On the development of intrinsically-actuated, multisensory dexterous robotic hands

Restoring human hand function by mechatronic means is very challenging in robotics research. In this paper, we first make a brief review on the development of dexterous robotic/prosthetic hands, and then detail our design philosophy of several robot hands. We make a concentration on a type of intrinsically-actuated robot hands, wherein the driving, transmission, and control elements are totally embedded in the hand. According to different application scenarios, we develop robot hands in two parallel lines, dexterous robotic hand and anthropomorphic prosthetic hand. In both, the hand’s actuation, sensing, and control subsystems are highly integrated and modularized. This feature endows our robot hands with compact appearances, simple integration, and large flexibilities. At last, we give some perspectives on the future development of dexterous hands from the aspects of structure, functionality, and control strategies.


Background
As a powerful tool, a large variety of robotic systems has been applied to help human beings explore unknown or hazardous areas such as outer space, deep sea, or contaminated nuclear plants. To achieve effective explorations, a dexterous end-effector with superior operation and perception capabilities is an urgent need. Although traditional grippers can deal with some simple, fixed tasks (grasping and transferring workpieces), their low commonality, humble perception and insufficient flexibility make them hardly competent to complex operations in unstructured environment. Then, dexterous robotic hands (DRHs) with multiple degrees of freedom (DOFs), superior operational and perceptional capabilities arouse great attentions in the robot society [1]. Currently, although a large progress has been made, the DRHs available on the market still cannot compete to biological hands due to current technical constraints on actuators, sensors and control means. It is indicated that, rather than simply imitating the human hand, the research should switch to fully exploiting the robot hand's advantages, while considering specific requirements (manipulative dexterity, grasp robustness, or human operability) that allow for successful, fluent, and dexterous operations [2].
As a branch of robotic hand research, the anthropomorphic prosthetic hand (APH) is a type of biomechatronic device used to restore hand motions for amputees or paralyzed patients. On this topic, great efforts have been made from both robotics and biomedical engineering. However, current prosthetic hands still cannot compete to a human hand in respect of structure, sensing, and control strategy. Only a few of prosthesis products can obtain their commercial success. Because of unintuitive control feelings, lack of sensory feedback, and poor hand functionality [3], a large portion of users often refuse to use their prosthesis. After analyzing human hand's activities of daily life (ADL's), a study reveals that a superior hand prosthesis should have more controllable functions, faster response/shorter reaction time, and an intuitive control and feedback strategy [4].
The advanced prosthetic hand systems are then characterized by its anthropomorphic appearance, congenital dexterity (including both mechanical structure and sensors), and high-level mechatronic integration. As for the hand's manipulation capability, it is normally held that the hand dexterity improves as the number of active joints increases. However, studies also shows that, as the number of the active joints increases, the dexterity of a prosthetic hand may even decrease due to the intensified control complexity. Therefore, the prosthetic hand design should consider more comprehensive factors, such as the compromise between dexterity and controllability, the suitability and adaptability of the sensory feedback, as well as essential neural rehabilitation principles [5].
After briefly reviewing some representative studies, in this paper, we detail our development process of several DRH and APH prototypes. From a view of biomechatronics, we also prospect some directions on the development of advanced robot hands, after fully acknowledging the challenges in front of us.

Representative studies
So far, over hundreds of robot hands, including DRHs and APHs, have been developed in academic colleges, research institutions and companies. Among these hands, the DRH is a special topic aiming to reproduce human hand's manipulation dexterity by mechatronic means. According to drive position (inside or outside the hand), the DRH can be mainly divided into two categories: intrinsic actuation pattern (IAP) or extrinsic actuation pattern (EAP). Some representative DRHs with specifications of number of fingers, number of active DOFs, actuation configuration, and transmission mechanism are shown in Table 1.
The design of robotic hands on the early stage, such as Stanford/JPL Hand and Utah/MIT Hand, are generally focused on the hand's anthropomorphism and multi-sensory integration. The DLR-I Hand [7] is a representative of the first generation DRH featured with independent actuation. To enhance the hand's appearance and operating flexibility, the DLR-II Hand [8] further introduces an extra DOF between the thumb and the ring finger for offering palm curling. Driven by air muscles, the shadow hand has more than 20 DOFs that endows the hand with noticeable grasp functions. Besides pure IAP and EAP, many DRHs (such as the iCub hand [15]) also adopt a hybrid driven pattern, wherein multimodal sensors (tactile, position, and force) are also integrated for providing more proprioception information.
Some DRHs are developed for special space and military applications, such as the NASA's Robonaut Hand I, II, and the DLR's Dexhand. Nowadays, the Robonaut hand II has been tested successful in the International Space Station (ISS) to assist astronaut. Meanwhile, the Dexhand also has some special design, such as its transmission system (Dyneema tendon plus harmonic reducer), control system (totally integrated into the hand), and communication system (CAN Bus and VxWorks controller), for properly working in the space environment. In addition, the DLR's HASy Hand is a new type multi-finger DRH to be used in the bionic Hand-Arm system [10]. It has a similar size, weight, and even behavior to a human hand. To reproduce the dynamic characteristic of the human hand, joints of the DLR HASy Hand are integrated with a special variable stiffness actuation system (VSA, consisting of servo modules and elastic elements) [33]. All actuation and electronic systems are embedded in the forearm, making it easy to be integrated in any concrete applications.
Together with DRH, the development of APH also gets a large promotion. During the last decades of 20th century, many multi-DOF prosthetic hands come into being, such as, Southampton prosthetic hand [34], Oxford Intelligent prosthetic hand [35], Stanford prosthetic hand [36], and NTU prosthetic hand [37]. Due to the actuation techniques and manufacturing level at that time, these prosthetic hands are generally large, heavy, and provided with very limited number of sensors. Upon entering 21st century, the development of APH shows a diverse tendency where the design guidelines are no longer simply "reforming" an existing DRH or totally "reproducing" the human hand. Both scope and depth of interdisciplinary fusion with relevant to mechanic, electronic, biology and control are getting strengthened in the APH development. Today, an ideal APH should possess a humanlike appearance, as well as high-level dexterity. As well, it should be comfortable to wear and, more importantly, easy to control. We list a collection of representative APH prototypes, as Table 2 shows.
In particular, the DARPA extrinsic hand adopts a special actuation mechanism named Cobot [49]. It consists of one power motor and 15 steering motors that is able to output 15 channels of motions. According to specific needs, the continuously variable transmission (CVT) device (including operating motor, position sensor, power transmitting ball, operating roller and synchronizing gear sets) is able to output varying torque moment and speed.
Comparing with IAP hands, the EAP prosthetic hands are superior in compactness, dexterity, actuation manner and power. Tendon actuation is usually adopted in EAP hand since there is sufficient space in the palm allowing for more active DOFs and sensors. In addition, the actuation components outside the hands are not limited by space anymore, by which motors with greater power can be used. On the other side, considering the overall volume and weight, IAP prosthetic hands usually use small-power direct current (DC) motors and tend to adopt a pre-tightening-free mechanism in actuation. Besides, the number of the sensors and embedded chips (CPU, ROM, etc.) are largely restricted such that sufficient information about the manipulation cannot be instantly processed. This problem could be solved along with the development of advanced electronic/computer engineering. One big merit of the IAP prosthetic hands is their application flexibility for different amputation degree. The individual difference of patients requires less re-design or re-configuration procedures for IAP prosthetic hand. Thus, these hands are more likely to be standardized, commercialized and maintained.

DLR/HIT dexterous robotic hands
The HIT-I hand (Fig. 1a) is our first dexterous hand prototype developed under a collaborative effort between Harbin Institute of Technology (HIT) and German Aerospace Center (DLR) in 2001. The HIT-I hand adopts a modular design concept that all four fingers (little finger excluded) are driven by embedded motors with tendons, because of which the degree of system integration is greatly improved and the size of the hand is well controlled at that time. Position sensor and force/torque gauges are embedded thus that the hand can accomplish some multisensory hand operations. However, due to the quality of the tendons and digital level of that time, the hand only promises a comparably low compatibility and robustness.
Based on HIT-I hand, refinement work for improving the mechanic/electronic reliability and human-like appearance is proposed in the design of DLR/HIT Hand I (Fig. 1b) [50]. Each finger is modularized as three joints with three active DOFs, wherein the metacarpophalangeal (MCP) joint has two active DOFs and the proximal interphalangeal (PIP) and distal interphalangeal (DIP) joints had one active DOF (coupled through a four-bar linkage). In the metacarpal joint (TM) of the thumb, an Belt + Tendon extra DOF is provided for realizing thumb opposition, thus that the relative position between thumb and four digits can be ensured in various grasping tasks. Gears, harmonic reducer, and linkages constitute to the hand's transmission system. The actuation and control system is totally embedded and digitalized as much as possible.
This design minimizes the hand's weight and reduces the number of tendons used for driving the joints. With collaboration of Schunk company, a commercialized version of the dexterous robot hand, SAH (Fig. 1c), is also available on the market and receives many success applications. The package design of the SAH largely improves its  appearance and effectively protects the electronic system and cables within the hand. In 2008, our newest generation DRH prototype, DLR/ HIT Hand II [51], is presented with five identical fingers and a human-like curved palm, as Fig. 2 shows. To improve the hand's manipulation dexterity and operation intelligence, a total of 15 active DOFs and 140 sensors (position, force, temperature, and tactile) are integrated based on IAP. Each finger can be divided into two modulated units, 2-DOF basic joint unit and 1-DOF finger unit, within each the motors, reducers, sensors and electronic systems are totally built-in. By adopting micro brushless DC motors, timing belt, harmonic drive, and tendon coupling, the size and shape of DLR/HIT hand are significantly reduced (similar to an average mature hand). To further save space, sensors are tried to be integrated with the hand's mechanical structures, such as, the torque gauge is a transmission linkage in the basic joint, and the 3D tactile grid is an embedded layer on the finger pads. Attributed to its multimodal sensors, the DLR/HIT Hand II has a superior perception capability during various operation tasks.

Multi-fingered prosthetic hands
Since 2001, five prototypes of HIT-DLR anthropomorphic prosthetic hands (HITAPH) have been developed, as Table 3 shows. Designed based on DLR/HIT hand II, the HITAPH I-III has five fingers, and each is composed of three knuckles (2 knuckles in the thumb of HITAPH I and II). The HITAPH I-III are actuated by three DC motors, which are installed at the TM joint of the thumb, MCP joint of the index finger, and MCP joint of the middle finger, respectively. The middle finger, ring finger and little finger are co-actuated through torsional springs and linkages. Taking advantaging of the underactuation principal [52,53], the inter-finger actuation of the hand is realized through elastic components, which provides the hands with an adaptive grasp to various objects. Among them, the HITAPH III [also called anthropomorphic robot (AR) hand III] [54] makes an improvement on its human-like appearance and grasp power.
To further improve the hand's dexterity, the HITAPH IV [5] is developed with five active DOFs (that is, all five fingers are individually actuated). Attributed to advanced actuation techniques, the volume of the IAP hand is only 85 % of that of HITAPH III. The total hand weight is about 450 g. The output force at the fingertip can reach up to 10N. Curved palm and scattered finger configuration endow the hand with more anthropomorphic characters in appearance and grasping. Reducing the number of the non-standardized mechanical and electronic elements is one critical request in the design of HITAPH IV, aiming to improve the hand's interchangeability and maintainability. Meanwhile, the packaging design of the hand is also considered in the design, in connection with its actuation capacity, thermal and life design.
To further improve the thumb finger's mobility, an additional DOF is provided at the TM joint for realizing opposition. Attributed to this extra DOF, the thumb can reach to each fingertip of the other four fingers. Another big revision is the reduced number of knuckles, (two, instead of three in HITAPH IV [55]), for briefing the mechanical structure while keeping the hand's functionality and reliability. A total of six DC motors are embedded, while the actuation force at the fingertip can reach up to 12N. The worm gear, instead of bevel gears, is adopted in the MCP joints, while tendon coupling mechanism is maintained in the DIP joints. With addition to the position and force sensors, a 3-D tactile sensor [56] is designed that can measure one perpendicular force and two tangential forces applied on the fingertips. The number of the parts of the HITAPH IV is largely reduced compared with former prototypes, making it very promising in commercialization.

Challenges and future work
There are mainly two trends for developing DRHs, one is anthropomorphism-oriented and the other is task-oriented. For the first one, the robot hands are devised with much more human-like properties, as in its kinematics (hand structure, DOFs, grasping functionality, etc.), dynamics (stiffness, damping, friction, etc.), and perception capabilities (position sensors, force/torque sensors, tactile grid, slipping sensors, etc.). While for the taskoriented trend, the DRHs are designed according to some specialized tasks or environments, such as the Robonaut 2 hand and Dexhand, both for extravehicular activities on the ISS. Generally speaking, the word "anthropomorphism" is a very complicated concept including numerous influential factors that lacking anyone of them may lead to an underperformance design. Under current conditions, how to make a compromise between the hand's appearance and functionality, or how to establish a proper performance index [57] to compare design and Attributed to its thumb's dexterity, the human hand can achieve so many versatile grasps and delicate operations. Decoding the thumb's movement in respect of DOF configuration is a big challenge in robotics research. We devote to analysis the thumb's movement based on human hand anatomy and biomechanics theory, and then to give a set of appropriate DOF configurations (TM-MCP-DIP, 2-1-1, 2-2-1 [58], 3-1-1 [59]) that can be used in DRHs for achieving a variety of dexterous operations. From the point of directional dexterity, we attempt to analyze these configurations on specific manipulation tasks, which can further facilitate our selection according to different application scenarios. How to arrange the thumb on the hand is another challenge. For achieving versatile and effective grasp patterns, an efficient method needs to be proposed for appropriately positioning the thumb on the palm.
The dexterous manipulations requested by the DRHs are not only promised by its anthropomorphic shape and motion, but also its high-speed processing system (sensor measurement, data analysis, kinematic calculation, etc.) and real-time control algorithms (task interpretation, motion planning, sensory feedback, etc.). Because of massive data calculation and interchange, selection of an appropriate control structure and platform, highlyintegrated hardware and software hierarchy and suitable communication protocol are critical for DRH realizing a real-time manipulation. Currently, based on EtherCAT, a real-time control design and validation platform [60] has been developed, on which a large variety of algorithms, such as the impedance control strategy with coordinated multi-finger manipulation and optimized grasping forces [61], can be been verified. For achieving delicate manipulations, the DRHs need to know necessary information about its outer environment (obstacles) and the object (stiffness, size, shape, and weight) to operate. The information is generally provides by our proprioception experiences (body schema) sensed by our central neural system through a long-term, multisensory stimulation (visual, tactile, force, tension, etc.). The tactile sensor [62] has been now widely integrated into the fingertip to acquire such information as the contact status, position, and some other physical properties (stiffness, texture, etc.) about the object. Even a primary haptic sense (object shape and category) can be reconstructed by using the tactile sensor and position sensors integrated on the robot hand. However, how to effectively fuse these information into the control scheme, thus to improve the hand's operation compliance, precision and intelligence, is still an open question [63].
The main task of a prosthetic hand is to to help physically disabled people restore hand functions in living environment (ADLs). General APHs should have three main features, as human-like appearance (size, weight, textures, compliance, etc.), mobility, and perception. Besides, stateof-the-art APHs request even more dexterous operations, given very limited choice on the actuation styles and DOF configurations. How to realize a large portion of human hand functions in very low actuation cost is very ambitious in APH research. Besides, high-precision position control (such as, to nip a needle) and accurate force control (such as, to grasp a fragile cup) are both frequently required in the daily use of APHs. How to devise a smart control strategy properly working on different conditions is another question. For controlling the prosthetic hands, the surface myoelectric signals (sEMG) collected from the residual neuromuscular system (stump) are widely accepted. Traditional mode-switching methods established on EMG amplitude only give very limited functions, discrete robot-like finger movements, and unintuitive control feelings. By introducing the pattern recognition method [64], a large progress has been made; however, there is still a big gap between the research and its real application [65,66]. Intrinsic timing-varying characters of the EMG signals, environmental change (electromechanical status, temperature, moisture, sweating, etc.) of the bio-machine interface, and confounding factors (body postures, contraction strength variations, involuntary EMG activations) largely affect the long-term usage of clinical APHs. In this case, the control of APHs should consider other alternative peripheral nervous signals, such as ultrasonic signal [67], mechanomyography [68], near-infrared spectroscopy [69] and electrocorticography [70], to be used in the control channel, and multi-sensory means [71], such as vision [72,73] and tactile sense [74,75], to be used in the feedback channel. With the big progress of the worldwide scientific research on artificial cognition and brain-computer interface, a fully-embodied hand avatar controlled by our brain with utmost ease will come soon.

Conclusions
This study focuses on the introduction of the development route of intrinsic actuation dexterous hands and prosthetic hands, giving a brief overview on the current artificial dexterous hands and prosthetic hands. With the progressing of science and technology, robotic hands are gradually approximating to human hands in dexterity and perception, based on which they can finish various complicated operations in the manufacturing process, activities of daily life, and exploration of unknown environment.
Authors' contributions HL conceived the manuscript, participated in the data collection and analysis, and helped revise the manuscript. DPY carried out the survey study and drafted the manuscript. SWF helped draft the manuscript. HGC helped supervise the study and revise the manuscript. All authors read and approved the final manuscript. Institute of Technology, where he currently holds the position of an associate researcher. His current research interests include anthropomorphic robot dexterous hand, prosthetics hand, tactile sensor, and grasping.Cai HG received the B.S. degree in mechanic engineering from the Harbin Institute of Technology, in 1958. Since 1985, he has been a professor at the School of Mechatronic Engineering, Harbin Institute of Technology. Since 1997, he has been an academician of the Chinese academy of engineering. Prof. Cai successfully developed China's first welding robots and spot welding robots, and solved critical problems in the robots' trajectory control accuracy and path predictive control. His current research interests include industrial robot, space robot and control algorithms.