Dexterous hand

A dexterous hand (also called a dexterous robotic hand or multi-fingered robot hand) is an anthropomorphic end effector that reproduces the grasping, in-hand manipulation, and tactile perception of the human hand, using many independently controllable joints packed into a hand-sized volume. It sits at the most difficult end of a spectrum of robot end effectors that runs from simple parallel-jaw grippers, through adaptive multi-finger grippers, to fully articulated five-finger hands with more than twenty degrees of freedom. The dexterous hand is widely regarded as the single hardest subsystem to build in a general-purpose robot, and after decades as a niche research device it became, in the 2020s, a mass-production race driven by the humanoid and embodied-AI boom.^[1][2][3]

This article is the general concept page spanning research hands, prosthetics, teleoperation rigs, industrial hands, and humanoid-integrated hands. For a catalog of hands fitted to specific humanoid robots, with a large per-robot comparison table, see Humanoid robot hands.

What is a dexterous hand?

A dexterous hand is defined less by its appearance than by what it can do: reposition and reorient an object within the grasp (in-hand manipulation), form many distinct grasp types, and modulate contact forces finely enough to hold a fragile object without crushing it. This distinguishes it from two simpler categories. A generic end effector is any tool mounted at a robot wrist, including welders, suction cups, and grippers. A gripper is a low-degree-of-freedom end effector, usually with two or three fingers and one or two actuators, that closes on an object to hold it but cannot reconfigure it. A dexterous hand adds enough actuated joints, and enough sensing, to manipulate rather than merely hold.^[1][4]

The benchmark is the human hand, which is usually modeled as having about 27 degrees of freedom (DOF): roughly 21 across the four fingers and thumb, plus 6 for the position and orientation of the wrist and palm. The thumb alone contributes about 5 DOF and is disproportionately important for opposition grasps.^[5][6] No robot hand yet matches the human hand across dexterity, sensing density, weight, and durability at once, which is why the human hand remains the reference point against which every design is measured.

How is a dexterous hand measured?

Engineers summarize a hand's capability with a handful of numbers. The most quoted is the count of degrees of freedom, but DOF alone is misleading because it does not distinguish actuated (independently driven) joints from passive or coupled ones, and vendor figures often mix the two. Other key metrics include the number of fingers, tactile density (the number of touch-sensing points, or taxels, and their force resolution), grip force and payload, positional repeatability, response time, and durability measured in open-close cycles.^[1][7]

Industry teardowns of humanoid-grade hands as of 2026 describe a rough envelope of roughly 12 to 22 active DOF, payloads of a few kilograms, positional repeatability on the order of 0.02 to 0.1 mm, and control response times in the low milliseconds. These figures should be read as a marketing-derived range rather than a standard: the best-documented recent products cluster there, but the numbers vary widely by generation and are not independently audited.^[3][8] Two examples illustrate the sensing frontier: Sanctuary AI's Phoenix hand senses forces as low as 5 millinewtons, close to the roughly 3-millinewton sensitivity of a human fingertip, while Figure's Figure 03 hand advertises fingertip sensors that can feel a 3-gram load, about the weight of a paperclip.^[9][10]

A short history of the dexterous hand

Multi-fingered mechanical hands predate modern robotics, but the research lineage that leads to today's hands began in the 1960s and matured through a series of landmark academic platforms.

Scroll sideways for more →

The 1980s foundational hands, the Stanford/JPL hand designed by Kenneth Salisbury and the Utah/MIT hand led by Stephen Jacobsen, established dexterous, force-controlled manipulation as a research field distinct from simple grasping.^[11][12] Through the 1990s and 2000s the emphasis shifted to integration and sensing: NASA's Robonaut hands were built to work in an astronaut's task space, and Robonaut 2, developed with General Motors, became the first humanoid robot in space when it reached the International Space Station in 2011.^[13] Germany's DLR produced a series of hands (DLR Hand I and II, and the DLR/HIT hands built with Harbin Institute of Technology) that packed all motors and electronics inside the hand and forearm, a design language visible in many Chinese commercial hands today.^[14]

The modern inflection came in 2018, when OpenAI's Dactyl project used a Shadow Dexterous Hand to learn in-hand manipulation through reinforcement learning, bridging the classic mechanical hand to learning-based control (see below).^[15] From about 2021 onward the general-purpose humanoid boom, led by Tesla Optimus, Figure, Sanctuary, 1X, and a wave of Chinese firms, turned the dexterous hand from a laboratory curiosity into a component with its own supply chain and mass-production roadmap.^[2][3]

How does a dexterous hand work? The technology stack

A high-DOF hand is a dense integration of miniature mechanical, sensing, and electronic subsystems. The table below lists the main building blocks and why each matters. The supplier names are drawn from published Chinese industry-chain maps and should be read as representative participants, not exclusive or sole suppliers; several are diversified firms for which robotics is an emerging rather than a core business.^[3][16][17]

Scroll sideways for more →

Two of these subsystems dominate the design conversation. Tactile sensing is the newest and, many argue, the most important: a hand that cannot feel is effectively blind at the moment of contact. Fingertip sensor arrays now reach very high taxel counts; PaXini's DexH13 four-finger hand, for example, covers its fingers and palm with about 1,140 intelligent tactile processing units and pairs them with an integrated camera.^[18] See Tactile sensing for the underlying technologies.

The tendon or cable-driven transmission is the other defining feature of many high-DOF hands. By placing the motors in the forearm and routing thin cables (tendons) to the finger joints, designers keep the fingers slim and light while allowing many actuated joints, mimicking how human forearm muscles pull tendons that move the fingers. Tesla's Optimus Gen 3 hand, revealed in early 2026, uses exactly this pattern, relocating its actuators into the forearm and reaching 22 DOF per hand with 25 actuators per forearm-and-hand assembly, roughly a 4.5-fold increase over the previous generation.^[8] The trade-off is durability: cables stretch and abrade, and traditional tendon hands can fail after only about 10,000 grasp cycles, which is why suppliers now publish cycle-life targets. Xynova, a Hangzhou startup that builds the entire actuator stack in-house, reports validating its tendon transmission past one million open-close cycles at rated load and its Flex 2 components past two million cycles, and Xiaomi's CyberOne hand program cites a 150,000-cycle target as a 15-fold improvement over typical designs.^[16][19] See Tendon-driven for the mechanism in detail.

The gear reduction that lets a tiny motor hold a finger against load usually comes from a micro harmonic drive, a strain-wave gear that delivers high reduction ratios (commonly in the tens to low hundreds to one) with near-zero backlash in a compact package. Harmonic reducers are a recognized supply-chain bottleneck for the whole robotics industry; fewer than five firms worldwide make them to the highest precision grades, with Japan's Harmonic Drive Systems the incumbent and China's LeaderDrive and Laifual scaling rapidly.^[20]

Actuation approaches compared

How a hand moves its joints is the central design decision, and the field has settled into several competing approaches, each with clear trade-offs.^[3][7]

Scroll sideways for more →

Most modern hands are tendon-driven or a hybrid of tendon and direct drive. Two outliers pursue fluid power. Sanctuary AI's Phoenix uses hydraulic actuation with coin-sized valves the company says are 50 times faster and 6 times cheaper than off-the-shelf parts, claiming an order of magnitude higher power density than cable or electromechanical systems and reporting that its valve actuators survived over 2 billion cycles without leakage; its 21-DOF hydraulic hand can reorient an object in-hand even under a sudden 500-gram load disturbance.^[9][21] Clone Robotics, a Polish company, goes further with water-powered Myofiber artificial muscles that contract when hydraulic fluid is pumped through them; its Alpha design wraps a polymer skeleton of 206 artificial bones in these muscles, driven by a 500-watt pump acting as an artificial heart, with muscles that contract about 30 percent in under 50 milliseconds.^[22]

A widely cited framing from the components maker AAC Technologies calls this design space an "impossible trinity" of degrees of freedom, size, and force output: any two can be maximized, but not all three at once. Tendon hands buy small size and high DOF at the cost of load capacity; linkage hands buy load and reliability at the cost of DOF and compliance. This constraint explains why suppliers often ship two product lines, one optimized for raw force and DOF and one for sensing and precision.^[3][16]

Sensing and the role of touch

Vision tells a robot where an object is; touch tells it what happens at contact. Rich sensing is what separates a hand that can grasp from a hand that can manipulate, because contact forces, slip, and object compliance are largely invisible to cameras once the fingers close around an object.^[1][9]

A dexterous hand typically fuses several sensing modes. Proprioception (joint position and torque from encoders and current sensing) tells the controller where the fingers are and how hard they push. Force/torque sensors, often six-axis units at the wrist or in the joints, measure interaction forces. Tactile arrays on the fingertips and palm sense pressure distribution, shear, texture, and slip. The tactile arrays themselves come in several physical types: resistive and piezoresistive skins, capacitive arrays, magnetic (Hall-effect) sensors that read the deformation of a magnet-loaded elastomer, barometric sensors, and vision-based (optical) fingertips such as the GelSight family, which put a small camera behind a soft gel to recover a high-resolution height map of the contact surface.^[23][24] Chinese supplier PaXini, spun out of tactile-sensing research at Waseda University, drove the cost of a multidimensional tactile sensor down from about 100,000 yuan to as little as 199 yuan, which is what makes covering a whole hand (or body) in touch sensors economically feasible.^[18] Some designs, including Figure's, add a camera in the palm or wrist for close-range visual feedback during grasping. For a fuller treatment see Tactile sensing.

AI and control: teaching a hand to manipulate

Controlling roughly twenty coupled joints in real time, under uncertain contact, is a problem that classical robotics never fully solved and that modern machine learning has transformed.

Classical control

At the lowest level, hands use position control (commanding joint angles for a preset grasp), force control (regulating contact force to handle fragile objects), and hybrid force-position control, which splits the task into directions that are position-controlled and directions that are force-controlled. A related family, impedance and admittance control, shapes the relationship between motion and force so the hand behaves compliantly rather than switching modes. These methods remain the foundation on top of which learned policies run.^[25]

Reinforcement learning and sim-to-real: the Dactyl landmark

The landmark demonstration that machine learning could master a real dexterous hand was OpenAI's Dactyl. In 2018, OpenAI trained a policy to reorient a block in a Shadow Dexterous Hand entirely in simulation, using reinforcement learning with heavy domain randomization (randomizing friction, object appearance, and dynamics), then transferred it zero-shot to the physical hand, an approach called sim-to-real. Learning the task took roughly 100 years of simulated experience compressed into about 50 wall-clock hours on a cluster of 6,144 CPU cores and 8 GPUs, and the resulting policy chained a median of 13 (and up to 50) successful reorientations on the real hand, discovering human-like finger-gaiting strategies with no human demonstrations.^[15][26] In October 2019 OpenAI extended the system to solve a Rubik's Cube one-handed, adding Automatic Domain Randomization, which grows the difficulty of the simulator as the policy improves. Notably, the cube-solving move sequence came from a classical algorithm; the hard, learned part was the physical manipulation, which succeeded about 60 percent of the time on easy scrambles and about 20 percent on the hardest.^[27][28] This line continued in academic work on in-hand reorientation, including UC Berkeley's proprioception-only rotation via rapid motor adaptation (2022) and touch-only in-hand rotation from UC San Diego (2023).^[29][30]

Imitation learning and teleoperation

The complementary paradigm is imitation learning: rather than discover behavior through trial and error, a hand learns from human demonstrations. The demonstrations are usually collected by teleoperation, in which a human pilots the robot hand through gloves, motion capture, or vision-based hand retargeting. Systems such as NVIDIA's DexPilot (2020) and the vision-based AnyTeleop (2023) let an operator drive a high-DOF hand by simply moving their bare hand in front of cameras, generating the contact-rich datasets that manipulation policies are trained on. Dense tactile feedback lets teleoperators perform touch-driven tasks and simultaneously produces higher-quality training data.^[31][32]

Vision-language-action models

The current frontier folds manipulation into vision-language-action models (VLAs), a class of foundation model that maps camera images and natural-language instructions directly to robot actions. Google's RT-2 (2023) showed that a web-pretrained vision-language model could be fine-tuned to output robot actions; Physical Intelligence's pi-0 (2024) added a flow-matching action expert that emits continuous actions at up to 50 Hz for dexterous, multi-stage tasks; and Figure's Helix (2025) used a dual-system design to control a 35-DOF humanoid upper body, including individual fingers, at high frequency. These models tie the dexterous hand into the broader programs of embodied AI, physical AI, and robot manipulation, where the hand is the point at which learned intelligence meets the physical world.^[33][34][35]

Notable dexterous hands and their makers

The table below compares well-documented hands across the research, humanoid-integrated, and standalone-supplier categories. Degrees of freedom are as stated by makers or in technical reports and often mix active and passive joints; treat them as approximate.

Scroll sideways for more →

Two structural observations follow from this list. First, the research and Western hands (Shadow, Allegro, PSYONIC, and the DLR and Robonaut lineage) established the field but are relatively expensive and low-volume. Second, the mass-market center of gravity has shifted to China, where a dense cluster of standalone suppliers (Inspire, Linkerbot, Xynova, Wuji, DexRobot, ZWHAND, and Unitree's captive Dex5) now sells hands as components to humanoid integrators worldwide, alongside a tactile-sensing specialist, PaXini, that supplies the touch layer. For the per-robot picture, see Humanoid robot hands.

Industry and supply chain

The dexterous-hand industry has organized into a three-tier structure. Upstream are component makers: motors and controllers, harmonic reducers, ball and roller screws, tactile and force sensors, structural parts, and flexible PCBs. Midstream are actuator and module makers who assemble those parts into joints and micro electric cylinders. Downstream are the dexterous-hand integrators who build finished hands, either as standalone suppliers (ZWHAND, Inspire, Linkerbot, Xynova, Wuji, DexRobot) or as captive programs inside humanoid makers (Tesla, Figure, Unitree, UBTECH, AgiBot).^[3][16][17]

A distinctive feature of the Chinese ecosystem is vertical integration and cross-investment. Xynova manufactures its own motors, controllers, planetary roller screws, reducers, and tendons, arguing that high-DOF hands are too sensitive to component matching to buy parts off the shelf.^[16] The components maker AAC Technologies reports roughly 80 percent self-sufficiency in key parts including coreless motors, six-axis force sensors, and inertial measurement units.^[3] Meanwhile large strategic investors have taken positions across the layer: Xiaomi, JD.com, and battery giant CATL have all backed hand or sensor startups, and PaXini's tactile-sensor rounds drew in BYD, JD.com, and even Meta.^[18][19] Several suppliers named on industry-chain maps are diversified public companies for which robotics is still an emerging line: Kedali's core business is battery structural parts, and Avary Holding and Dongshan Precision are consumer-electronics PCB makers, so their robotics exposure should not be overstated.^[17]

How big is the dexterous-hand market?

Estimates of the dexterous-hand market vary enormously, and the variance is mostly a matter of definition: a narrow "humanoid multi-fingered hand" market, a broad "dexterous hands" market spanning industrial and prosthetic uses, and a "tactile" or "robot end-effector" framing all produce very different numbers. Any single figure should be treated with caution, and reported CAGRs for this niche span roughly 12 percent to 87 percent depending on scope and source.^[44][45][46]

Scroll sideways for more →

The most defensible reading is a range, not a point: a broad dexterous-hand market on the order of $0.8B to $1.2B in the mid-2020s, growing at anywhere from about 12 percent (tactile/industrial framing) to roughly 40 percent (broad framing) or higher for the narrow humanoid segment.^[44][45] A widely circulated "72.38% CAGR" figure appears in no verifiable market report; it sits amid a cluster of real but low-reliability QYResearch-derived numbers (64.6 to 74.4 percent) and should be treated as unsourced.^[46]

Unit-shipment data from Chinese industry trackers is more concrete. China's dexterous-hand shipments exceeded 30,000 units in 2025, a figure derived from an estimated 15,000-plus humanoid robots (up from about 2,000 in 2024) carrying two hands each; Inspire Robots alone delivered 10,000 hands in 2025.^[3][40] Forecasts diverge: the China Commercial Industry Research Institute projects capacity of about 1.41 million units and more than $3 billion in revenue by 2030, while the tracker GGII projects Chinese shipments rising from about 19,200 units in 2025 to over 430,000 by 2030 at roughly 87 percent CAGR.^[3][46]

The strongest single anchor is cost share. According to a Morgan Stanley teardown of Tesla's Optimus, the dexterous hands account for about 17.2 percent of the robot's total cost and are the single most expensive component; industry commentary places the range at roughly 15 to 25 percent of the bill of materials, potentially exceeding 30 percent if very high-DOF hands become standard.^[46] This is why the hand, long treated as an afterthought, has become a strategic battleground: it is both the hardest part to build and one of the most expensive.

What are dexterous hands used for?

• Humanoid robots. The largest emerging market. General-purpose humanoids need hands that can use human tools and handle human-designed objects, which is precisely what dexterous hands provide.^[2]
• Manufacturing. Automotive final assembly, fastening, and 3C (computer, communication, consumer) electronics assembly, where dexterity and tactile feedback allow delicate, variable tasks that fixed grippers cannot.^[18]
• Logistics. Picking, packing, and palletizing of mixed items, a priority for warehouse operators such as JD.com.^[18]
• Prosthetics. Advanced myoelectric hands such as the PSYONIC Ability Hand restore grasping and, increasingly, a sense of touch to amputees, and share technology with robot hands.^[39]
• Teleoperation. Hands piloted remotely for hazardous work and, crucially, for collecting the demonstration data used to train autonomous policies.^[31]
• Research. Hands such as Shadow, Allegro, and DEX-EE remain the platforms on which manipulation science advances.^[36][37]

Open challenges

Despite rapid progress, several hard problems remain unsolved. In-hand manipulation, reorienting an object within the grasp without dropping it, is still limited; most hands grasp far better than they manipulate. Durability is a chronic weakness, especially for tendon-driven designs whose cables wear; Figure cited durability as its main reason for building tactile sensors in-house.^[10] Cost remains high for capable hands, keeping them out of price-sensitive applications. Tactile-sensor robustness is difficult: sensors that are sensitive enough to be useful are often too fragile to survive real work. And generalization, building hands and controllers that handle novel objects rather than a fixed set, is the central obstacle between today's demos and general-purpose deployment.^[24][47] The trajectory of the field suggests these are engineering and data problems rather than fundamental ones, but each is still a genuine barrier to hands that work reliably outside the lab.

ELI5: What is a dexterous hand?

A dexterous hand is a robot hand that works a lot like a human hand. A normal robot "gripper" is like a claw that can only open and close to hold something. A dexterous hand has many fingers and lots of little joints, so it can do fiddly things: turn a key, pick up a single grape, hold an egg without breaking it, or use a tool. It is one of the hardest robot parts to build, because you have to fit many tiny motors, gears, and touch sensors into something the size of a real hand and then control them all at once. To teach these hands what to do, engineers either let a computer practice millions of times in a simulation and then copy that onto the real hand, or they have a person "puppet" the hand to show it how, and the robot learns by copying. Human hands are still better than every robot hand, but robot hands are catching up fast, and companies in the United States, Europe, Canada, and especially China are now racing to build and sell them.

See also

• Bionic hand

• ORCA Hand

• TetherIA

• Kyber Labs

• The Robot Studio

• Sarcomere Dynamics

• PSYONIC

• Humanoid robot hands

• Tactile sensing

• Tendon-driven

• Robot manipulation

• Degrees of freedom

• Harmonic drive

• Dactyl

• Shadow Robot

• Allegro Hand

• Inspire Robots

• Xynova

• ZWHAND

• PaXini Technology

• Embodied AI

• Vision-language-action model

• Humanoid robot

References

1. "From Prototypes to Production: Dexterous Hands Kick Off a Mass-Production Race." Gasgoo Auto News, January 28, 2026. https://autonews.gasgoo.com/articles/news/from-prototypes-to-production-dexterous-hands-kick-off-a-mass-production-race-2016425582734970881 ↩
2. "Humanoid Robots 101." Bank of America Institute, April 29, 2025. https://institute.bankofamerica.com/content/dam/transformation/humanoid-robots.pdf ↩
3. "Humanoid Robots: Still Learning to Work, but Dexterous Hands Turn into a Hot Business." 36Kr, June 22, 2026. https://eu.36kr.com/en/p/3864146944414980 ↩
4. "A Review of the Research Status and Development Trends of Dexterous Hand Technology." International Journal of Mechanical and Electrical Engineering, 2025. https://wepub.org/index.php/IJMEE/article/view/5756 ↩
5. "Kinematics of human hand and robotics applications." International Journal of Science and Research Archive, 2024. https://ijsra.net/sites/default/files/IJSRA-2024-1441.pdf ↩
6. "The human hand has 27 degrees of freedom (DOF)." ResearchGate scientific figure, 2021. https://www.researchgate.net/figure/Finger-flexion-extension-Scene-The-human-hand-has-27-degrees-of-freedom-DOF-7-4-in_fig1_355898836 ↩
7. "Humanoid dexterous hands from structure to gesture semantics for enhanced human-robot interaction: A review." ScienceDirect, 2025. https://www.sciencedirect.com/science/article/pii/S266737972500049X ↩
8. "Tesla Optimus Gen 3 Hands: 22-DoF, 50 Actuators Explained." Basenor, 2026. https://www.basenor.com/blogs/news/tesla-optimus-gen-3-hands-22-dof-50-actuators-explained ↩
9. "Sanctuary AI integrates tactile sensors into Phoenix general purpose robots." The Robot Report, December 2024. https://www.therobotreport.com/sanctuary-ai-integrates-tactile-sensors-into-phoenix-general-purpose-robots/ ↩
10. "Introducing Figure 03." Figure AI, October 2025. https://www.figure.ai/news/introducing-figure-03 ↩
11. "Kinematic and Dynamic Analyses of the Stanford/JPL Hand." U.S. Department of Energy OSTI. https://www.osti.gov/servlets/purl/5658755/ ↩
12. Jacobsen, S. C., et al. "The UTAH/M.I.T. Dextrous Hand: Work in Progress." International Journal of Robotics Research, 1984. https://journals.sagepub.com/doi/10.1177/027836498400300402 ↩
13. "NASA and GM Develop Dexterous Humanoid Robonaut2." IEEE Spectrum, 2010. https://spectrum.ieee.org/nasa-and-gm-develop-dexterous-humanoid-robonaut2 ↩
14. Butterfass, J., Grebenstein, M., Liu, H., Hirzinger, G. "DLR-Hand II: Next Generation of a Dextrous Robot Hand." IEEE ICRA, 2001. https://www.dlr.de/en/rm/research/robotic-systems/hands/hand-ii ↩
15. "Learning Dexterity." OpenAI, July 30, 2018. https://openai.com/index/learning-dexterity/ ↩
16. "Xynova Defines New Height of Dexterous Operation with Flex2 after Securing Hundreds of Millions of Yuan in Series A." 36Kr, May 2026. https://eu.36kr.com/en/p/3829065404211456 ↩
17. "Where China leads and lags in humanoid joint architecture." Interesting Engineering, 2025. https://interestingengineering.com/ai-robotics/china-humanoid-robots-actuators ↩
18. "PaXini Unveils the Tactile Infrastructure for Embodied AI at CES 2026." PR Newswire, January 2026. https://www.prnewswire.com/news-releases/paxini-unveils-the-tactile-infrastructure-for-embodied-ai-redefining-full-stack-product-matrix-at-ces-2026-302655238.html ↩
19. "Unicorn born in record time amid 'arms race' among China's robotic hand developers." South China Morning Post, May 30, 2026. https://www.scmp.com/tech/article/3355365/unicorn-born-record-time-amid-arms-race-among-chinas-robotic-hand-developers ↩
20. "Scaling the humanoid robotics supply chain into billion-dollar wins." McKinsey & Company, 2025. https://www.mckinsey.com/industries/industrials/our-insights/turning-humanoid-supply-chain-constraints-into-billion-dollar-wins ↩
21. "Sanctuary AI Demonstrates In-Hand Manipulation Capabilities for Improved General Purpose Robot Dexterity." Sanctuary AI / PR Newswire, December 2024. https://www.prnewswire.com/news-releases/sanctuary-ai-demonstrates-in-hand-manipulation-capabilities-for-improved-general-purpose-robot-dexterity-302329804.html ↩
22. "Alpha: Water-powered humanoid robot with synthetic organs, muscles unveiled." Interesting Engineering, 2024. https://interestingengineering.com/innovation/clone-alpha-humanoid-robot-unveiled-poland ↩
23. "Tactile Sensor Technologies Explained." LOOMIA Soft Electronics. https://www.loomia.com/blog/tactile-sensor-technologies-explained ↩
24. "Learning with Less: Optimizing Tactile Sensor Configurations for Dexterous Manipulation." arXiv, 2024. https://arxiv.org/abs/2409.20473 ↩
25. "Multifingered robot hands: Control for grasping and manipulation." Annual Reviews in Control / ScienceDirect, 2010. https://www.sciencedirect.com/science/article/abs/pii/S1367578810000416 ↩
26. Andrychowicz, M., et al. "Learning Dexterous In-Hand Manipulation." arXiv:1808.00177, 2018. https://arxiv.org/abs/1808.00177 ↩
27. "Solving Rubik's Cube with a Robot Hand." OpenAI, October 15, 2019. https://openai.com/index/solving-rubiks-cube/ ↩
28. OpenAI. "Solving Rubik's Cube with a Robot Hand." arXiv:1910.07113, 2019. https://arxiv.org/abs/1910.07113 ↩
29. Qi, H., Kumar, A., Calandra, R., Ma, Y., Malik, J. "In-Hand Object Rotation via Rapid Motor Adaptation." arXiv:2210.04887 (CoRL 2022), UC Berkeley. https://arxiv.org/abs/2210.04887 ↩
30. "Rotating without Seeing: Towards In-hand Dexterity through Touch." arXiv:2303.10880 (RSS 2023), UC San Diego and HKUST. https://arxiv.org/abs/2303.10880 ↩
31. Handa, A., et al. "DexPilot: Vision-Based Teleoperation of Dexterous Robotic Hand-Arm System." NVIDIA, ICRA 2020. https://research.nvidia.com/publication/2020-05_dexpilot-vision-based-teleoperation-dexterous-robotic-hand-arm-system ↩
32. Qin, Y., et al. "AnyTeleop: A General Vision-Based Dexterous Robot Arm-Hand Teleoperation System." arXiv:2307.04577 (RSS 2023). https://arxiv.org/abs/2307.04577 ↩
33. "RT-2: New model translates vision and language into action." Google DeepMind, July 28, 2023. https://deepmind.google/blog/rt-2-new-model-translates-vision-and-language-into-action/ ↩
34. "pi-0: Our First Generalist Policy." Physical Intelligence, October 31, 2024. https://www.pi.website/blog/pi0 ↩
35. "Helix: A Vision-Language-Action Model for Generalist Humanoid Control." Figure AI, February 2025. https://www.figure.ai/news/helix ↩
36. "Shadow Dexterous Hand." Shadow Robot Company, 2025. https://www.shadowrobot.com/dexterous-hand-series/ ↩
37. "Our latest advances in robot dexterity." Google DeepMind, 2024. https://deepmind.google/blog/advances-in-robot-dexterity/ ↩
38. "Allegro Hand." Wonik Robotics. https://www.allegrohand.com/ ↩
39. "Ability Hand." PSYONIC. https://www.psyonic.io/ ↩
40. "INSPIRE ROBOTS Delivers Answers with Ten-Thousand-Unit Delivery." Inspire Robots, 2026. https://en.inspire-robots.com/news/7719.html ↩
41. "Unitree Unveils Dex5 Dexterous Hand: 20 Degrees of Freedom in a Single Hand." AIbase, 2025. https://www.aibase.com/news/16777 ↩
42. "Freakishly Lifelike Robotic Hands Surge to Market." Mike Kalil, 2025. https://mikekalil.com/blog/freakishly-lifelike-robotic-hands/ ↩
43. "ZWHAND Shines at CES 2026: Redefining Robotic Hands for Mass Embodied AI." PR Newswire, January 12, 2026. https://www.prnewswire.com/news-releases/zwhand-shines-at-ces-2026-redefining-robotic-hands-for-mass-embodied-ai-302659035.html ↩
44. "Multi-Dimensional Tactile Robotic Dexterous Hand Market to Reach USD 3.8 Billion by 2036." Future Market Insights / openpr, January 2026. https://www.openpr.com/news/4528350/multi-dimensional-tactile-robotic-dexterous-hand-market ↩
45. "Dexterous Hands Market Size to Reach USD 10.3 Billion by 2031 Driven by Humanoid and Industrial Robots." Valuates Reports / PR Newswire, January 20, 2026. https://www.prnewswire.co.uk/news-releases/dexterous-hands-market-size-to-reach-usd-10-3-billion-by-2031-driven-by-humanoid-and-industrial-robots--valuates-reports-302665587.html ↩
46. "Humanoid Robot Multi-fingered Dexterous Hand Research." QYResearch, February 2025. https://www.qyresearch.com/news/10074/humanoid-robot-multi-fingered-dexterous-hand ↩
47. "ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning." arXiv:2504.04259, 2025. https://arxiv.org/abs/2504.04259 ↩