# End effector

> Source: https://aiwiki.ai/wiki/end_effector
> Updated: 2026-07-14
> Categories: AI Hardware, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

An **end effector** is the device mounted at the working end of a robot arm or wrist: the part that actually touches, grips, welds, or otherwise acts on the world, while the rest of the arm exists only to position and orient it. The term pairs "end," marking it as the terminal link of a [manipulation](/wiki/manipulation) chain, with "effector," a word physiology borrowed from the Latin verb *efficere* ("to bring about" or "to effect") for any organ or muscle that carries out a command from the nervous system; robotics engineers later adopted the same word for the mechanical equivalent [1][2]. End effectors range from a pair of steel jaws worth a few hundred dollars to a 20-plus-joint anthropomorphic hand worth as much as a luxury car, and the choice between simple and complex is one of the most consequential, and least settled, design questions in [humanoid robot](/wiki/humanoid_robot) engineering.

In brief: everything else on a robotic arm (the motors, the [degrees of freedom](/wiki/degrees_of_freedom), the control software) exists to get the end effector to the right place with the right force. On a factory arm that end effector is as likely to be a suction cup, a magnet, or a welding torch as a hand; on a humanoid it is almost always some kind of hand, because doorknobs, hand tools, and packaging were all designed for one.

## How it works

In the formal language of robot kinematics, a manipulator is a chain of rigid links joined by degrees of freedom, and the end effector is the last link in that chain: whatever attaches to the wrist after the joints that position and orient it have done their job [1]. That framing separates two different engineering problems. Getting an end effector to a location and orientation in space is a matter of arm design and [robot manipulation](/wiki/robot_manipulation) control; what happens once it arrives (whether it can grip a wine glass without crushing it, sense that a bolt has seated, or hold a tool steady against a moving part) is a separate problem that lives entirely inside the end effector itself.

Every end effector has to solve three linked problems: how to generate a holding force or apply a process (gripping, suction, magnetism, adhesion, penetration, heat, or a jet of material), how to sense what it is touching, and how to route both power and data through the wrist without the cabling becoming a point of failure. [Tactile sensing](/wiki/tactile_sensing) closes the loop between the first two: without some way to measure contact force, slip, or position, an end effector can only be commanded open or closed rather than controlled intelligently. Recent comparisons of robotic and human hands increasingly point to that sensing and control gap, rather than raw mechanical complexity, as the main thing separating today's hardware from human-level dexterity [3].

## Types of end effectors

Industrial engineering divides end effectors into two broad families: grippers, which hold a workpiece so the arm can move it, and process tools, which act on a workpiece without necessarily holding it, such as a welding gun, a paint sprayer, or a drill [1][4]. Grippers are further organized by the physical principle they use to hold on. The most widely cited classification comes from the engineering reference text *Robot Grippers*, by Gareth Monkman, Stefan Hesse, Ralf Steinmann, and Henrik Schunk, which splits grippers into four categories: impactive, astrictive, ingressive, and contigutive [5].

### The four-category taxonomy

**Impactive** grippers grasp through direct mechanical contact: jaws, claws, or fingers close onto an object and hold it by friction or a form fit. This is the largest and most familiar category, running from simple two-jaw pneumatic grippers through underactuated adaptive grippers to fully articulated [dexterous hand](/wiki/dexterous_hand)s with a dozen or more independently sensed joints [5].

**Astrictive** grippers hold an object through an attractive force applied to its surface rather than a mechanical pinch: vacuum suction, magnetism, or electroadhesion [5]. Vacuum grippers are the industrial default for flat, non-porous, single items: a vacuum generator drops the pressure inside a suction cup so ambient atmospheric pressure holds the part against it, and releasing the vacuum drops the part [6]. Magnetic grippers work only on ferromagnetic materials such as steel, iron, and nickel, but need no seal and tolerate holes, uneven surfaces, and some surface contamination that would defeat a suction cup [7]. Electroadhesion is the newest and least common of the three: a voltage applied across closely spaced electrodes induces an opposing charge in the target surface, generating a clamping force that works on a wide range of materials, including irregular and delicate ones, using very little power, though generally at lower force per unit area than vacuum [8].

**Ingressive** grippers penetrate the surface of the target rather than holding it from outside, using pins, needles, or hackles. Textile, fiberglass, and carbon-fiber handling in composite manufacturing rely on ingressive tooling because those materials are too limp or fibrous for a rigid jaw to grip reliably [5].

**Contigutive** grippers rely on direct surface-contact bonding, such as glue, surface tension, or freezing, rather than a force field or mechanical closure. Adhesive and cryogenic grippers occupy this category; they are far less common in general manufacturing but appear in niche applications such as handling wet or irregularly shaped food products [5].

Beyond these four, a handful of specialized non-contact techniques (Bernoulli grippers that float an object on an air cushion, gecko-inspired dry-adhesive grippers, and capillary or ultrasonic grippers) extend astrictive and contigutive principles to delicate or micro-scale parts, though they remain uncommon outside specialty electronics and laboratory automation [1].

| Category | Holding principle | Common subtypes | Typical use | Humanoid relevance |
| :---- | :---- | :---- | :---- | :---- |
| Impactive | Direct mechanical grip (friction or form fit) | Parallel-jaw grippers; underactuated adaptive grippers; multi-finger [dexterous hands](/wiki/dexterous_hand) | General pick-and-place; assembly; manipulation | Dominant approach for humanoid hands |
| Astrictive | Attractive force on the surface | Vacuum/suction; magnetic; electroadhesive | Flat panels, boxes, glass; ferromagnetic sheet metal; delicate or irregular items | Rare on humanoids; occasional task-specific attachment |
| Ingressive | Surface penetration | Pins, needles, hackles | Textiles, carbon and glass fiber sheets | Not used; humanoid hands avoid piercing tools |
| Contigutive | Direct-contact bonding | Adhesive; surface tension; cryogenic (freezing) | Wet, delicate, or irregular items (e.g. food) | Not used on production humanoid hands |

### Task-specific tools and quick-change systems

Not every end effector grips at all. Spot-welding guns, MIG and TIG welding torches, paint spray guns, adhesive and sealant dispensers, deburring and polishing heads, and screwdriving spindles are all end effectors in the strict sense: the device that does the arm's actual job [4]. In fixed industrial cells, a single robot often needs to run several such tools across a shift, so manufacturers standardized automatic tool changers: a robot-side and tool-side coupling pair that locks and unlocks in seconds and passes pneumatic, electrical, and data lines through the mechanical connection. ATI Industrial Automation and Schunk are the two most established suppliers of this hardware, and their couplers are rated for millions of cycles at full rated load [9].

Humanoid platforms have been slower to adopt quick-change tooling, because most current hands are built as permanent, integrated parts of the wrist and forearm. [Apptronik](/wiki/apptronik) is a rare public exception: its Apollo humanoid ships with a quick-release wrist that lets a simple gripper be swapped for a different end effector in the field, a deliberate hedge that lets the company sell narrower, cheaper hand modules for specific jobs rather than a single expensive dexterous hand for everything [10].

## Tradeoffs and key evaluation criteria

Because "end effector" spans everything from an inexpensive pneumatic clamp to a six-figure robotic hand, engineers evaluate them against a common set of criteria rather than a single "best" design.

**Dexterity** measures how many degrees of freedom the end effector can control on a grasped object, not just on itself. A simple gripper controls exactly one: open or closed. A fully articulated hand can, in principle, control an object's complete position and orientation once grasped; most real end effectors fall somewhere between those extremes [1][3].

**Robustness** asks whether the end effector survives repeated, unplanned contact with the environment: drops, collisions, dust, and thousands of duty cycles without maintenance. For production deployments, robustness usually matters more than peak dexterity, which is one reason simple grippers remain dominant on factory floors [3].

**Force-to-weight ratio** compares how much holding or lifting force an end effector produces relative to its own mass. The human hand is the usual benchmark: it weighs roughly 0.4 kilograms yet can produce grip forces exceeding 400 newtons, a ratio robotic hands have not matched [11]. Weight matters disproportionately on a humanoid because a heavier hand adds inertia at the far end of the arm, which the actuators and structure further up the chain must then be sized to control.

**Controllability** is a caution against counting degrees of freedom as a proxy for capability: a hand with 24 independently actuated joints is only as good as the software driving it. Several recent studies find no reliable correlation between a hand's mechanical complexity (its finger and joint count) and the range of tasks it can actually perform, which is why sensing and control, not actuator count, are usually the binding constraint [3].

**Cost** varies by roughly three orders of magnitude across the category. A basic single-acting pneumatic parallel gripper can be bought for well under $200 [40]; a popular adaptive two-finger electric gripper such as Robotiq's 2F-85 lists at around $5,000 [12]; a 16-degree-of-freedom research-grade hand such as the Allegro Hand runs from roughly $15,000 to $24,000 [13]; and flagship humanoid hands with dense tactile sensing, such as the Shadow Dexterous Hand or Sharpa's SharpaWave, are priced from roughly $50,000 to $100,000 for a single hand, or well over $150,000 for a matched pair integrated into a humanoid [14][15][16]. For most commercial applications, a simpler, cheaper end effector wins on unit economics even if it does less.

## Use in humanoid robots

Humanoid robots make the end effector question unusually hard, because the whole premise of the form factor, a machine built to work in spaces designed for humans, points toward hands, while the whole discipline of engineering points toward the simplest mechanism that gets the job done. As of mid-2026 there is no industry consensus on where that tradeoff should land, and the major humanoid developers have made visibly different bets. For a deeper look at the hand side of this question specifically, see [humanoid robot hands](/wiki/humanoid_robot_hands).

### Anthropomorphic hands versus simpler grippers

At one end of the spectrum, [Tesla](/wiki/tesla) has committed to a highly anthropomorphic five-fingered hand for [Tesla Optimus](/wiki/tesla_optimus). Patents published in 2026 for the [Tesla Optimus Gen 3](/wiki/tesla_optimus_gen_3) hand describe 22 degrees of freedom per hand, roughly double the prior generation, driven by around 25 actuators per arm, all relocated from the hand into the forearm and connected to the fingers by tendons in a layout that mimics human forearm muscles [17].

At the other end, [Sunday Robotics](/wiki/sunday_robotics) builds its Memo home robot around a simplified three-finger, dual-gripper end effector rather than a five-fingered hand, a decision the company ties directly to its data pipeline: Sunday collects manipulation demonstrations with a wearable "Skill Capture Glove" and then designed the robot's end effector to match the glove's own geometry rather than a full human hand [18].

Between those poles, [Boston Dynamics](/wiki/boston_dynamics) has visibly changed its mind in public. In late 2025 the company detailed a deliberately non-anthropomorphic three-fingered gripper for its electric Atlas humanoid, arguing that fewer, simpler digits could still cover most industrial tasks [19]; by its CES 2026 production unveiling it had switched to a more conventional four-digit design (three fingers and an opposable thumb), reportedly because the earlier gripper's unusual geometry struggled with tools designed around the five-fingered human hand [20]. Apptronik's Apollo takes minimalism further still, using a one-degree-of-freedom pincer gripper for most of its production units rather than solving general dexterity up front, on the theory that a robot can ship and earn revenue on narrower tasks while dexterous manipulation matures separately [10].

Academic research increasingly backs the minimalists. A 2025 study from the German Research Center for Artificial Intelligence (DFKI) and Leibniz University Hannover reviewed robot hand designs against the tasks they could actually demonstrate and found no significant relationship between a hand's number of fingers or degrees of freedom and its measured skill repertoire. Three fingers, the authors concluded, is "a good compromise between simplicity and dexterity" that matched or beat both two- and five-fingered designs across the tasks surveyed, while wrist flexibility and the ability to spread fingers sideways (abduction and adduction) correlated with capability more strongly than raw finger count [3]. That lines up with older evidence from robotics competitions: entries at the DARPA Robotics Challenge mostly used underactuated three- or four-finger hands rather than fully anthropomorphic ones, and the first Amazon Picking Challenge was won by a team using plain suction rather than fingered grippers at all [3][21].

Part of the reason simple end effectors go further than they look is that a hand does not have to do all the work alone. Research on "extrinsic dexterity" at MIT showed that a basic two-finger parallel gripper can reorient a grasped object by pressing it against a table edge, wall, or other environmental fixture, effectively borrowing a third contact point from the world instead of building it into the hand [22]. A robot that can plan those environmental contacts intelligently needs less built-in mechanical dexterity to begin with.

### Number of fingers

How many fingers an end effector needs remains an open engineering question, and current humanoid and gripper designs span the full range from two to more than five.

**Two fingers** are the default for pick-and-place and for much of today's foundation-model manipulation research. Simple parallel-jaw grippers dominate widely used research arms and data-collection rigs, including the handheld Universal Manipulation Interface (UMI), a portable parallel-jaw gripper fitted with an action camera that Stanford, Columbia, and Toyota Research Institute developed for collecting manipulation demonstrations in the wild for training robot policies [23]. Two fingers are cheap and mechanically simple, and combined with extrinsic dexterity they can cover a surprising share of manipulation tasks [22].

**Three fingers** add a stable tripod grasp and enough in-hand adjustment for basic tool use without the cost and control complexity of a full hand. Sunday Robotics' dual-gripper and Boston Dynamics' first-generation electric Atlas gripper both landed here [18][19].

**Four fingers** typically mean a thumb plus three fingers, dropping the little finger, which contributes least to most human grips. Boston Dynamics' production Atlas hand, unveiled at CES 2026, followed this pattern after its three-fingered prototype struggled with tools designed for human hands [20].

**Five fingers** are chosen when a design targets direct compatibility with human tools and environments, or eventual prosthetic-grade dexterity. This is the most crowded category in 2026. It includes Tesla Optimus; [Figure AI](/wiki/figure_ai)'s [Figure 03](/wiki/figure_03), whose hand third-party trackers estimate at around 20 degrees of freedom, though Figure has not published an official specification [24]; [Unitree](/wiki/unitree)'s Dex5 series, with 20 degrees of freedom and a list price around $25,000 [25]; [XPeng Iron](/wiki/xpeng_iron)'s 22-degree-of-freedom hand [26]; [Shadow Robot](/wiki/shadow_robot)'s long-running Dexterous Hand, with 24 degrees of freedom, tendon-driven actuation, and a historical price around $100,000 [14][15]; and [Sharpa](/wiki/sharpa)'s SharpaWave, a 22-degree-of-freedom hand with more than 1,000 tactile sensing points per fingertip that reportedly costs roughly $50,000 for a single hand or $150,000 to $200,000 for a matched pair integrated into a humanoid [16]. Five-fingered hands remain expensive, thermally constrained, and rarely demonstrated on tasks as complex as their joint counts suggest, a gap several researchers attribute to control and sensing limits rather than mechanical ones [3].

**More than five fingers** is largely unexplored outside research, though it is not an absurd idea. A widely cited 2019 study of people with six-fingered hands (polydactyly) found they could coordinate the extra digit with the rest of the hand for tasks that normally require two hands, controlled by a distinct region of motor cortex dedicated to the sixth finger, evidence that more actuated digits than the human baseline can be controlled without a net loss of function [27]. No humanoid manufacturer had shipped a hand with more than five fingers as of mid-2026.

For comparison, the human hand itself has 27 bones and 27 degrees of freedom (four per finger, five in the thumb, and six at the wrist), packed with roughly 17,000 mechanoreceptors in the skin of the palm and fingers alone [28][29]. No robotic hand approaches that sensor density, which is part of why the finger-count debate is really a proxy for a harder, unsolved sensing problem.

| Fingers | Typical grasp | Representative examples | Notes |
| :---- | :---- | :---- | :---- |
| 2 | Parallel pinch; relies on environment for reorientation | Robotiq 2F-85; UMI data-collection gripper | Cheapest and simplest to control; widely used in foundation-model manipulation research |
| 3 | Tripod grasp; basic tool use | Sunday Robotics Memo gripper; Boston Dynamics' 2025 Atlas gripper | DFKI review found three fingers a strong complexity-to-capability compromise |
| 4 | Thumb plus three fingers, little finger dropped | Boston Dynamics' 2026 production Atlas hand | Added mainly to grip human-designed tools |
| 5 | Full anthropomorphic hand | Tesla Optimus, Figure 03, Unitree Dex5, XPeng Iron, Shadow Dexterous Hand, Sharpa Wave | Most expensive, heaviest tier; control and sensing are the binding constraint, not mechanics |
| 6+ | Experimental only | None shipped as of 2026 | Human polydactyly research suggests extra digits are controllable without a net loss of dexterity |

### Actuation architecture

Separate from finger count, how an end effector's joints are powered has just as much effect on cost, weight, and behavior.

**Fully actuated** designs give every degree of freedom its own motor, roughly one actuator per joint, so a three-jointed finger needs four motors once abduction is included. This maximizes independent control and force output per joint, but packing that many motors into a hand-sized volume is mechanically brutal: the motors either sit inside the fingers, adding bulk and inertia exactly where designers most want to minimize it, or in the palm or forearm connected by extra linkages. Fully actuated hands remain largely confined to laboratories and high-end research platforms [3].

**Underactuated** designs use fewer motors than degrees of freedom, linking multiple joints through springs, tendons, or mechanical linkages so a single actuator drives a coordinated, adaptive closing motion. The Robotiq 3-Finger Adaptive Gripper, a commercial example with 4 actuators driving 10 degrees of freedom, closes each finger until it meets resistance, letting the fingers passively wrap around whatever shape they encounter without needing to know that shape in advance [30]. This self-correcting behavior is a major reason underactuated grippers dominate commercial deployments: they trade the ability to control every joint independently for mechanical simplicity, lower cost, and grasp reliability on unmodeled objects.

[**Tendon-driven**](/wiki/tendon_driven) designs move the motors out of the hand entirely, usually into the forearm, and pull the fingers with cables, echoing the layout of human forearm muscles and tendons. This keeps the hand itself slim and light, which matters because hand mass sits at the far end of the arm and multiplies actuator demands upstream, but it introduces friction, cable stretch, and routing complexity that can degrade precision over time and adds maintenance burden. The Shadow Dexterous Hand is the longest-standing example, packing 20 motors and 24 degrees of freedom into a human-sized envelope this way [14]. Tesla's Optimus hand and 1X's Neo hand both adopted the same forearm-tendon layout; 1X specifically uses a low gear ratio, roughly 5:1 to 15:1, that keeps the hand backdrivable and able to sense external force through the tendons themselves rather than through added sensors [17][31].

**Hydraulic and pneumatic** actuation replaces electric motors with fluid power, an approach only a few humanoid developers have pursued for hands specifically. [Sanctuary AI](/wiki/sanctuary_ai)'s Phoenix hand uses miniaturized hydraulic valve actuators about the size of a coin, which the company says deliver an order of magnitude higher power density than comparable electric or cable-driven designs while surviving more than two billion test cycles without leaking [32]. [Clone Robotics](/wiki/clone_robotics) has gone further, building a 27-degree-of-freedom hand actuated by dozens of proprietary artificial muscles, a hydraulic descendant of the classic McKibben pneumatic muscle, run by a hybrid hydraulic-pneumatic system and a dedicated onboard pump [33]. Both companies argue fluid actuation is the only practical way to combine human-scale speed, force, and compliance in a hand-sized package; the tradeoff is the added complexity, weight, and potential failure modes of pumps, valves, and fluid lines that electric designs avoid entirely.

A cross-cutting design choice inside any of these architectures is the actuator's gear ratio. High-ratio geared actuation, whether tendon-driven or built into each finger, maximizes torque but tends to mask outside forces from the motor's own current sensing. Low-ratio, or quasi-direct-drive, actuation sacrifices some torque density to keep the joint backdrivable and torque-transparent, so the controller can infer contact force directly from motor current rather than needing a separate force sensor at every joint [34]. [RobotEra](/wiki/robotera)'s XHand1 and 1X's Neo hand both use this approach [35][31]. Wuji Tech takes a related but distinct path, embedding direct-drive micro-actuators inside each finger phalanx rather than routing tendons from the forearm at all, trading the compactness of a slim wrist for the elimination of tendon friction and stretch [36]. [Xynova](/wiki/xynova)'s Flex 2 hand blends both philosophies in a single hybrid-drive system, combining backdrivable actuation with a claimed force-control resolution of 0.05 newtons across 23 degrees of freedom in a 400-gram hand [37].

| Architecture | Motor location | Typical actuator-to-DOF ratio | Strengths | Weaknesses | Examples |
| :---- | :---- | :---- | :---- | :---- | :---- |
| Fully actuated | In-hand or forearm; one motor per joint | About 1:1 | Maximum independent control and force per joint | Hard to package; adds hand mass and inertia; mostly lab-only | Research-grade multi-finger hands |
| Underactuated | In-hand, shared across linked joints | Fewer motors than DOF (e.g., 4 actuators for 10 DOF) | Adaptive, self-correcting grasp; simple, cheap, robust | Cannot control every joint independently; less precise in-hand manipulation | Robotiq 3-Finger Adaptive Gripper |
| Tendon-driven | Forearm, cables routed to fingers | Varies; often near 1:1 at low gear ratios | Slim, light fingers; low finger inertia; mimics human tendon layout | Cable friction and stretch; routing complexity; maintenance | Shadow Dexterous Hand; Tesla Optimus hand; 1X Neo hand |
| Hydraulic / pneumatic | Forearm or hand, via valves and fluid lines | Varies by design | High power density; strong force-to-weight; smooth compliant motion | Pumps, valves, and fluid management add complexity and failure modes | Sanctuary AI Phoenix hand; Clone Robotics hand |

## Suppliers and landscape

The commercial market for hands and grippers spans decades-old industrial suppliers and a wave of well-funded humanoid-focused startups, with China in particular producing several companies competing directly on cost and degrees of freedom; known Chinese dexterous-hand shipments passed 30,000 units in 2025, and domestic hand prices have fallen from several hundred thousand yuan to roughly 30,000 to 80,000 yuan (an estimated $4,000 to $11,000) as a local supply chain matured [38].

| Maker | Product | Approach | Notable spec | Notes |
| :---- | :---- | :---- | :---- | :---- |
| Robotiq | 2F-85 adaptive gripper | Underactuated, 2-finger | About $5,000; 20 to 235 N grip force | Best-selling gripper for collaborative arms [12] |
| [Shadow Robot](/wiki/shadow_robot) | Dexterous Hand | Tendon-driven, 5-finger, 24 DOF | About $100,000; 20 motors, 100+ sensors | Longest-running commercial dexterous hand [14][15] |
| [Sharpa](/wiki/sharpa) | SharpaWave | Tendon and tactile hybrid, 5-finger, 22 DOF | About $50,000 per hand; 1,000+ tactile points per fingertip | In mass production as of mid-2026 [16] |
| [Sanctuary AI](/wiki/sanctuary_ai) | Phoenix hand | Hydraulic valve actuation, 5-finger, 21 DOF | Coin-sized valves; 2 billion-plus test cycles | Claims order-of-magnitude power-density gain over electric designs [32] |
| [Clone Robotics](/wiki/clone_robotics) | Clone Hand | Hydraulic/pneumatic artificial muscle, 5-finger, 27 DOF | 36 artificial muscles; 500 W onboard pump | Among the most biomimetic commercial hands attempted to date [33] |
| [Tesla](/wiki/tesla) | Optimus Gen 3 hand | Tendon-driven, 5-finger, 22 DOF | About 25 actuators per arm | Production targeted for 2026 [17] |
| [1X Technologies](/wiki/1x_technologies) | [1X Neo](/wiki/1x_neo) hand | Quasi-direct-drive tendon, 5-finger, 25 DOF | 5:1 to 15:1 gear ratio; fully force-controlled | Manufacturing line targeted 10,000 hands in 2026 [31] |
| [Figure AI](/wiki/figure_ai) | [Figure 03](/wiki/figure_03) hand | Electromechanical, 5-finger | About 20 DOF (third-party estimate) | Tuned to work with Figure's [Helix](/wiki/figure_helix) manipulation stack [24] |
| [Unitree](/wiki/unitree) | Dex5-1 | Tendon/direct hybrid, 5-finger, 20 DOF | About $25,000; 94 tactile sensors | Priced well below Western flagship hands [25] |
| [XPeng Iron](/wiki/xpeng_iron) | Iron hand | Harmonic-geared, 5-finger, 22 DOF | Integrated with roughly 200 total body DOF | [26] |
| [RobotEra](/wiki/robotera) | XHand1 | Quasi-direct-drive, 5-finger, 12 DOF | 80 N max force; 100+ tactile points per finger | [35] |
| Wuji Tech | [Wuji Hand](/wiki/wuji_hand) | Direct-drive-in-finger, 5-finger, 20 DOF | 0.9 kg; about 15 N fingertip force; 300,000+ cycles tested | Motors embedded in finger phalanges rather than the forearm [36] |
| [ZWHand](/wiki/zwhand) | ZWHAND DM17 | Micro-drive modules, 5-finger, 17 to 20 DOF | E-skin tactile covering | Spun out of gear and micro-motor maker Zhaowei [39] |
| [Xynova](/wiki/xynova) | Flex 2 | Hybrid drive (direct and tendon), 5-finger, 23 DOF | 400 g; 0.05 N force-control resolution | Raised over $13 million in an angel round led by CATL Capital [37] |
| [PaXini Technology](/wiki/paxini_technology) | Tactile dexterous hands | Hall-effect tactile sensing focus | Core supplier to multiple humanoid integrators | One of at least four Chinese dexterous-hand unicorns as of 2026 [38] |
| Wonik Robotics | [Allegro Hand](/wiki/allegro_hand) | Fully actuated, research-grade | 16 DOF; about $15,000 to $24,000 | Long-standing academic research standard [13] |

## See also

- [Dexterous hand](/wiki/dexterous_hand)
- [Tendon-driven](/wiki/tendon_driven)
- [Tactile sensing](/wiki/tactile_sensing)
- [Humanoid robot hands](/wiki/humanoid_robot_hands)
- [Robot manipulation](/wiki/robot_manipulation)
- [Degrees of freedom](/wiki/degrees_of_freedom)
- [Shadow Robot](/wiki/shadow_robot)

## References

1. Wikipedia, "Robot end effector." https://en.wikipedia.org/wiki/Robot_end_effector
2. Wikipedia, "Effector (biology)." https://en.wikipedia.org/wiki/Effector_(biology)
3. Alexander Fabisch, Wadhah Zai El Amri, Chandandeep Singh, and Nicolas Navarro-Guerrero, "Do Robots Really Need Anthropomorphic Hands? A Comparison of Human and Robotic Hands," German Research Center for Artificial Intelligence (DFKI) and Leibniz University Hannover, arXiv:2508.05415, 2025. https://arxiv.org/abs/2508.05415
4. Wevolver, "What are End Effectors? Types of End Effectors in Robotics and Applications." https://www.wevolver.com/article/end-effector
5. Gareth J. Monkman, Stefan Hesse, Ralf Steinmann, and Henrik Schunk, Robot Grippers (Wiley-VCH, 2007), sample chapter "Introduction to Prehension Technology." https://application.wiley-vch.de/books/sample/3527406190_c01.pdf
6. Schmalz, "Vacuum grippers: definition, types and applications." https://www.schmalz.com/en/support/know-how/glossary/vacuum-gripper
7. Goudsmit Magnetics, "Electronic magnetic grippers." https://www.goudsmitmagnetics.com/en-us/products/gripping-handling-and-moving/electronic-magnetic-grippers
8. "Electroadhesion Technologies for Robotics: A Comprehensive Review," ResearchGate. https://www.researchgate.net/publication/338283460_Electroadhesion_Technologies_for_Robotics_A_Comprehensive_Review
9. ATI Industrial Automation, "Automatic / Robotic Tool Changers." https://www.ati-ia.com/products/toolchanger/robot_tool_changer.aspx
10. John Koetsier, "Apptronik has a totally different approach to building humanoid robots." https://johnkoetsier.com/humanoid-robots/
11. "Integrated linkage-driven dexterous anthropomorphic robotic hand," Nature Communications, 2021. https://www.nature.com/articles/s41467-021-27261-0
12. Robotics Center (SVRC), "Robotiq 2F-85 Gripper" (specifications and pricing). https://www.roboticscenter.ai/en/hardware/robotiq-2f-85
13. Wonik Robotics, "Purchasing," Allegro Hand Wiki. http://wiki.wonikrobotics.com/AllegroHandWiki/index.php/Purchasing
14. Shadow Robot Company, "Shadow Dexterous Hand Technical Specification," 2022. https://www.shadowrobot.com/wp-content/uploads/2022/03/shadow_dexterous_hand_e_technical_specification.pdf
15. RobotWale, "The Grasp Reality: Analyzing Shadow, Allegro, and Inspire in the Race for Dexterous Hands." https://www.robotwale.com/article/dexterous-hands-shadow-allegro-inspire-race-to-5-finger-dexterity
16. CNX Software, "Sharpa Wave is a high-end dexterous robotic hand with 22 DoF, high-sensitivity dynamic tactile array," 2026. https://www.cnx-software.com/2026/06/02/sharpa-wave-high-end-dexterous-robotic-hand-with-22-dof-high-sensitivity-dynamic-tactile-array/
17. Teslarati, "Tesla Optimus V3 hand and arm details revealed in new patents," 2026. https://www.teslarati.com/tesla-optimus-v3-hand-arm-details-revealed-new-patents/
18. Bain Capital Ventures, "Home Robots Finally Arrive: Inside Sunday Robotics." https://baincapitalventures.com/insight/robots-finally-at-home/
19. Boston Dynamics, "Atlas' Evolution From Research Robot to Industrial Humanoid." https://bostondynamics.com/blog/atlas-evolution-from-research-robot-to-industrial-humanoid/
20. Thomasnet, "Boston Dynamics Upgrades Atlas Robot with High-Precision, 3-Fingered Hands." https://www.thomasnet.com/insights/boston-dynamics-atlas-3-fingered-hands/
21. Nicolas Correll et al., "Analysis and Observations From the First Amazon Picking Challenge," arXiv:1601.05484. https://arxiv.org/pdf/1601.05484
22. MIT News, "Giving robots a more nimble grasp," 2015. https://news.mit.edu/2015/giving-robots-more-nimble-grasp-0804
23. Cheng Chi et al., "Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots," Stanford University, Columbia University, and Toyota Research Institute. https://umi-gripper.github.io/
24. Humanoid.press, "Figure 03 Humanoid Robot by Figure AI." https://www.humanoid.press/database/database-figure-03-humanoid-robot/
25. OpenELAB, "Unitree Dex5-1 Tactile Hand Review and Specs." https://openelab.io/blogs/learn/unitree-dex5-1-tactile-hand-review-20-dof-robotic-manipulation
26. Humanoid.guide, "Xpeng IRON Specs and Price." https://humanoid.guide/product/iron/
27. Carsten Mehring et al., "Augmented manipulation ability in humans with six-fingered hands," Nature Communications, 2019. https://www.nature.com/articles/s41467-019-10306-w
28. "The human hand has 27 degrees of freedom (DOF)," ResearchGate figure. https://www.researchgate.net/figure/Finger-flexion-extension-Scene-The-human-hand-has-27-degrees-of-freedom-DOF-7-4-in_fig1_355898836
29. A.B. Vallbo and R.S. Johansson, "Tactile sensibility in the human hand: relative and absolute densities of four types of mechanoreceptive units in glabrous skin," Journal of Physiology, 1979, via PubMed. https://pubmed.ncbi.nlm.nih.gov/439026/
30. Emanuele Franchi et al., "Technical Report: Use of Hybrid Systems to model the RobotiQ Adaptive Gripper," University of Illinois. https://motion.cs.illinois.edu/papers/IUCS-TR711-Franchi-gripper.pdf
31. John Koetsier, "Human-Level Hands? 1X Just Gave Humanoid Robot Neo Something Close," Forbes, 2026. https://www.forbes.com/sites/johnkoetsier/2026/07/09/human-level-hands-1x-just-gave-humanoid-robot-neo-something-close/
32. Sanctuary AI, "Sanctuary AI Demonstrates In-Hand Manipulation Capabilities for Improved General Purpose Robot Dexterity." https://www.sanctuary.ai/blog/sanctuary-ai-demonstrates-in-hand-manipulation-capabilities-for-improved-general-purpose-robot-dexterity
33. Clone Robotics, "Hand." https://clonerobotics.com/hand/
34. "Development of a Novel Impedance-Controlled Quasi-Direct-Drive Robotic Hand," arXiv:2405.18730. https://arxiv.org/abs/2405.18730
35. ManufacturingTomorrow, "Humanoid Robot Developer Robot Era Debuts Dexterous, Versatile XHand for Object Manipulation," 2024. https://www.manufacturingtomorrow.com/news/2024/05/20/humanoid-robot-developer-robot-era-debuts-dexterous-versatile-xhand-for-object-manipulation/22783/
36. Humanoids Daily, "Experts Weigh In on Wuji Tech's Robotic Hand: A Remarkably Robust Direct-Drive Contender." https://www.humanoidsdaily.com/news/experts-weigh-in-on-wuji-techs-robotic-hand-a-remarkably-robust-direct-drive-contender
37. Interesting Engineering, "Chinese firm unveils Flex-2 arm to advance humanoid robots toward more versatile tasks." https://interestingengineering.com/ai-robotics/chinas-new-robotic-hand-combines-hybrid-actuation-for-smarter-robot-manipulation
38. CMRA, "China's Dexterous Hand Sector Produces Four Unicorns Amid Rapid Embodied Intelligence Expansion." https://cnmra.com/chinas-dexterous-hand-sector-produces-four-unicorns-amid-rapid-embodied-intelligence-expansion/
39. ZWHAND (Shenzhen ZWHAND Technology Co. / Zhaowei), official product site. https://www.zwhand.com/en
40. Standard Bots, "The best 5 pneumatic grippers for your Industrial robot." https://standardbots.com/blog/the-best-5-pneumatic-grippers-for-your-industrial-robot