Humanoid Robot Hands: How They Work, Compared

How robot hands work

Every robot hand does three jobs: move the fingers, sense what they touch, and decide what to do next. (For the rest of the robot, see what is a humanoid robot.) The differences between hands come down to how each job is solved.

The drive: tendon, linkage, direct drive or hydraulic

Tendon drive. Motors sit in the forearm and pull thin cables that run through the wrist into the fingers, much like human forearm muscles and tendons. The fingers stay light and slim, and the hand can be backdrivable (it gives when pushed). The cost is complexity: cables stretch, rub and need tensioning. The Shadow Dexterous Hand has used tendons for years, 1X uses its own tendon drive in the NEO hand with low gear ratios of about 5:1 to 15:1, and Tesla patent filings describe a tendon-driven Optimus hand with actuators moved into the forearm (Teslarati).

Linkage drive. Small linear actuators push rigid rods that curl each finger. Joints are coupled, so one motor moves two or three joints along a fixed path. Linkage hands are strong for their weight, cheaper and easier to maintain, but they cannot place each joint independently. Inspire’s RH56DFX uses six miniature linear servo actuators to drive 12 joints and weighs 540 g (Robotics Center). This style is common on Chinese humanoids and research platforms.

Direct drive. Tiny motors sit in the palm or finger segments and drive the joints with little or no transmission. Control is precise and response is fast, but motors inside the fingers add mass at the end of the arm. RobotEra’s XHAND1, used on the RobotEra L7, is described as fully direct-driven with 12 active DoF.

Hydraulic. Sanctuary AI uses miniaturized hydraulic valves in the Phoenix hand. The company claims an order of magnitude more power density than cable and electromechanical systems and says the valves passed two billion cycles without leaking (Sanctuary AI). Nobody else in our database uses hydraulics in the hand.

Degrees of freedom: active, passive and coupled

“Degrees of freedom” (DoF) counts independent ways a hand can move. Read spec sheets carefully, because vendors count differently:

• Active (actuated) DoF have their own motor. This is the number that matters most.
• Passive DoF move but are not driven, usually springs or couplings that let a finger conform to an object.
• Total joints counts every hinge, driven or not.

Unitree’s Dex5-1 is a good example: 20 DoF in total, of which 16 are active and 4 passive (Robotics 24/7). The Shadow Hand has 20 actuated DoF and 24 joints. The Inspire RH56DFX has 6 actuated DoF across 12 joints. A human hand is usually described as having about 27 DoF including the wrist.

Tactile sensing

Position sensors tell the hand where its fingers are. Tactile sensors tell it what it is touching, how hard and whether the object is slipping. Figure says each Figure 03 fingertip sensor detects forces as small as three grams and that each palm carries a wide-angle camera for close-up grasping (Figure AI). 1X says NEO’s fingertips measure normal force, contact location and shear, and detect slip. Sharpa’s Wave hand uses what it calls a Dynamic Tactile Array on each finger (CNX Software). Cheaper hands often ship with no tactile sensors at all, or sell them as an upgrade.

Weight, materials and protection

Every gram at the fingertips costs payload and battery at the shoulder, so hand weight is a real spec. The full Shadow Hand weighs 4.3 kg, while its 7-motor Super Lite version weighs 1.8 kg (Shadow Robot). The Inspire RH56DFX weighs 540 g and the Sharpa Wave 1.3 kg. Moving motors into the forearm, as 1X and Tesla do, is mostly a way to keep the hand itself light.

Makers rarely publish materials in detail, but fingertip compliance is a recurring theme. Figure says Figure 03’s fingertips were made softer and given more contact area for stable grasps. Home robots add protection: 1X rates the NEO hand IP68 and says it uses food-safe materials, which matters for a robot that loads dishwashers.

Humanoid robot hands compared

The table lists what makers have published. Blank fields mean no public number; we do not fill gaps with guesses.

Robot or hand	Fingers	Active DoF per hand	Drive	Sensing	Force or payload
1X NEO	5	22 (plus 3 at wrist)	Tendon, forearm motors	Fingertip normal force, contact location, shear	Up to 45 N distal flexion force; IP68
Figure 03	5	Not published	Not published	Fingertip sensors to about 3 g, palm cameras	Robot payload 20 kg
Figure 02	5	16	Not published	Not published	“Human-equivalent strength,” per Figure
Tesla Optimus Gen 2	5	11	Not published	Tactile fingers reported	Not published
Tesla Optimus (next-gen hand, patents)	5	22 incl. 2 at wrist, per filings	Tendon, forearm actuators	Not published	Not published
Sanctuary Phoenix	5	20 (2023 launch) / 21 (later releases)	Hydraulic	Haptic and tactile sensing claimed	Not published
Unitree G1 with Dex3-1	3	7	Electric	Optional tactile arrays	Not published
Unitree Dex5-1	5	16 (20 total)	Electric	94 sensors on Dex5-1P	3.5-4.5 kg grasp; 10 N fingertip
Kepler K2	5	11	Electric	25 force points per finger, 6-axis wrist sensor	Not published
Fourier GR-2	5	12	Electric	Six array tactile sensors per hand	Not published
PUDU D9 with DH11	5	11 (6 motors)	Electric	1,018 tactile pixels in 12 areas	30 N grip
Humanoid HMND 01	5 or gripper	12 (or 1-DoF gripper)	Electric	Not published	Not published
Boston Dynamics Atlas	Gripper	2	Electric	Tactile sensing	14 kg grip
Apptronik Apollo	Not published	Not published	Not published	Not published	Robot payload 55 lb
Shadow Dexterous Hand	5	20 (24 joints)	Tendon, 20 DC motors	Over 100 sensors at up to 1 kHz	4.3 kg hand weight
Inspire RH56DFX	5	6 (12 joints)	Linkage, linear servos	None on DFX	15 N thumb, 10 N finger
Sharpa Wave	5	22	Electric	Dynamic tactile array	20 N fingertip; 1.3 kg

Sources: 1X, Figure (Figure 02 launch), Ars Technica for Optimus Gen 2, Sanctuary AI’s Phoenix launch, Unitree, The Robot Report for Kepler, Interesting Engineering for Fourier, the PUDU D9 flyer, Humanoid, the Atlas spec sheet, and the hand makers’ pages cited above. Sanctuary’s own materials give 20 DoF at the 2023 launch and 21 DoF in later announcements; we show both.

Two things stand out. First, the industrial robots with the most real deployment hours, Agility Digit and Atlas, use simple end effectors, not five-finger hands. Second, the hands with the best published specs belong to robots that are not yet in broad paid service. For a ranking that weighs demonstrated manipulation rather than specs, see our most dexterous humanoid robots list.

Why hands are the hardest and costliest part

Musk has called the hand “the majority of the engineering difficulty of the entire robot” (Teslarati). That matches what you see across the industry, for four reasons.

1. Packaging. Shadow Robot puts it simply: each finger works “like a small robot arm” squeezed into a much smaller space (Shadow Robot). Twenty motors, gearboxes, sensors and cables have to fit inside something the size of your hand.
2. Durability. Hands hit things. Fingers jam, tendons fray and sensors take impacts thousands of times a day. 1X says it tests its hands for millions of cycles and that its wrist joints have passed 2 million cycles.
3. Sensing. A hand without touch has to rely on cameras, which are blocked by the hand itself at the moment of grasp. That is why Figure added palm cameras.
4. Control and data. A 22-DoF hand has a huge space of possible movements. Learning to use it requires many hours of demonstrations; Figure trained Helix on about 500 hours of teleoperated data across the whole upper body (Figure AI).

Cost follows complexity. Shadow priced its full hand at €110,000 in 2022 and, citing the International Federation of Robotics, noted that end-of-arm tooling can triple the price of a robot arm. You can see the same effect in humanoid prices. On RoboStore, a Unitree G1 EDU Standard lists at $43,900 while EDU Ultimate versions, which differ in hand type and compute, run $65,900 to $73,900. The base G1 at $13,500 has no dexterous hands at all.

When a simple gripper is the better choice

It is tempting to rank hands by DoF and stop there. In paid work, the opposite pattern shows up. Atlas ships with a 2-DoF gripper rated at 14 kg of grip, per its spec sheet, and Agility built Digit’s end effectors around totes. 1X’s older EVE robot uses 1-DoF grippers. HMND 01 from Humanoid offers buyers either a 12-DoF hand or a 1-DoF parallel gripper.

Grippers win when the object set is known and repetitive: totes, bins, boxes, car parts with designed grasp points. They are cheaper, stronger for their size, easier to keep running and far easier for a control policy to learn. Five-finger hands earn their cost only when the robot must handle many unknown objects, use human tools, or work in homes where nothing is standardized. If you are scoping a deployment, start by listing the objects the robot will touch; if a handful of known shapes covers the job, a gripper is probably enough.

Which tasks need which hand

Task	Hand you actually need	Why
Moving totes and bins	1-2 DoF gripper or purpose-built end effector	Agility Digit does paid tote work with custom end effectors
Picking boxes and parcels	Gripper or suction, sometimes a simple 5-finger hand	Shape varies, but grip is mostly power grasp
Loading parts into fixtures	6-12 DoF hand or a tactile gripper	Needs repeatable positioning more than finger dexterity
Laundry, dishes, tidying	12-22 DoF with tactile fingertips	Soft, deformable and fragile objects need force control
Tool use (drills, screwdrivers)	16+ DoF with an opposable thumb and wrist strength	Holding and triggering a tool at once
In-hand manipulation (turning a key, rotating a pen)	20+ DoF with finger abduction and fast tactile feedback	Sanctuary and Shadow target this; few humanoids show it reliably

If you are buying a humanoid for research, the hand choice often matters more than the robot. Several research platforms, including the Unitree G1 EDU and Booster T2, are sold with a choice of hands; see our research humanoids list.

Buying a standalone dexterous hand

Prices below are what vendors or credible reports state. Most makers sell on quote, so treat these as anchors.

Hand	Maker country	Active DoF	Price	Price source
Shadow Hand Plus	UK	20	€110,000 incl. installation, training, support	Shadow Robot (2022)
Shadow 2-finger system	UK	7	€45,000	Shadow Robot (2022)
Sharpa Wave	China (Shanghai)	22	About $50,000, reported	CNX Software, citing a 36kr interview
Unitree Dex5-1	China	16	$15,000	Humanoid.guide listing
Inspire RH56DFX	China	6	$10,500	Robotics Center store
Inspire RH56E2 (heavy load, optional tactile)	China	6	$10,530	Robotics Center store
Linker Hand O6 Lite	China	Not stated	About 3,900 yuan ($550), reported	CMRA

The spread is enormous, and DoF explains only part of it. The Shadow price includes installation, training and support, and the hand is built in small numbers for labs. Chinese linkage hands are built at volume for humanoid makers. CMRA reports that AgiBot priced its OmniHand around $2,000 on the logic that a hand cannot cost more than the body it sits on; we could not confirm that figure with AgiBot.

Checklist before you order:

• Left, right or a pair? Most hands are sold one side at a time.
• Interface: CAN, RS-485, EtherCAT or Ethernet. Match your robot’s controller.
• Mounting: does it bolt to your arm’s flange, or do you need an adapter?
• Tactile sensing: included, optional or unavailable?
• Simulation model: is there a URDF or MJCF file for training in sim?
• Repair: can a single finger or tendon be replaced on site, or does the whole hand go back to the factory?
• Lead time and support region. The FCC’s July 2026 rule restricts new foreign-made humanoid robots in the US; read our China vs US guide before importing Chinese robot hardware.

Research directions to watch

• Forearm tendon designs. Tesla and 1X both moved hand motors into the forearm. Expect more hands to follow, since it cuts mass at the fingertips.
• Vision-based touch. Palm cameras (Figure) and camera-backed tactile fingertips (Sharpa) give the AI model close-range views exactly when the head cameras are blocked.
• Artificial muscles. Clone Robotics is building hands with its Myofiber artificial muscles and claims human-level finger strength (Clone Robotics). The Clone Alpha is still a prototype.
• Hydraulics. Sanctuary’s bet is that hydraulic valves beat electric motors on power density at hand scale.
• Cheaper hands at volume. Chinese suppliers are pushing six-actuator hands toward commodity prices, which makes five-finger hands viable on low-cost humanoids.
• Learning from human video. Training hands from footage of people working, instead of only teleoperation, is the main hope for closing the data gap.

To see how these hands perform on whole robots, compare models on the leaderboard, or line two up side by side, for example 1X NEO vs Figure 03. For prices of the robots themselves, see humanoid robot prices.