16 Humanoid Robot Hands Compared: DOF, Sensing and Evidence
Sixteen humanoid robot hands compared by fingers, degrees of freedom, actuation, tactile sensing, force control, demonstrated tasks and availability in 2026.
Introduction
A five-finger silhouette says almost nothing about a robot hand’s useful capability. One hand may have 22 independently driven degrees of freedom and dense fingertip touch. Another may use five fingers coupled to six motors. A three-finger hand can outperform both in repetitive industrial picking because it is lighter, simpler and easier to control. Published specifications also use incompatible counting rules: some companies report joints, some active degrees of freedom, some total hand-and-wrist axes and some only the number of actuators.
This comparison covers sixteen hands built for humanoid robots, humanoid upper bodies or direct integration into humanoid research platforms. It excludes parallel grippers and purely cosmetic hands. The table records only values found in manufacturer pages, technical reports or laboratory documentation. Missing fields remain missing. The categories at the end are evidence-based groupings, not a universal ranking, because no common protocol measures fingertip force, in-hand manipulation, tactile resolution, durability and task success across all sixteen systems.
Key findings
- Sharpa Wave and Shadow publish the most detailed standalone commercial hand specifications in this set.
- Figure 03 and 1X NEO integrate vision or whole-body design decisions that cannot be judged from finger count alone.
- Three-finger systems such as Unitree Dex3-1 trade human-like morphology for lower mass and simpler control.
- Tactile sensing ranges from a few contact or pressure sensors to dense arrays with more than 1,000 pixels per fingertip.
- Several major humanoid companies do not publish enough hand data for a defensible dexterity ranking.
Sixteen humanoid and humanoid-platform hands
Active DOF and joint counts are kept separate where the source distinguishes them. “Undisclosed” is preferable to inferring mechanics from appearance.
| Hand / robot | Country | Fingers and DOF | Actuation and sensing | Documented task or status |
|---|---|---|---|---|
| 1X NEO hand | United States / Norway origin | Five fingers; main NEO page lists 22 DOF per hand, while the July 2026 dedicated hand release describes a newer 24/25-DOF design. | Tendon-driven, low-inertia, IP68 hand. Published joint torque and durability figures; cameras are in NEO’s head rather than palm. | Integrated into NEO. Home manipulation and fast finger motion shown. Early Access product. |
| Figure 03 hand | United States | Five fingers; active DOF count not publicly disclosed. | Compliant fingertips, custom tactile sensors detecting a claimed 3 g pressure and an embedded palm camera. | Dishwasher loading, bedroom reset and industrial part handling under Helix demonstrations. Not sold separately. |
| Unitree Dex3-1 | China | Three fingers, 7 active DOF: thumb 3, index 2, middle 2. | Force-position hybrid control; optional tactile sensor arrays. About 710 g on Unitree product documentation. | Optional on G1 EDU and other Unitree platforms. Commercial sales channel. |
| Unitree Dex5-1 | China | Five fingers; exact current DOF requires configuration confirmation. | Manufacturer presents it as a five-finger dexterous hand; public detailed specification is less complete than Dex3-1. | Offered in Unitree’s current hand catalog. Ranking withheld because sensing and force data are incomplete. |
| Sharpa Wave | Singapore / China manufacturing | Five fingers, 22 active DOF. | Joint-space actuation, Dynamic Tactile Array, over 1,000 tactile pixels per fingertip, claimed 0.02 N sensitivity and over 20 N fingertip force. | Shipping standalone hand; integrated into Sharpa North and NVIDIA’s H2-based reference humanoid. |
| Shadow Dexterous Hand | United Kingdom | Five fingers, 20 actuated DOF and 24 joints including wrist/palm coupling. | Tendon-driven, EtherCAT/ROS, more than 100 sensors; optional expanded fingertip tactile sensing. | Commercial research hand used on humanoid and arm platforms. Published 4 kg power-grasp rating in 2025 specification. |
| Fourier GR-2 dexterous hand | Singapore / China | Five fingers, 12 DOF per hand on the enhanced configuration. | Six array tactile sensors; force and material sensing described by Fourier. | Integrated into GR-2 for manipulation research and demonstrations. Quote-based platform. |
| Sanctuary Phoenix hand | Canada | Five fingers; Gen 6 described 20 total hand DOF, later hydraulic hand releases use different generations. | Hydraulic actuation, proprietary haptics and new tactile sensors. Generation must be stated. | In-hand manipulation and industrial task research. Current standalone availability not public. |
| iCub hand | Italy | Five fingers; 9 motors and 19 joints in the hand, plus a 3-DOF wrist in catalog documentation. | Tendon-driven; tactile skin on palm and fingertips, encoders and force/torque sensing in the platform. | Long-running open research platform for grasping, learning and human-robot interaction. |
| IIT R1 hand | Italy | Five fingers, 4 DOF hand. | Series-elastic actuation and distributed pressure sensing; designed for safe service manipulation. | Integrated into R1 research humanoid. Not a high-DOF in-hand manipulation system. |
| DLR Hand II | Germany | Four fingers, three DOF per finger plus one reconfigurable palm DOF. | Integrated motors, joint torque, position, motor-position and temperature sensors. | Used on Rollin’ Justin and Agile Justin for bimanual research. Historical research hardware. |
| DLR-HIT Hand II | Germany / China | Five modular fingers, 15 DOF and four joints per finger with distal coupling. | Integrated brushless motors, harmonic drives, position, force/torque and temperature sensing. | Used on SpaceJustin with shared autonomy. Research platform rather than current mass-market product. |
| Toyota T-HR3 hand | Japan | Five-finger humanoid hand; detailed active hand DOF not separately published in the reviewed corporate material. | Torque Servo modules and master maneuvering system for remote operation. | Teleoperated fine hand and arm movements in partner-robot research. Not commercially offered. |
| UBTECH Walker S2 hand | China | Five-finger industrial dexterous hand; current public product page does not provide a clean hand-only DOF table. | UBTECH reports successive fourth- and fifth-generation dexterous hand development; sensing detail varies by corporate report. | Integrated into mass-produced Walker S2 factory platform. Information insufficient for rank. |
| Apptronik Apollo hand | United States | Five-finger appearance; current public active DOF and tactile specification not disclosed. | Apptronik describes dexterous manipulation but does not publish a complete hand data sheet. | Industrial pilot platform. Information insufficient for hand ranking. |
| AgiBot G-series dexterous hand | China | Multi-finger dexterous hand; model-specific public DOF and sensor fields are incomplete on English product pages. | Integrated actuation and embodied-AI manipulation stack; generation-specific technical sheet not located. | Used across AgiBot full-size and industrial platforms. Information insufficient for rank. |
Degrees of freedom are useful only when the counting method is clear
A human hand contains many anatomical joints, but robotic designs often mechanically couple distal joints or drive several joints from one tendon. A specification may therefore list 24 joints and 20 controllable degrees of freedom. Another company may add two wrist axes to the hand total. The comparison should preserve the manufacturer’s terms rather than normalize them into a false single number.
Independent actuation supports more hand shapes and in-hand manipulation, but increases motor count, wiring, calibration and control complexity. Coupled fingers can produce robust enveloping grasps with fewer actuators. Industrial reliability may favor fewer degrees of freedom when the task is carrying a tote, feeding a part or pressing a button.
Three-finger hands deserve separate consideration. Unitree Dex3-1 does not imitate human anatomy, yet its seven active axes and optional tactile arrays can cover pinch, tripod and power grasps at much lower mass than many five-finger systems. It should not be mislabeled as a five-finger hand or compared on appearance.
Touch sensing is not one feature
A pressure switch can confirm contact. A strain gauge can estimate joint torque. A force/torque sensor can measure the wrench at a wrist or fingertip. Dense tactile arrays can localize pressure across a surface and detect slip patterns. Companies frequently use the word tactile for all of these systems, so the sensor type and location matter.
Figure publishes a clear threshold claim for its fingertip sensor and adds palm cameras to preserve close-range vision when the head view is occluded. Sharpa publishes tactile pixel density and pressure sensitivity. Shadow publishes a broad sensor count and optional fingertip configurations. iCub uses distributed skin as part of a research platform. Those disclosures are more useful than a generic statement that a hand can feel.
Tactile sensing also needs software. Raw pressure maps do not automatically produce stable grasps. A policy must detect contact, estimate slip, regulate force and adapt as an object deforms. Few manufacturers publish task-level ablations showing how much tactile input improves success over vision alone.
Actuation changes strength, speed and safety
Tendon-driven hands place motors away from the fingers and can reduce distal mass. Tendons also introduce stretch, friction, routing complexity and maintenance. Direct-drive or geared finger modules simplify some control paths but add mass inside the hand. Hydraulic designs can deliver high force and speed, with added pumps, valves and sealing requirements.
Compliance can be mechanical, software-controlled or both. A soft fingertip increases contact area. A backdrivable actuator yields when pushed. Force control limits contact load. These properties matter in a home or shared workspace more than a maximum grip-force figure measured in a power grasp.
Durability is rarely tested under the same conditions. Sharpa publishes cycle and shock tests. Shadow publishes payload and typical movement frequency. 1X publishes nominal and peak load-cycle claims. Those numbers cannot be ranked without knowing duty cycle, maintenance, failure definition and whether the test covered the complete hand or one module.
Evidence-based categories instead of a false overall ranking
Most human-like morphology
Sharpa Wave, Shadow Dexterous Hand, 1X NEO and DLR-HIT Hand II publish high-DOF anthropomorphic architectures. “Human-like” refers to kinematics and scale, not equal human dexterity.
Strongest documented sensing
Sharpa Wave, Figure 03, Shadow and iCub provide the clearest public detail about tactile or distributed sensing. Their sensor modalities and resolutions differ, so this is a documentation category rather than a performance winner.
Most accessible for research
Shadow sells a standalone research system, Sharpa accepts orders and Unitree offers Dex3-1 with supported humanoid configurations. iCub is an established institutional platform. Price and delivery require current quotations except where a store lists them.
Simplest industrial design
Unitree Dex3-1 and lower-DOF hands such as IIT R1 reduce actuation complexity. Their suitability depends on the object set and cannot be inferred from DOF alone.
Best documented dexterity
Shadow has extensive technical specifications and research history. Sharpa publishes current hardware, tactile and durability figures. Figure publishes demanding integrated tasks but not a complete hand data sheet. No shared benchmark supports one absolute winner.
Information insufficient for ranking
Apptronik Apollo, UBTECH Walker S2, AgiBot’s current hands and Unitree Dex5-1 lack enough consistent public hand-only data for a defensible cross-platform rank.
Limitations and missing information
- Manufacturers count joints, motors, active DOF and wrist axes differently.
- Several hands have changed between robot generations; Sanctuary, 1X and UBTECH specifications must be tied to a date and generation.
- Grip force, fingertip force and arm payload are different measurements and were not merged.
- Company demos use different objects, speeds, training data and success criteria.
- Some research hands are commercially orderable components; others exist only as parts of an institutional robot platform.
Conclusion
The best hand depends on the task and on how much evidence the buyer needs. Shadow and Sharpa provide unusually detailed standalone specifications and direct research integration paths. Figure 03 and 1X NEO show how a hand can be designed with the whole robot: cameras, soft contact, tendon routing, safety and learned control are part of the manipulation system. Unitree Dex3-1 demonstrates why three fingers can be a rational engineering choice when weight, cost and control simplicity matter.
The largest gap is not finger count. It is comparable testing. Buyers need repeated trials on the same object sets, force limits, slip tests, in-hand rotation, cable insertion, deformable materials, impact cycles and recovery from failed grasps. Until those protocols exist, the most honest comparison separates architecture, sensing, documentation and availability instead of declaring a universal dexterity champion.
Frequently asked questions
Which humanoid robot has the most dexterous hand?
There is no verified universal winner. Sharpa Wave and Shadow publish high-DOF architectures and detailed sensing information. Figure and 1X show integrated whole-robot manipulation. Dexterity also depends on control policies, calibration, wrist motion, tactile feedback and the task, so finger count alone cannot answer the question.
How many degrees of freedom does a robotic hand need?
A narrow task may need only one or two closing axes. A three-finger hand with seven active DOF can perform many useful grasps. High-DOF five-finger hands are valuable for human tools and in-hand manipulation, but they demand more data, control bandwidth and maintenance.
Do humanoid robot hands have touch sensors?
Some do. Figure, Sharpa, Shadow, iCub, Sanctuary and optional Unitree configurations publish tactile, pressure or force sensing. The word tactile covers very different hardware, from sparse contact measurements to dense fingertip arrays, so the sensor type and location should always be stated.
Can I buy a humanoid robot hand separately?
Shadow Dexterous Hand and Sharpa Wave have commercial inquiry or order channels. Unitree offers hands as options for its platforms. Other hands, including Figure 03 and 1X NEO, are integrated products and are not advertised as standalone components. Institutional platforms such as iCub require a quotation.
Is a five-finger hand always better than a three-finger hand?
No. Five fingers provide more human-like contact patterns, but add mass, cost and control complexity. A three-finger design can be faster to calibrate, easier to protect and sufficient for boxes, tools and many industrial parts. The object set and required failure rate should decide.
Why do robot hand DOF figures conflict between sources?
Sources may count active axes, passive coupled joints, wrist axes or total mechanical joints. Products also change between generations. A comparison should quote the manufacturer’s exact term and date rather than silently converting every number into a single DOF field.
Sources and methodology
A hand qualified when it was developed for a full humanoid, humanoid upper body or documented humanoid integration. Parallel industrial grippers were excluded. The comparison uses official product pages, manufacturer technical documents and laboratory pages. No missing specification was estimated from photographs or third-party databases.
- NEO hands — 1X Technologies · July 9, 2026
- Introducing Figure 03 — Figure AI · October 9, 2025
- Unitree G1 and Dex3-1 — Unitree Robotics · Accessed July 11, 2026
- Unitree Dex3-1 — Unitree Robotics · Accessed July 11, 2026
- Sharpa Wave — Sharpa · Accessed July 11, 2026
- NVIDIA GR00T reference humanoid — NVIDIA · May 31, 2026
- Shadow Dexterous Hand series — Shadow Robot Company · Accessed July 11, 2026
- Shadow Dexterous Hand technical specification — Shadow Robot Company · 2025
- Fourier GR-2 — Fourier · Accessed July 11, 2026
- Sanctuary Phoenix — Sanctuary AI · May 16, 2023
- iCub product catalog — Italian Institute of Technology · Accessed July 11, 2026
- IIT R1 robot — Italian Institute of Technology · Accessed July 11, 2026
- DLR Hand II — German Aerospace Center · Accessed July 11, 2026
- DLR-HIT Hand II — German Aerospace Center · Accessed July 11, 2026
- Toyota T-HR3 — Toyota Motor Corporation · November 21, 2017
- UBTECH Walker S2 — UBTECH Robotics · Accessed July 11, 2026
- Apptronik Apollo — Apptronik · Accessed July 11, 2026
- AgiBot products — AgiBot · Accessed July 11, 2026
Related TechniaHQ guides
Official image recommendations
- A verified 4×4 grid using one official photograph for each named hand.
Sixteen labeled humanoid robot hands shown in a four-by-four comparison grid — Official manufacturers and research laboratories listed in the source section - Sharpa Wave palm and fingertip tactile close-up.
Sharpa Wave five-finger robotic hand showing its tactile fingertip surfaces — Sharpa - Figure 03 hand with palm camera and compliant fingertip detail.
Figure 03 robotic hand with embedded palm camera and soft tactile fingertips — Figure AI - Unitree Dex3-1 three-finger architecture.
Unitree Dex3-1 force-controlled three-finger robotic hand — Unitree Robotics - Hand architecture map: tendon, geared, hydraulic and coupled designs.
Taxonomy of humanoid robot hand actuation and tactile sensing designs — TechniaHQ original graphic
Fact-check report
Verified: July 11, 2026
Confirmed
- Sixteen distinct hand systems are listed; no duplicate branding of the same hand is counted twice.
- Parallel grippers are excluded.
- Generation conflicts are explicitly identified for 1X and Sanctuary.
Not confirmed or incomplete
- Current hand-only DOF and tactile specifications remain incomplete for Apptronik, UBTECH, AgiBot and Unitree Dex5-1.
- No independent common dexterity benchmark covers all sixteen hands.
Fast-changing information
- Hand configurations, optional tactile packages and order availability can change with robot revisions.