Thomas J. Cairnes, Christopher J. Ford, Efi Psomopoulou, Nathan Lepora
©STUTTERSTOCK.COM/GORODENKOFF
The development of robotic grippers is driven by the need to execute particular manual tasks or meet specific objectives in handling operations. Grippers with specific functions vary from being small, accurate, and highly controllable, such as the surgical tool effectors of the Da Vinci robot (designed to be used as noninvasive grippers controlled by a human operator during keyhole surgeries), to larger, highly controllable grippers like the Shadow Dexterous Hand (designed to recreate the hand motions of a human). Additionally, there are less finely controllable grippers, such as the iRobot-Harvard-Yale (iHY) Hand or iRobot-Harvard-Yale (IIT)-Pisa SoftHand, which, instead, leverage natural motions during grasping via designs inspired by observed biomechanical systems. As robotic systems become more autonomous and widely used, it is becoming increasingly important to consider the design, form, and function of robotic grippers.
While the designs of robotic grippers are varied, there are common descriptors that can be used to separate them into groups that are informative about their operating principles (Fig. 1).
Fig 1 Examples of robotic grippers. (a) The Yale OpenHand Model O is a three-fingered anthropomimetic hand open sourced from customizing the iHY Hand, developed at Yale University (Ma and Dollar, 2017). (b) The SDM Hand is a four-fingered anthropomimetic gripper utilizing compliant passive joints, developed at Harvard University (Dollar et al., 2007). (c) The Shadow Dexterous Hand is a fully actuated five-fingered anthropomorphic hand developed by the Shadow Robotics Company. (d) The Pisa/IIT SoftHand is a five-fingered anthropomorphic hand utilizing rigidÐpassive links and a single actuator to achieve grasps; this version has a biomimetic tactile fingertip developed at Bristol Robotic Laboratory (Lepora et al., 2021). (e) A gecko-inspired adhesive gripper. This biomimetic gripper was developed at the University of California to utilize electrostatic forces to adhere to objects (Glick et al., 2018). (f) An inchworm-inspired soft gripper developed in at Guangdong University in China. The claws mimic the muscle structure of an inchworm for climbing robots (Li et al., 2017). (g) A universal jamming gripper. This gripper, developed at Cornell University, was designed to use the unique jamming properties of granular materials to grasp a range of objects (Brown et al., 2010). (h) This origami-inspired folding gripper, developed at the Harbin Institute of Technology, uses rigidÐflexible links to grasp (Liang et al., 2021).
Anthropomimetic grippers aim to replicate the motion of the human hand and are inspired by the success of human dexterity to provide a familiar reference point for intended motion. Alternatively, biomimetic but nonanthropomimetic grippers take inspiration from other biomechanical structures to inform their motion and control. Lastly, there are nonbiomimetic grippers. These have not been investigated as much as grippers inspired by the huge range of diversity in natural organisms, but this category still provides examples of effective grippers.
There are four general types of actuation in modern robotic grippers: pneumatic grippers driven by compressed air, hydraulic grippers driven by compressed fluid, electric grippers using motors, and vacuum grippers relying on some form of suction.
As mentioned earlier, it is important to consider grippers in two overarching categories: fully actuated and underactuated grippers. These terms indicate the degree of control that can be expected from a gripper. A fully actuated gripper has every degree of freedom (DOF) directly controlled by a motor or other actuation mechanism. Conversely, an underactuated gripper has only some DOFs actively controlled, while the rest react passively according to the mechanical structure of the gripper.
The simplest robotic grippers are fully actuated 2-DOF grippers, often referred to as parallel jaw grippers. The earliest was mounted on the Stanford Arm (1969) and was originally hydraulic but later operated through dc motors. In the following decades, industrialized adaptations of parallel jaw grippers were used widely in various picking and packing applications, and they are still widely used today because of they are capable and straightforward to use.
This section focuses on anthropomimetic grippers, i.e., grippers that use opposing multijointed fingers to form grasps similar to those humans may perform.
An early example of an anthropomimetic gripper was the Graspar Hand (1996), a three-fingered hand composed of mostly passive joints driven by an internal pulley system. The Schunk and Barret three-fingered hands (2007) were rigid and fully actuated. This made them very controllable; however, they had issues when applied to a wide range of objects, as their motions required precise control of every joint, which made them difficult to control. These hands relied almost exclusively on power grasps—i.e., grasps where the hand seeks to completely envelope an object—as opposed to precision grasps, where contact with the object is made with the fingertips of the hand.
A desire for smooth anthropomimetic motions to reduce control complexity led to underactuated hands utilizing compliant passive joints. The four-fingered SDM = Shape Deposition Manufacturing Hand (2010) was constructed almost entirely of soft compliant materials, all driven by a single motor. Its design focused on creating anthropomimetic power grasps that maximize contact between the fingers and an object, with the finger configuration resembling two soft, compliant parallel jaw grippers. Passive, dislocatable joints enabled hand designs that could produce natural and stable power grasps. In this way, one joint could stop moving after making contact with an object, and the connected joints in the chain could continue to move with the actuation until they contacted the object. This produces a grasp that maximizes the contact area without prior knowledge of the object, which is a property commonly referred to as being adaptive.
There then followed hands optimized for power grasps using passive joints, such as Robotiq’s Adaptive Gripper (2013) and Kinova’s Jaco (2013). However, those adaptive grippers were still fairly rigid in construction and had not capitalized on the strengths shown by the SDM Hand.
Consequently, the iHY Hand (2013) was designed for a range of specific tasks to be achieved, not just grasping objects. This additional functionality required the ability to perform precision and power grasps, requiring individually actuated fingers and enough rigidity to perform tasks like flicking switches. The iHY utilized five motors but was underactuated, as it still had 8 DOF, making use of a rigid–flexible construction for compliant fingers. The iHY Hand was well suited for research applications, being simply fabricated, and was open sourced as the GRABLAB Model-O while additionally being commercialized as the Reflex Hand from RightHand Robotics.
Full anthropomorphism is achieved when a hand is five fingered; for example, the Pisa/IIT SoftHand (2014) is a semisoft, underactuated gripper intended to be easily controllable while also applicable to a variety of grasping tasks. A single motor results in the hand only having one grasping mode (a whole-hand enveloping grasp) with tendons routed through the hand; this results in a grasp that follows a “postural synergy,” which closely mimics the motion of a human enveloping grasp. The rigid–flexible construction, dislocatable joints, and anthropomimetic design result in a gripper that is highly adaptive yet simple to control. More recent versions have included additional control from a second postural synergy, and an easily fabricated 3D-printed version of the SoftHand has been released, known as the BRL/Pisa/IIT SoftHand (2022).
At the opposite extreme, in terms of actuation, is the Shadow Dexterous Hand, a highly actuated anthropomorphic gripper demonstrated as capable of replicating human grasps during teleoperation due to its highly actuated nature. This was demonstrated in a teleoperation task where an operator solved a Rubik’s Cube using a pair of the hands.
Anthropomimetic grippers offer a familiarity of control and design not found in other types, most notably due to the human hand development focused on tool use and task completion. While the capabilities of other biomimetic grippers are much more specialized, anthropomimetic grippers offer a much wider range of application cases.
Traditionally, robotic grippers have been composed of rigid links and hard components, which offer precise control, better overall actuation strength, and efficient transfer of power through the system. However, natural systems make use of rigid–flexible couplings, for example, in the joints of the human body. One of the earliest examples of a soft robotic gripper utilizing rigid–flexible links is Hirose’s soft gripper (1978), which resembled an octopus tentacle and was the first gripper to passively adapt itself to an object’s shape. Such rigid–flexible links would later become an important component in some anthropomimetic grippers, such as the Pisa/IIT SoftHand.
There are also many examples of entirely compliant grasping solutions in nature, such as suction cups on cephalopod tentacles or the adhesive tongues of anteaters. Nonanthropomimetic grippers inspired by mechanisms like these have long been widely studied in the application of soft robotics, leading to grippers that depart from the soft–rigid principle. For example, NASA’s gecko gripper (2016) is based on gecko feet and used for holding flat objects in space. The microscopic hairs on a gecko’s feet generate intermolecular forces (van der Waals forces), allowing adhesion without any liquid or surface tension. NASA’s gripper operates on the same principle, making it suitable for moving and orientating large panels or objects that would otherwise require a large gripper to power grasp that would otherwise not be well suited for space applications.
As discussed earlier, these biomimetic grippers offer grasping capabilities that are highly specialized and capable of grasping and manipulating objects that anthropomimetic grippers could not. This, however, comes with its own issues regarding the design, control, and application of these grippers.
There are currently only a few examples of grippers that are not based on the human hand or other biomechanical structures in living organisms. For example, vacuum grippers used suction cups to create vacuums to adhere to an object, which are reminiscent of cephalopod suckers but were not inspired by them. Early vacuum grippers used industrial vacuums and cycling airflow to provide constant suction, but these progressed quickly to more energy efficient equivalents of vacuum gripping technology. In particular, rather than cycling air for constant suction, the vacuum gripper presses against an object’s surfaces and then has the air in the cup expelled, creating a pressure imbalance for adhesion.
An interesting example of a nonbiomimetic gripper is the universal “jamming” universal gripper, which utilizes the phenomenon that granular materials “jam”—i.e., pack together very tightly when forced together. In one design, an elastic container (like a balloon) is filled with a granular material and pushed against an object; then, vacuum suction removes the air from the balloon and jams the material together, which maintains the same shape but is much stiffer, so it can grasp the object securely.
The end effectors of surgical robots, such as the Da Vinci or Smart Tissue Autonomous Robot, are miniaturized surgical tools that can have gripping functions. For some applications, like surgical pliers, they can be considered grippers, but, in others, they are more specialized tools like scalpels.
Grippers in this category are once again specialized even further for specific applications but lend themselves to being applied and controlled easily. Vacuum grippers and jamming grippers are incredibly simple and straightforward, but they cannot accomplish the more complex tasks that anthropomimetic grippers are aiming toward.
An important capability for robotic grippers is to perceive that they are interacting with an object as they come into contact with or are holding the object. This could involve interoceptive sensing inside the hand itself due to pressure, friction, and tension in actuated components, which are analogous to the human perception of grasping through proprioceptive and other signals. This information combined with exteroceptive information, such as from vision, creates a feedback loop for interacting with objects and forms the building blocks of an automated system for a robotic gripper to interact with the world.
This separation of perception into visual and tactile components is important because grasping does not entirely concern sensing in the gripper alone: knowledge of the gripper exclusively is not enough to make intelligent decisions regarding interacting with objects or completing tasks.
Proprioception encompasses a suite of senses internal to the body that enables humans to accurately estimate the position of their own limbs, head, and trunk without being able to see them. These senses originate from muscles, tendons, and joint ligaments and detect changes in skeletal muscle contraction/stretch, body/limb position, and joint capsule stress.
The proprioceptive sense of a robotic gripper depends on the type of gripper. For example, a two-fingered gripper driven by electric motors can sense proprioceptively by using encoders in the motor, with measurements of how far a servo has rotated and how much torque is applied being common indications of effective grasping. A gripper driven pneumatically or hydraulically, however, can present challenges for integration with proprioceptive sensing. For example, the previously mentioned Graspar Hand had no proprioceptive capability at all. Simply measuring that the actuator or piston is active provides more detail, but additional sensors, such as barometers, are usually needed for detailed proprioceptive information.
Fully actuated robotic grippers, such as the Shadow Dexterous Hand, have controllable servos for each joint in the fingers. Therefore, from the encoder readings on the servos, an operator or autonomous system can ascertain the state and position of each finger.
Exteroception is the sensing of external stimuli and can be seen as a system-level capability of the entire robot rather than of the gripper alone. By utilizing cameras and depth sensors along with a model of the platform, the system can localize itself and construct a representation of its environment.
The human sense of touch is essential for our grasping and dexterous capabilities, relating specifically to sensations carried through the skin. Tactile sensing allows humans to evaluate a broad range of physical contact properties, including shape, texture, softness, temperature, and noxious stimuli (pain). It is also used in combination with proprioception to make estimates of weight and how much force is required to grasp an object. One can view touch as forming the sensory boundary between proprioception and exteroception. Therefore, when progressing toward intelligent grippers, the role of tactile sensing and the link it provides between outward and inward perception are completely crucial.
Tactile sensors range in size and complexity, but their baseline function is the same: determining whether the gripper has made contact with something. These range from simple pressure sensors that enable contact checking; to piezoelectric tactile arrays; to optical tactile sensors, such as the TacTip or GelSight sensors. Optical tactile sensors rely on a camera observing what is usually a representation of a layer of human skin. The TacTip traces white papillae on the inside that deflect and displace in response to surface stimuli. The GelSight traces shading and dots printed directly under the “skin” to observe stretch and topography changes at a small scale.
The research field of autonomous robotics considers levels of autonomy (LoAs) that can separated into six groups that delineate where the control and responsibility of an operator lie in relation to the type of autonomous system (Fig. 2).
Fig 2 Examples of LoAs. (a) LoA 0: a pair of kitchen tongs, a gripper with no input or control on the gripper side. (b) LoA 1: a Da Vinci surgical robot, teleoperated with feedback control loops to eliminate instrument shake. (c) LoA 2: an autonomous robot used in a smelting factory, which waits for human input and then performs a preprogrammed set of instructions (d) LoA 3: a factory robot assembling cars. The beginning of operation and emergency stops are still under human control. (e) LoA 4: a surface-mounted device pick-and-place system using a pneumatic gripper to assemble devices. The machine operates in a workspace separate from humans and, therefore, has control over its own safety measures. (f) LoA 5: the Port of Rotterdam is an almost entirely autonomous shipping port; automated stacking cranes perform the necessary grasping. In this situation, the system can accept tasks, fulfill them, access resources it needs autonomously, and reasonably grasp any object it will encounter.
In any area where a gripper must interact with a human being, it is generally assumed that most interactions are under human control, which can be referred to as kinesthetic guidance. This may either be through direct control by an operator or through the system being prompted by a human to act and then wait for further instructions. There must also be sufficient feedback to the human operator to allow reliable control.
A shared autonomy situation refers to tasks where a robot may operate autonomously, but final control of the system rests with a human. For example, a prosthetic robot hand could be LoA 1 if the hand is capable of regulating itself to close around an object but requires human instruction and decision making to do so.
The Da Vinci surgical robot is directly controlled by a human operator, who may be in the same room, via teleoperation. Such surgical robots have closed-loop control systems that offer the stable control of surgical tools. This capability differentiates them from being under direct kinesthetic guidance. Likewise, solutions for hazardous/dangerous environments often involve teleoperation in a similar manner.
A progression in the LoA would be, for example, a warehouse scenario where a human inputs an item request for a robot to collect an item from some storage bin. The arm maneuvering to the bin, identifying the item, determining how to grasp it, and retrieving it would be automated. However, a human still controls when the task is undertaken as well as the details of the task and can make a decision to stop or cancel the task. Such operations are at LoA 2.
Giving such systems, the capability to initiate tasks autonomously in a self-determined order to just retrieve and deliver items would give the robot sufficient autonomy to reach LoA 3. The human operation is limited to interventions during emergencies and ceasing operations. In such scenarios, a robot can handle some safety features itself but is still ultimately dependent on a human operator. Many driverless cars can be viewed as LoA 3.
A full LoA occurs when there is no human involvement at all. Achieving this LoA does not imply that the system is safe; it just means that the system is responsible for its own safety measures.
Consider, for example, the automated warehouse described previously and imagine that a section of the warehouse has been closed off so that the robot can operate independently. This section contains storage bins and containers for certain item types that the system is capable of handling autonomously. The robot has lists of tasks that it performs without supervision, making all relevant decisions regarding the collection of items and the item order. This is a fully autonomous system that has been limited to operate in a specific scenario: that is LoA 4. This scenario separation may be because of the operational limits of the system itself; for example, the grippers on the arms may not be able to handle objects larger than the grippers, so the system is constrained to prevent it from attempting tasks at which it may fail.
If, instead of limiting the item types or section of the warehouse where system can operate, it is allowed to operate for all items types in the entire warehouse, then an LoA 5 system has been created. Fully automated shipping yards are good examples of LoA 5 systems, as they can handle all standard shipping containers and do so without supervision from a human.
The ethical considerations of a robotic gripper being used on or near humans are important but complicated by the gripper being just one component of an entire robotic system. In many regards, safety standards that apply to systems like robotic arms can also be applied to robotic grippers.
In the early development of workplace and factory robots, the simplest way to ensure human safety was to have a clearance zone that a person should not enter or to isolate the robot entirely within a cage. However, separating humans and robotics is proving to be limiting in modern applications of robotics.
The development of collaborative industrial robots has brought them into human workspaces, demanding a much higher degree of safety to allow close proximity to humans. Safety measures can involve monitored safety stops, i.e., a stopping operation when a human enters the workspace; speed and separation monitoring that causes the robot to slow operations when humans are nearby; and power limiting. Power and force limiting require either the monitoring of actuation mechanisms or limiting through inherent design. For example, soft grippers in collaborative scenarios offer a straightforward and implicit method of providing operational safety limits.
Since the primary function of a gripper is to secure the position and orientation of an object in relation to the robotic platform, a reasonable safety standard would be to require that a gripper should not change state upon losing power. A gripper dropping something fragile, heavy, or sharp during an emergency stop or power cut could be harmful and/or costly.
Developing grippers that are more capable of performing like human hands—or, perhaps, are superior to humans for specific situations/tasks—will lead to the development of autonomous platforms that can perform tedious/dangerous tasks in place of humans. How this affects the job market is a matter of debate, as the shifting economic growth and recession of the past decade have made it difficult to evaluate the impact that automating labor has actually had. An important question is whether increasing the automatization of human physical labor would simply remove jobs or create more skilled and fulfilling work in other areas, such as the installation and maintenance of those robots.
Whatever the answer, future robotic technologies should be used responsibly and to improve human society. One clear responsible use in this scenario would be to target jobs and tasks that are considered dangerous or tedious so that the humans doing those jobs can be safer and focus on more productive or creative work.
A key problem for future of robotic grasp is how to grasp intelligently. Modern robotic hands have become highly dexterous and articulated, but the challenge of controlling them remains unsolved.
The use of more sophisticated and human-like sensing in a robotic gripper will lend itself to enabling intelligence, but the also creates a need to process and interpret more data on top of a potentially already very complex gripping system. For example, during the development of the iHY Hand in the early 2010s, highly articulated dexterous hands were available and had the same issues that they still face now with regard to uptake and application in industrial spaces. Simpler grippers are currently a more accessible option, but they have their own issues with the integration of a greater sensing capacity into structurally simple hands, often requiring a complete change in design to be compatible with a given sensing technology; for example, vision-based tactile sensors are difficult to integrate into soft grippers.
So-called universal grippers offer a cheap and effective way to grasp many objects; they contain a particle material enclosed in a fine rubbery membrane for enclosing around objects. However, they can fail to offer a reliable solution for many tasks, such as grasping in clutter or object manipulation. Likewise, a gripper designed for a specific task, such as two-fingered grippers for picking and placing simple objects, may fail to work outside of that task. So far, the workaround has been to change the tasks given to certain grippers or to limit their application space. While this separation has enabled progress, it has led to a huge diversity in gripper types. The development of a gripper that can be used universally and controlled accurately—which, to some extent, the human hand solves—is still ongoing.
Unknown, unstructured, and cluttered environments are problematic for any robotic system. Finding ways to explore spaces and deal with clutter is not only a challenging perception task but also requires intelligent planning. The system needs to be capable of differentiating between unimportant clutter that can be ignored or displaced and dangerous disturbances that must be safely dealt with. High proprioceptive and tactile-sensing capabilities are needed for gripping systems that can operate in such environments.
In terms of the LoAs as we discussed earlier, at present, LoA 5 systems exist in very predictable and structured environments, such as shipping yards.
While collaborative robots are beginning to make more of an industrial appearance, their uptake remains slow. From a legal and practical standpoint, it is currently far simpler and safer for a company to isolate its robots from human workers. It remains an important challenge to improve the productivity of humans working alongside robots while maintaining their safety. As grippers develop toward more intelligent grasping alongside human workers, their operational safety will remain a key consideration.
The development of robotic grippers has advanced enormously over the past few decades, which is exhibited by the diverse range of manipulator designs and applications. The next steps for the use of robotic grippers need to address the fact that highly articulated hands are still very costly, are difficult to control, and have not seen as much application as their potential suggests. Currently, cheaper and less complex grippers are finding more industrial uses. The developers of dexterous robot hands need to make their grippers easier for commercial users to utilize; otherwise, those users will stick with current options and use case scenarios. Moreover, further advancements in the perceptual and control capabilities of simpler grippers are much-needed stepping stones to integrating intelligent gripping technologies into fully autonomous systems.
• B. Young, “The bionic-hand arms race: High-tech hands are complicated, costly, and often impractical,” IEEE Spectr., vol. 59, no. 10, pp. 24–30, Oct. 2022, doi: 10.1109/MSPEC.2022.9915629.
• L. Odhner et al., “A compliant, underactuated hand for robust manipulation,” Int. J. Robot. Res., vol. 33, no. 5, pp. 736–752, Feb. 2014, doi: 10.1177/0278364913514466.
• A. Dollar and R. Howe, “The highly adaptive SDM hand: Design and performance evaluation,” Int. J. Robot. Res., vol. 29, no. 5, pp. 585–597, Feb. 2010, doi: 10.1177/0278364909360852.
• B. Ward-Cherrier et al., “The TacTip family: Soft optical tactile sensors with 3D-printed biomimetic morphologies,” Soft Robot., vol. 5, no. 2, pp. 216–227, Jan. 2018, doi: 10.1089/soro.2017.0052.
• R. Ma and A. Dollar, “Yale OpenHand Project: Optimizing open-source hand designs for ease of fabrication and adoption” IEEE Robot. Autom. Mag., vol. 24, no. 1, pp. 32–40, Mar. 2017, doi: 10.1109/MRA.2016.2639034.
• A. Dollar and R. Howe, “The SDM hand as a prosthetic terminal device: A feasibility study,” in Proc. IEEE 10th Int. Conf. Rehabil. Robot, Noordwijk, The Netherlands, 2007, pp. 97–983, doi: 10.1109/ICORR.2007.4428542.
• N. Lepora et al., “Towards integrated tactile sensorimotor control in anthropomorphic soft robotic hands,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), May 30–Jun. 5, 2021, pp. 1622–1628. [Online] . Available: https://entrepreneurship.ieee.org/session/2021-icra/, doi: 10.1109/ICRA48506.2021.9561350.
• P. Glick, S. Suresh, D. Ruffatto, M. Cutkosky, M. Tolley, and A. Parness, “A soft robotic gripper with gecko-inspired adhesive,” IEEE Robot. Autom. Lett., vol. 3, no. 2, pp. 903–910, Apr. 2018, doi: 10.1109/LRA.2018.2792688.
• M. Li et al., “Development of a bio-inspired soft gripper with claws,” in Proc. IEEE Int. Conf. Robot. Biomimetics (ROBIO), Macau, Macao, 2017, pp. 828–833, doi: 10.1109/ROBIO.2017.8324520.
• E. Brown et al., “Universal robotic gripper based on the jamming of granular material,” Proc. Nat. Acad. Sci., vol. 107, no. 44, pp. 18,809–18,814, Oct. 2010, doi: 10.1073/pnas.1003250107.
• D. Liang, Y. Gao, H. Huang, and B. Li, “Design of a rigid-flexible coupling origami gripper,” in Proc. IEEE Int. Conf. Robot. Biomimetics (ROBIO), Sanya, China, 2021, pp. 1283–1287, doi: 10.1109/ROBIO54168.2021.9739578.
Thomas J. Cairnes (pd18523@bristol.ac.uk) is a Ph.D. candidate at the University of Bristol, BS8 1QU Bristol, U.K. His main area of research is the grasping and perception of soft, deformable objects with underactuated hands.
Christopher J. Ford (chris.ford@bristol.ac.uk) is a Ph.D. candidate at the University of Bristol, BS8 1QU Bristol, U.K. His research interests are focused on the use of tactile perception and feedback for the real-time grasping and manipulation of multifingered robotic hands.
Efi Psomopoulou (efi.psomopoulou@bristol.ac.uk) is an assistant professor in the Department of Engineering Mathematics at the University of Bristol, BS8 1QU Bristol, U.K. Her research interests include physical robot interaction, grasping, and manipulation.
Nathan Lepora (n.lepora@bristol.ac.uk) is a professor of robotics and artificial intelligence in the Department of Engineering Mathematics at the University of Bristol, BS8 1QU Bristol, U.K. His research interests include robot dexterity, biomimetics, and tactile sensing.
Digital Object Identifier 10.1109/MPOT.2023.3236143