. Most dexterous hands use a combination of tendon-driven mechanisms and direct-drive motors.
Gripper-style hands use fewer fingers (typically 2-3) with simpler mechanics. They trade dexterity for reliability, grip strength, and lower cost. Robots like the Atlas (3-finger gripper) and Unitree H1 use this approach.[1] Grippers are well suited for pick-and-place tasks in structured environments.
Some robots use unconventional hand designs optimized for specific applications:
| Hand type | Number of robots |
|---|---|
| Dexterous | 14 |
| Gripper | 5 |
| 20-dof dexterous hands with tactile sensing | 1 |
| Three-fingered dexterous grippers | 1 |
| Dexterous multi-finger | 1 |
| Ball gripper (no dexterous hands) | 1 |
| Interchangeable | 1 |
| Magic-hand dexterous end effector | 1 |
The following table lists all 120 humanoid robots with documented hand specifications, sorted by degrees of freedom per hand.[1]
| Rank | Robot | Manufacturer | DOF per hand | Notes |
|---|---|---|---|---|
| 1 | MATRIX-3 | Matrix Robotics | 27 | 5 fingers |
| 2 | Clone Alpha | Clone Robotics | 26 | N/A |
| 3 | 1X Neo | 1X | 22 | 5 fingers, dexterous |
| 4 | NEO | 1X Technologies | 22 | 5 fingers |
| 5 | NEO Gamma | 1X Technologies | 22 | N/A |
| 6 | North | Sharpa | 22 | N/A |
| 7 | Optimus Gen 3 | Tesla | 22 | N/A |
| 8 | Walker S2 | UBTECH Robotics | 22 | 5 fingers |
| 9 | XPENG IRON | XPENG Robotics | 22 | 5 fingers, dexterous |
| 10 | Agile One | Agile Robots SE | 21 | 5 fingers |
| 11 | Phoenix | Sanctuary AI | 21 | 5 fingers, dexterous |
| 12 | Figure 03 | Figure | 20 | 5 fingers |
| 13 | Friday | Holiday Robotics | 20 | N/A |
| 14 | QUANTA X2 | X Square Robot | 20 | 5 fingers, 20-DoF dexterous hands with tactile sensing |
| 15 | A2 Ultra | AgiBot | 19 | N/A |
| Rank | Robot | Manufacturer | Grip force (N) |
|---|---|---|---|
| 1 | ALLEX | WIRobotics | 40 |
| 2 | MenteeBot V3 | Mentee Robotics | 30 |
86 robots feature five-finger hands, the closest analog to the human hand.[1] Among these, the most dexterous are:
| Robot | Hand DOF | Grip force (N) | Hand type |
|---|---|---|---|
| MATRIX-3 | 27 | N/A | N/A |
| 1X Neo | 22 | N/A | dexterous |
| NEO | 22 | N/A | N/A |
| Walker S2 | 22 | N/A | N/A |
| XPENG IRON | 22 | N/A | dexterous |
| Agile One | 21 | N/A | N/A |
| Phoenix | 21 | N/A | dexterous |
| Figure 03 | 20 | N/A | N/A |
| QUANTA X2 | 20 | N/A | 20-DoF dexterous hands with tactile sensing |
| RoBee | 18 | N/A | N/A |
| TORA DoubleOne | 18 | N/A | N/A |
| DexBot | 17 | N/A | N/A |
| Figure 02 | 16 | N/A | dexterous |
| ALLEX | 15 | 40 | N/A |
| Unitree G1 | 14 | N/A | Three-fingered dexterous grippers |
| Honda ASIMO | 13 | N/A | dexterous |
| Atom Max | 12 | N/A | N/A |
| HIVA Haiwa | 12 | N/A | N/A |
| HMND 01 Alpha Bipedal | 12 | N/A | N/A |
| Motion 2 | 12 | N/A | N/A |
The most-cited 2025-2026 hand designs each took a distinct approach to actuation and touch sensing. Tesla's Optimus Gen 3 hands use a tendon-driven layout with 22 degrees of freedom across both hands (11 per hand) and 25 actuators per forearm and hand, which Tesla says is roughly a 4.5x increase over Gen 2; the fingertips add force-feedback sensors so the robot can register how hard it is gripping.[4] Figure's third-generation Figure 03 introduced proprietary fingertip sensors with 3-gram tactile sensitivity, enough to feel the weight of a paperclip, after the company concluded that off-the-shelf tactile sensors could not withstand real-world use.[5] Sanctuary AI's Phoenix hand uses hydraulic actuation, which Sanctuary says offers an order of magnitude higher power density than cable and electromechanical systems, with 21 degrees of freedom and 5-millinewton tactile sensitivity that enables blind picking and in-hand reorientation.[2][6]
The Shadow Dexterous Hand from Shadow Robot Company remains a long-standing research benchmark. It provides 24 degrees of freedom (20 actuated), with the thumb contributing 5 DOF, and ships with Pressure Sensor Tactiles in the fingertips as standard; a fully instrumented hand carries 129 sensors, more than 100 of them running at up to 1 kHz.[3]
| Hand | DOF | Actuation | Tactile sensing |
|---|---|---|---|
| Shadow Dexterous Hand | 24 (20 actuated) | Tendon (forearm motors or pneumatics) | Fingertip pressure sensors; up to 129 sensors at 1 kHz[3] |
| Phoenix hand | 21 per hand | Hydraulic | 5 mN sensitivity, in-hand manipulation[2] |
| Figure 03 hand | 20 per hand | Tendon/electric | 3-gram fingertip sensitivity, palm camera[5] |
| Optimus Gen 3 hand | 11 per hand | Tendon, 25 actuators per arm | Force-feedback fingertip sensors[4] |
While many humanoid robots use captive hands built by the same company that integrates the rest of the body, a parallel ecosystem of standalone dexterous hand suppliers sells hands as components to outside humanoid integrators, research labs, and industrial automation customers. These suppliers do not appear in the per-robot table above because their hands are not tied to a single platform; the same product may ship on multiple unrelated humanoids and inside teleoperation rigs.
As of mid-2026 the leading external suppliers in the high-DOF segment are concentrated in China, with the following representative products.[7]
| Supplier | Product | Year revealed | Active DOF | Total DOF | Palm weight | Whole-hand grasp | Drive |
|---|---|---|---|---|---|---|---|
| Xynova | Flex 1 | August 2025 | 20 | 25 | 380 g | More than 30 kg | Pure tendon |
| Xynova | Flex 2 | May 13, 2026 | Not separately disclosed | 23 | 400 g | 12 kg (4 kg rated continuous) | Hybrid cable plus direct drive |
| Wuji Tech | Wuji Hand | September 2025 | 20 | 20 | 580 g (skeleton) | 10 kg (static) | In-hand direct drive |
| Linkerbot | LinkerHand L30 | 2025 | 22 | 22 | Not published | Not published | Tendon, forearm motors |
| Inspire Robots | RH56 series | 2023 to 2025 | 6 to 12 | 6 to 12 | Not published | Not published | Linkage drive |
| Shadow Robot | Shadow Hand (research) | 2005 | 24 | 24 | About 4 kg with forearm | Not published | Tendon, forearm pneumatics or motors |
These component suppliers have become a structural feature of the humanoid market. The Chinese dexterous hand market shipped more than 30,000 units in 2025, with humanoid-grade hands accounting for an estimated 15,000 units (up from about 2,000 the year before), and forecasts published in early 2026 expect the global market to reach roughly 1.4 million units and about 3 billion US dollars in revenue by 2030.[7] The Xynova Flex 2 in particular drew attention at launch for combining a 23-DOF hybrid drive layout with +/-0.1 mm positional repeatability, 0.05 N force control, and a 400 g palm, on a back-drivable mechanism with multimodal tactile, force, and wrist-camera sensing and a "cerebellum"-style adaptive grasping layer.[10][11][12]
The supplier ecosystem also matters because it shapes the strategic decisions of integrated platform makers. Xiaomi Strategic Investment, for example, joined Xynova's December 2025 angel round (alongside CATL Capital, Zhengxuan Investment, Orient Renaissance Capital, SEARI Capital, and the L2F Ray Entrepreneur Fund) and increased its position in the March 2026 Pre-A, even though Xiaomi continues to develop a captive bionic hand for its CyberOne platform.[9][13] Genesis AI's bimanual dexterous demonstration in May 2026 used a Wuji-built variant of the Wuji Hand renamed Genesis Hand 1.0. These cross-cutting investments and partnerships suggest the market is converging on a layered structure in which a small number of component suppliers serve a much larger number of humanoid integrators.
Robot hands use several actuation methods:
Effective manipulation requires rich sensory feedback:
Hand control approaches include:
Several open problems remain in humanoid robot hand design: