Journey from human hands to robot hands: biological inspiration of anthropomorphic robotic manipulators

The development of robotic hands that can replicate the complex movements and dexterity of the human hand has been a longstanding challenge for scientists and engineers. A human hand is capable of not only delicate operation but also crushing with power. For performing tasks alongside and in place of humans, an anthropomorphic manipulator design is considered the most advanced implementation, because it is able to follow humans’ examples and use tools designed for people. In this article, we explore the journey from human hands to robot hands, tracing the historical advancements and current state-of-the-art in hand manipulator development. We begin by investigating the anatomy and function of the human hand, highlighting the bone-tendon-muscle structure, skin properties, and motion mechanisms. We then delve into the field of robotic hand development, focusing on highly anthropomorphic designs. Finally, we identify the requirements and directions for achieving the next level of robotic hand technology.


Introduction
The human body is regarded as one of the most sophisticated structures on earth, with the human hand being a highly exquisite manipulator that can perform everything from the delicate touch of fragile objects to the power grasping of a bench press.What enables this dexterity, and how can we replicate it in the form of robotic hands, prosthetic hands, or other comparably accurate and versatile manipulators?These questions have been at the forefront of the minds of scientists, engineers, and neurosurgeons [1][2][3][4][5][6], and have even influenced popular culture.Why have we been drawn to the creation of incarnate beings?In the classic novel 'The Adventures of Pinocchio' [7], Geppetto was compelled to carve a living boy from a miraculous crying log, ultimately creating Pinocchio.From a sentimental perspective, Geppetto's actions may have been driven by a desire to have a lovable child and secure the future generation.However, from a more plausible standpoint, Pinocchio-an impersonating structure-could be seen as the most advanced version of creation that humans are capable of imagining.
Whether driven by practical or artistic motivations, the goal of the biomimetic human hand has produced a large body of research literature over the past 50 years.In this article, the direction of past and current development of biomimetic hand manipulators is reviewed to build an understanding of nextlevel challenges.Then, by investigating human hands with respect to anatomy, structure and function, we consider the future scope of development for biomimetic robotic hands.Based on this anatomical study of human hands, we lay out artificial counterparts of biological structures for robotic hand manipulators in terms of functions, materials, and viability.

Trends in anthropomorphic hand development
Because the interest in robotic hands has been growing steadily for a long time and includes nonbiomimetic approaches, 'robotic hands' is too broad of a search term for this review, which focuses on biomimetics, materials, and integration.Our initial search methodology used the keywords 'biomimetic robotic hand' , 'anthropomorphic robotic hand' , and 'bionic hand' with an 'OR' operator, collecting articles for sorting within the scope of this review.As shown by the gray bars in figure 1(1) and ( 2), the publication rate on these topics has increased across the years.The largest number of publications is in robotics journals, but articles also appear across multiple engineering domains and in neuroscience.
Although there were attempts to emulate human hands as iron prosthetics in the early Roman era [8] and other studies on robotic hands in the 1960s [9], research into robotic hands with human-like shapes intensified in the 1980s.The most well-known examples are the Salisbury hand (1982) from Stanford [10] and the Jacobsen hand (1984) from Utah/MIT [11].Figure 1(3) illustrates how those initial studies, shown as large dots on the upper left of the citation map, influenced subsequent research, with lines indicating citations between works.Moreover, figure 1(3) displays colors representing each study's focus.Major focus areas include design, mechanism, and control.
In the 1980s and 1990s, a sophisticated design framework evolved, based on an anatomical point of view.The intention was to make the system more dexterous [12][13][14][15].In recent years, advances in materials science, sensor technology, and artificial intelligence have continued to drive the development of more advanced and sophisticated robotic hands [16][17][18][19][20][21].Today, there are a wide variety of robotic hands and prosthetic devices that are capable of mimicking the complex movements and dexterity of the human hand, and these devices are being used in a variety of applications, from manufacturing and logistics to healthcare and rehabilitation [22].
Despite these significant advancements in design, fabrication, control, and dexterity, there still exists considerable room for improvement.Furthermore, figure 1(3) demonstrates a paucity of research that integrates three or more distinct areas of focus, indicating that there are untapped opportunities to explore in this multidisciplinary field.This analysis underscores the potential for new breakthroughs and further acceleration of progress.

Overview of human hand anatomy-biological inspiration
Alongside development of prosthetic hands and robotic hands, there have been concurrent studies of human hand anatomy, functions, and mechanisms geared toward a better understanding of how our hands work.This topic is crucial to improvement of biomimetic artificial hands [2,24,25].Therefore, in this section, human hands' structure is briefly reviewed in terms of the bone-tendon-muscle framework as well as the skin.We also cover previous research into understanding human hands' motion in both functional and mechanical aspects.

Bone-tendon-muscle structure
Each human hand consists of 27 bones: 3 in the thumb, 8 in the wrist and, 16 in the rest of four fingers like figure 2. Bones are the physical support for our hands [23].These bones are connected with other bones and muscles by tendons.One of the main anatomical patterns is paired main tendons to open (extensor) or close (flexor) each finger, with the muscles connected to both extensor and flexor tendons located in the forearm.Interestingly, however, the muscles for side to side finger motions (adduction, abduction) lie in the palm [26].By distributing those muscles and tendons in the right position with respect to tendons and bones, each finger joint is capable of its full range of motion.
In order to better mimic human hand motion, there have been mechanical and clinical investigations to measure tendon excursion and moment arm of finger muscles since those motions and forces define finger dynamics and explain internal forces and joint torques [25,[27][28][29]].It's not just a matter of pulling on a tendon by distance X and getting a fingertip displacement of Y, because of the elastic properties of biological materials.An interesting feature specific to hands is that finger tendons are a bit more viscoelastic and less elastic compared to other tendons in our body such as legs, allowing for precise control of movement and passive motion against external disturbances [30].On top of that, the synovial sheath covers and guides tendons' path so that tendons and muscles are able to move smoothly and efficiently with natural lubricant [31].Likewise, the bone and tendon of the human hand have specific physical properties in terms of their location, material, shape, and even texture which makes for integrated versatile motion.Therefore, it is essential to understand those mechanical characteristics and mechanisms in human hands and expand the features to modern biomimetic robotic hands.This subject will be expanded in later sections.
The end of the tendon is connected with the muscle as mentioned before.In figure 3, the skeletal muscle located in the hand and forearm is composed of a bundle of muscle fibers.The function of the human muscle is obviously making motion.Researchers have developed sliding filament theory to describe the microscopic mechanism by which muscles move [32,33].Sliding filament theory is summarized in the bottom left of figure 3, which illustrates how human muscle contracts and relaxes in molecular units.Described in a further main section, methods to actuate robotic hands are one of the major research topics in anthropomorphic robotic hand design.

Skin
Human hands' adaptability and dexterity stem from not only the above anatomical framework, but also from sensory neural receptors at the skin surface and below [34].Figure 4 shows the structure of thick (palmar side) skin, which is distinguished by the thickness of the epidermis.The function of skin is not just limited to obtaining tactile information, but to deforming around objects.
There are three layers of the skin (epidermis, dermis, hypodermis) as shown in figure 4. Epidermis itself has low mechanical strength, yet it plays a key role to protect underlying skin layers from abrasion.Furthermore, due to the unique ridges (fingerprints) on the outer surface of the epidermis, hand manipulation exerts contact friction that leads to a more stable grip.The dermis is the dominant underlayer of the skin, and it determines the skin's overall mechanical properties.Because the dermis has high elasticity and tensile strength, it serves as an absorber of external shock along with the hypodermis.Thus, thanks to the physical properties of each stratum, the skin helps human hands conform to objects during grasping and spring back to their original shape by absorbing shock and elastically restoring.
Meanwhile, most of the tactile sensory receptors (mechanoreceptors) lie in between the epidermis and dermis section to accept mechanical deformation of the skin easily, while staying protected by surrounding layers [35].Interestingly, many current artificial skin studies have tried to implement and exploit this human skin stratum [36][37][38][39].

Motion mechanism
In the pursuit of highly versatile robotic hands, the fundamental question is how to effectively grasp, manipulate, and interact with objects.Consequently, researchers often describe the remarkable versatility  of the human hand function by classifying its various prehensile actions [40][41][42].However, it is important to note that human prehension is a complex process that integrates the intricate interplay of the bone-tendon-muscle structure, the sensory system embedded in the skin, and even cognitive processing including planning for multiple tasks.Napier points out that power grasping and precision manipulation are not mutually exclusive [42], implying that not only can human hands make separate postures for different tasks, but they can also perform multiple postures at once like in figure 5.The overall classification system shows clearly the versatility and range of human hands.Based on this physiological standpoint, many current researches present posture versatility of artificial hands as a performance standard similar to the grasping classification in figure 5 [43][44][45][46][47].

Robotic hand functional realization
The examination of human hand anatomy assumes pivotal significance in advancing robotic manipulators due to the inherently human-centric design of our daily tasks.This paradigm extends beyond commonplace implements like kitchen utensils  Selected grasping taxonomy from simple and complex daily life motions; subset of a larger taxonomy presented in [41].
to encompass industrial equipment and research instruments.These tools are meticulously crafted with ergonomic considerations, ensuring compatibility with the average human hand in terms of size and shape.Ranging from delicate fine-tipped tweezers to robust hammers and cordless power tools, the diversity of tasks necessitates a multifaceted approach.While the robot's adherence to dexterity and versatility is paramount, safety considerations become equally crucial in handling fragile and heavy devices, requiring the ability to modulate between soft and strong grips, as well as subtle and forceful touches.
Moreover, the interconnection of signals from tools with sensory feedback assumes key importance in manipulation control.These signals, encompassing pressure, temperature, and vision feedback, enable humans to seamlessly interact with their environment, establishing benchmarks for evaluating the success of manipulation tasks.In essence, understanding of human hand anatomy not only informs the design of robotic manipulators but also underscores integrating safety features and sensory feedback mechanisms to improve their overall effectiveness in many operational contexts.
The ultimate objective is to transfer human labor to robot labor, prevent workplace injuries, and move beyond human physical limitations.With industries such as food & beverage, aerospace & defense, and energy & power emphasizing 'smart factories' where supplies and outputs are tracked throughout the manufacturing process, a highly sophisticated robotic manipulator could also raise the manufacturing performance, for example by sensing increases in assembly forces of injection-molded parts that indicate a mold is wearing out.
Therefore, inspired by human hands, there have been efforts to endow robotic manipulators with human-like design, motion and function without completely duplicating all features of human hands.Different implementations geared toward different objectives have come out, including factory automation, clinical prosthetics, humanoid robots, and exploration robots, with various design features emphasized depending on the system's purpose [48].Since cataloging the full diversity of these manipulators is beyond the scope of this article, this section is bounded to review highly anthropomorphic designs of robotic hands.
Recent developments often focus on new materials to implement state-of-the-art designs.Compared to 1980s-1990s designs, the 2010s-2020s designs are concerned not only with solid structures but also the softer components of robotic hands.This development is made possible by advanced fabrication processes, improved computing performance and multidisciplinary collaboration [49,50].In this section, by investigating robotic counterparts of human hands from a mechanical point of view, we show how anthropomorphic hands have been advanced in terms of design, material, modeling and, control.

Main framework
Analogous to human bones, hard frames in robot hands provide support for other structures, sensors, and actuators.Such frames are able to perform highly coordinated movements as well as retain their own shape.Several early stage developments relied on frames, mainly focusing on simple manipulating functions with only rigid bodies [14,25,51].In order to make the system adapt to the environment, active control strategies based on concrete modeling analysis have been investigated [52,53].Although these previous approaches led to fairly straightforward and intuitive drive and control methods, in some industries such as those dealing with delicate or fragile materials, robotic hands need to perform tasks in a conformable, passive manner [54,55].Accordingly, there is a trend to adopt shapes, sizes, and structures that resemble human hands' yielding the bone, tendon, sheath, ligament, and skin [56][57][58].This trend is shown in table 1, where the majority of the highly anthropomorphic designs feature elastomers in one or more components to realize active and passive mechanism mutually.
In applications where precision and accuracy of motion are critical, typically rigid body materials are more commonly utilized such as metals, hard plastics, strong fibers or cables.This choice facilitates the analysis of kinematics and dynamics, allowing for the isolation of linear terms while neglecting nonlinear factors like friction, hysteresis, and viscoelastic features [60,61].In contrast to the somewhat human bone-shaped structures, recent investigations indicate that joint structures in robotic systems tend to be simplified, featuring 1 or 2-axis rotational joints to connect and mobilize individual phalanges.Additionally, innovative designs incorporating pulleys, either within or external to the hand, have been proposed.These designs not only guide tendons and reduce friction but also address underactuation challenges, occasionally at the expense of sacrificing one degree of freedom (DOF) in the finger joint.Notably, some prosthetic hands continue to adhere to rigid bodies and relatively stiff cable wires for reasons of simplicity in control setups, cost efficiency, and compactness [6,61,64].
However, a paradigm shift is observed in biomimetic robotic manipulators and end-effectors, where a departure from rigid materials is embraced.Leading the way, these systems explore the use of softer, lighter, and safer materials.This strategic move towards compliant materials not only promotes biomimicry but also holds promise for enhanced adaptability and safety in various applications [65,66].These soft materials whose motion is coupled to rigid bodies can expand the range of function of existing robotic hands.Several works have gone even softer, adopting a compliant and passive framework instead of rigid bones so that the manipulator can have a higher DOF [20,59].On top of that, some studies in soft robotic hands show more flexible sensorsembedded designs or smart-materialized designs, which are able to control the finger position or stiffness straightforward [63,67].But, in those cases, a certain level of pressure inside the finger is required to maintain its stiffness corresponding to rigid structures.In addition, understandably, the higher the DOF, the more complex the control system.
Therefore, many recent researches has shown how to interlace active and passive motion by selecting rigid and soft materials properly according to its main objectives.Besides, advanced 3D printing and laser cutting technologies excel sophisticated human finger design [66,68].

Actuation and transmission
Once the main framework of the hand is established, it is time to generate motion.Depending on the objective of the system, the force transmission method, force generation method, and actuator placement will vary.Table 2 describes the major force transmission methods and their pros and cons.Some of these transmission categories are closely integrated  [47].
Pneumatic [63] Optical fiber, Nylon fabric, Silicone elastomer -Optoelectronic sensor for touching and grasping  with force generation methods (for example, artificial pneumatic or polymer muscles that actuate in place), while tendon and linkage based systems may be driven by a variety of force generators.As illustrated in the figure 6, the prevailing architecture for robotic hand mechanisms, particularly in biomimetic applications, is predominantly tendondriven.In the tendon-driven system, tendon cables placed in tightly confined spaces transmit the force to move the digits at the user's command.Since the multiple tendons can be connected directly to each phalange, it is also able to be applied to comparatively precise position control likewise the linkage-driven types.Due to weight, compliance and size limitations, however, the linkage-driven mechanism is less popular these days even though it has easy-to-design and higher motion accuracy features than other methods.
Conversely, the tendon-driven method offers a measure of system compliance owing to the elastic properties of tendons.While this introduces complexities in control and imposes limitations on maximum force transmission due to elongation and hysteresis, the benefits often outweigh the drawbacks.In recent times, advancements within the tendondriven category have seen the integration of soft and robust materials.Twisted cables, commonly referred to as twisted string actuators (TSA), and twistedcoiled polymers (TCP) have emerged as innovations combining force generation and transmission.These materials exhibit superior power or weight efficiency compared to traditional tendon systems [21,46,80].Given that this actuation and transmission system mirrors the muscle and tendon interplay in the human hand, tendon-driven mechanisms have been explored with alternative power sources, including motors and hydraulics/pneumatics [81].
Moreover, recent researches suggests the potential for replacing essential components such as main bodies, linkages, joints, and even tendons with alternatives employing pneumatic systems, and shape memory alloy (SMA)/shape memory polymer (SMP) [82].These materials serve not only as mediums for force transmission but also as integral elements of the power source [67,83].As interest in human-machine interaction intensifies, there is a growing focus on soft and conformable bodies.These adaptive structures have the capability to encompass a broad spectrum of forces, ranging from delicate grips to power grasps, while dynamically adjusting to the shape and stiffness of various objects.
Likewise, each force generation method brings a new control challenge as well as advantages due Capacitive [87] Ionic hydrogel, Fabrics, Elastomers 1.5 kPa Reproduced with permission from [90].
Vision [91] Elastomer, plastic 640 × 480 (image) High spatial resolution, Miniaturizable to distinct physical properties as mentioned previously.Precise control of these soft, continuous, high-DOF actuation methods leads to another challenge: proprioception, or internal detection of the actuators' state, for closed-loop control.[84,85] Therefore, in the next section, it is introduced how researchers are addressing the challenge using embedded sensors.

Sensing
As mentioned earlier, human hands' dexterity comes from the interplay of frame structures, actuators, and sensory receptors.Sensor systems for measuring tactile information have been developed not only to meet demands of robotic hands, but also for other applications including medical devices and pressure mapping surfaces [39,86].Table 3 collects recently developed tactile sensors, focusing on those inspired by the structure of human skin.Most use soft and flexible materials so that they can conform to the objects being touched, as well as interact with external inputs within the force range of human hand manipulation tasks [38,[92][93][94][95][96].
While the inspiration and physical properties of various projects are aligned, table 3 highlights distinctly different transduction mechanisms, materials, and components.The majority of soft-type sensors commonly detect changes in electrical signals, such as alterations in resistance, capacitance, or electrical charge in response to mechanical deformation [87,88,96].These sensors can be seamlessly embedded onto elastomers in compact sizes, facilitating attachment to surfaces with flexible and stretchable features.While electric sensing mechanisms boast high accuracy and sensitivity, they may be susceptible to a high signal-to-noise ratio (SNR) or interference.
In contrast, optic-based sensing mechanisms are less susceptible to electrical or magnetic field interference and offer water resistance.Many optical sensors employ flexible optical fibers, enabling interfacing with soft-type polymers for use as skin sensors.Although these sensors are cost-effective and easy to fabricate, they may encounter issues related to optical attenuation due to ageing or analog signal calibration [97].
A more sophisticated version of optical sensors employs a vision-based sensing mechanism, exemplified by systems like TacTip [92], GelSight [98], and DIGIT [91].In these systems, an image-capturing camera or optic-based vision sensor embedded within elastomeric material facilitates the detection of objects' shape and texture.While such systems were initially bulky, they have undergone continuous refinement in recent years.Additionally, magnetic sensing mechanisms have garnered attention, especially with the integration of magnetic particles into polymeric materials to design soft, thin, skin-type tactile sensors [90].
The main concern of these studies is obtaining normal force, detecting lateral motion, surface contact information and collecting other signals within the capability of human skin such as temperature.The skin sensor community assesses and compares projects on their force resolution, minimum and maximum measurable force, and tactile sensor element size [99].

Modelling and control
The primary focus of robotic hand modeling and control lies in emulating human hand unit motions, encompassing the regulation of contact touch force, grasping force, finger stiffness, response speed, and fingertip positions within constrained workspaces.Researchers pursuing anthropomorphic robotic hands have made notable strides in novel frame designs, diverse actuation mechanisms, and skin-type sensor systems, as previously discussed.However, substantial challenges persist, with each section exhibiting significant advancements yet lacking cohesive system integration and the attainment of seamless motions.Like figure 7, a biomimetic scheme for motion control might be intricate as much as the human nervous system.
According to the classic engineering concept, a controller is constructed by starting with a solid model of the system.Indeed, such attempts have been made to control robotic hands accurately, with active control concepts added to make the system more robust [43,101].Based on a thorough forward and inverse kinematic analysis, the position and orientation of the system can be successfully controlled.Dynamic modelling allows researchers to go further, simulating the hand's behavior under different loading conditions and aiding in the development of control strategies [102].These modeling processes involve techniques of not only dynamics and kinematics algorithms based on geometric and numerical optimization, but also data-driven approaches based on experimental measurements [22,[103][104][105].
Such data-driven approaches are the current trend.As the system gains components with different physical properties, such as when semi-rigid frameworks are combined with soft tendons and actuators, the variables multiply and the system becomes too complex for exact description.Thanks to advanced imaging technologies and improved computing performance, partial data-driven modeling methods have been actively researched to make the system more accurate without becoming computationally intractable [106,107].
On the design side, underactuated designs help researchers reduce the system complexity because they have fewer actuators than DOF [6,108].And on the controls side, a concept called 'synergies' has been studied in the context of achieving dexterous and efficient movements [70,109,110].Synergies provide coordinated patterns of joint movements that allow the hand to perform grasping and manipulation tasks effectively.Since the underactuated design allows for shared mechanical elements, and since synergies give a simplified variable of motion, researchers have connected the two concepts closely to improve robotic hands' capabilities [20,111].

Discussion
Over the past 50 years, advancements in materials science, sensor technology, and artificial intelligence have propelled the field of biomimetic hands, producing a wide variety of robotic hands and prosthetic devices.Key advances in the past 10 years include soft sensors and actuators, machine learning coupled with sensors, and a new appreciation for the mechanical intelligence of compliant and underactuated mechanisms.
However, despite significant progress, there are still areas that require improvement, including design, fabrication, control, and practicality.Notably, there remains a dearth of research on the seamless integration of disparate technologies for the purpose of effective sensing, processing, and dexterous motion.For instance, as previously discussed, while significant progress has been made in the development of anthropomorphic design, tactile skin sensors, and various types of actuation, there is still a need to explore their intimate and natural connection between those systems not only mechanically but also computationally in order to advance the field of robotic manipulators to the next level.
Evolution has shaped our current hand structure, while humans have subsequently created the diverse environments that we interact with on a daily basis.The need to manipulate objects in varied settings has necessitated the transfer of this dexterity to robots, allowing for the delegation of mundane and repetitive tasks.
Ongoing research focuses on understanding the anatomy and functions of the human hand to guide the development of robotic counterparts.Fueled by emerging materials, fabrication methods, and computing techniques, one day we may develop manipulators that can mimic or even outperform the complex movements and capabilities of the human hand.
In conclusion, the journey from human hands to robot hands has been marked by steady progress, but there is still much room for improvement.By studying and understanding the human hand's structure and capabilities, researchers and engineers can continue advancing robotic hand technology, aiming for even greater dexterity, functionality, and integration with human users.

Figure 1 .
Figure 1.(1, Top left) Fraction of publication count in top 10 selected research areas from Web of Science (publications in two or more areas are counted in multiple wedges).(2, Top right) All 1517 publications in this analysis, distributed across years 1971-2023 (Web of Science, WoS).(3, Bottom) Development trend for 60 selected publications including the Salisbury hand and Jacobsen hand (two large dots at upper left).Publications are organized from left to right by year (1977-2022), and organized top to bottom by number of citations (also indicaed by dot size), plotted using Litmaps.Dots are color coded to indicate one or more focus areas.

Figure 2 .
Figure 2. (Left) Illustration of human hand bone structure, (Right) Example of a possible artificial hand component arrangement, inspired by human hands.

Figure 3 .
Figure 3. (Top) Illustration of tendon-muscle structure of human hand.Partially reproduced from [23].(Bottom) Sequential zoom on the 'belly' structure of a skeletal muscle.Motion generation starts from the molecular unit (bottom left).Motion is reinforced by multiple units that form larger skeletal muscle units and ultimately drive the whole hand.Adapted from [23], Copyright (1967), with permission from Elsevier.

Figure 4 .
Figure 4. (Left) Illustration of human skin structure and mechanoreceptors for tactile sensing, (Right) Example of artificial skin.

Figure 5 .
Figure5.Selected grasping taxonomy from simple and complex daily life motions; subset of a larger taxonomy presented in[41].

Figure 6 .
Figure 6.Proportion of transmission methods out of 196 research articles dealing with the development of the anthropomorphic robotic hand.Search keywords example : 'anthropomorphic' AND 'robotic' AND 'hand * ' AND 'tendon * ' from WoS.

Figure 7 .
Figure 7. Human response control process from skin to muscle.Brain and spinal cord response time are around 200 ms and 30 ms respectively [100].Each possible robot system counterpart is presented with Italic font.

Table 1 .
Recent developments in anthropomorphic robotic hand design by years.

Table 2 .
Advantage and disadvantage of each force transmission method.

Table 3 .
Recent skin-mimetic sensors use an array of mechanisms to produce tactile sensor elements.