Latest / Embodied AI 101 / REL Hand: A Mechanical Shortcut to More Human-Like Humanoid Dexterity
Transcript
- REL Hand: A Mechanical Shortcut to More Human-Like Humanoid Dexterity. Design and evaluation of a tendon-and-linkage hybrid-driven humanoid dexterous hand • Ke Li, Fanlin Meng, Linjie Liu, Raviraj Nataraj, Ning Sun, and Na Wei • Shandong University, Stevens Institute of Technology, Shandong Normal University, and Qilu Hospital of Shandong University • Scientific Reports • 2026. The newest generation of humanoid manipulation research is dominated by learning: vision-language-action models, reinforcement learning, teleoperation, diffusion policies, and large-scale simulation. But none of those methods can escape the physical interface through which a robot acts. If the hand has tendon slack, poorly coupled joints, insufficient force, or weak state estimation, the policy inherits those problems as hidden dynamics. The REL Hand addresses this hardware layer directly. It is a five-digit, approximately human-scale robotic hand with 20 joint degrees of freedom, 15 of them actively driven, that combines tendon transmission, rigid linkages, passive spring return, adjustable tendon pretension, fingertip force sensing, and phalange-mounted inertial sensing. Its central idea is not merely to place tendons and linkages in the same device. The more important contribution is to allocate them to different motions and route them so that flexion and finger spreading remain largely decoupled. The paper was published online on August 4, 2026, after being received on April 18 and accepted on July 22. As of August 9, the publisher identifies the available manuscript as an unedited Article in Press, meaning copyediting and final production may still correct terminology or presentation errors. That caveat matters here because some mechanical labels in the manuscript are not always used consistently. The short version: REL Hand is a thoughtfully integrated grasping platform rather than a breakthrough in high-force actuation or autonomous in-hand manipulation. Its strongest ideas are geometric transmission decoupling, serviceable tendon pretensioning, human-like kinematic structure, and an evaluation that connects mechanism design to actual object grasping. The Physical Problem Behind Dexterous-Hand Learning. Tendon-driven and linkage-driven hands offer almost complementary advantages. Tendons are easy to route through crowded geometries. They allow an actuator to be separated from the joint it drives, and they can pass through several articulated segments without requiring a rigid transmission chain. That makes them attractive for human-like hands, where many axes must fit inside narrow fingers. Their disadvantages are equally familiar: stretch, creep, friction, hysteresis, slack, uncertain pretension, and coupling caused by tendon path-length changes as neighboring joints move. Rigid linkages provide more deterministic geometric transmission. They can carry relatively high loads, tolerate impacts, and avoid some of the compliance and maintenance problems of tendons. But linkages consume volume, have fixed motion paths, can collide with neighboring components, and become difficult to package when motion must be transmitted across multiple finger segments. The REL Hand is based on the premise that neither transmission should be asked to do everything. Tendons are used where flexible routing is most valuable—primarily flexion and extension—while linkages handle the abduction and adduction that spread the four fingers laterally. The tendon paths are then placed relative to the linkage axes so that operating one transmission does not substantially perturb the other. This is a useful distinction. A mechanism can contain both tendons and linkages and still be a poor hybrid if every movement changes every transmission’s geometry. A genuinely useful hybrid should simplify the input-output map seen by the controller. In other words, the mechanical Jacobian should not vary unnecessarily just because a neighboring actuator moved. That point connects directly to learning-based control. If commanding index-finger abduction also shortens its flexor tendon, a policy must learn that cross-axis effect, including how it changes with tendon wear and pretension. Mechanical decoupling removes part of that burden before training begins. Anatomy of the REL Hand. The hand consists of five modular digits and a palm. Each digit has four joint degrees of freedom but only three motors. Across the complete hand, that produces 20 mechanical degrees of freedom, 15 active coordinates, and five passive coordinates. For each of the four fingers, the actively commanded motions are metacarpophalangeal flexion and extension, metacarpophalangeal abduction and adduction, and flexion of the coupled distal chain. The proximal and distal interphalangeal joints do not receive fully independent actuators; one follows through the underactuated distal mechanism. The thumb similarly has three active coordinates and one passively coupled coordinate, with two motions at the carpometacarpal joint and a coupled metacarpophalangeal–interphalangeal flexion chain. That layout is a deliberate compromise between anatomical fidelity and control dimensionality. It preserves finger spreading and thumb opposition—motions that strongly affect grasp class—while avoiding the cost and packaging difficulty of independently actuating all 20 joints. | Design property | REL Hand | |---|---| | Digits | Five | | Mechanical joint DoFs | 20 | | Active DoFs and motors | 15 | | Passive, coupled DoFs | 5 | | Motors per digit | 3 | | Flexion transmission | Steel tendons | | Extension | Joint-mounted return springs | | Finger abduction/adduction | Parallelogram linkage | | Fingertip force sensors | 5 | | Phalange posture modules | 10 | | Prototype mass | Approximately 600 g |. The joint ranges are broadly anthropomorphic. Finger distal joints and the thumb interphalangeal joint reach roughly 80 degrees of flexion; finger proximal joints and the thumb metacarpophalangeal joint reach around 100 degrees; finger metacarpophalangeal flexion reaches approximately 105 degrees. Finger abduction and adduction spans about minus 18 to plus 18 degrees, while thumb opposition through the carpometacarpal abduction axis reaches roughly 90 degrees. Most segment lengths were selected from adult male anthropometry. The conspicuous exception is the little finger, which was made longer and wider than the corresponding human digit to accommodate actuators and sensors. This is a reminder that “anthropomorphic” is always qualified by packaging constraints. The hand reproduces the broad proportions and kinematic organization of a human hand, but it is not a mechanically exact anatomical replica. The prototype is manufactured primarily from 3D-printed white resin and weighs about 600 grams with its mechanical and electrical components installed. That mass is comparable to a human hand, although equality in mass does not imply equality in inertia: actuator placement, gearbox mass, and the distribution of material along the fingers and palm all affect the load seen by a humanoid wrist. How the Hybrid Drive Actually Works. Every digit contains three small geared DC motors. The manuscript identifies the actuators as GA12-N20 motors with 298-to-one planetary gearheads. One motor drives the coupled distal flexion mechanism. In a finger, that means the proximal and distal interphalangeal chain; in the thumb, it means the metacarpophalangeal and interphalangeal chain. A second motor drives flexion at the base joint. A third controls finger spreading or thumb opposition. Flexion is produced by pulling a 0.3-millimeter, 7-by-7-strand steel wire rope. Extension is not actively driven by an antagonistic motor or second tendon. Instead, springs embedded around the joints restore the digit toward extension when flexor tension is released. This architecture reduces actuator count and simplifies routing, but it also establishes an asymmetric actuation model. Flexion is actively force-controlled through the tendon, while extension depends on passive spring torque. Such a hand can close firmly and reopen reliably, but it cannot independently prescribe arbitrary bidirectional torque at every joint. That is often acceptable for grasp acquisition, where flexion against an object is dominant, but it is more limiting for dynamic in-hand manipulation and explicit impedance shaping. For the four fingers, abduction and adduction are generated by a parallelogram linkage. A motor rotates a shaped input wheel, and the linkage carries that motion to the finger base. A parallelogram is a sensible choice here because it constrains the orientation relationship between the input and output links while packaging the drive within the palm. The thumb is structurally different. Its opposition axis is connected more directly to the third motor rather than using the same finger linkage. That reflects the thumb’s distinct placement and workspace. Geometric Decoupling Is the Core Contribution. The main transmission problem appears where the tendon-driven flexion axis crosses the linkage-driven abduction axis. Suppose the flexor tendon passes beside, rather than through, the center of the abduction joint. When the finger spreads laterally, the tendon’s entry and exit points move relative to one another. The tendon path becomes longer or shorter even though the flexion motor has not moved. The result is unintended flexion, tendon slack, additional tension, or some combination of all three. REL Hand routes the flexion tendon through the rotational center of the abduction axis. If the cable intersects that center, lateral rotation changes the cable’s direction but ideally not the relevant path length. In the manuscript’s geometric analysis, the cable length before and after abduction remains constant when the tendon passes through the center. Moving it off-axis produces an angle-dependent path-length change. This is a mechanical form of coordinate separation. Instead of estimating the cross-coupling online and compensating in software, the design suppresses it geometrically. Of course, nominal geometric decoupling is not the same as perfect physical decoupling. Tendon compliance, guide friction, pulley radius, resin deformation, assembly tolerances, and gearbox backlash can all reintroduce cross-axis effects. A particularly valuable follow-up experiment would measure a complete empirical coupling matrix: command each motor independently and record displacement, force, and tension changes at every other joint. Even with those qualifications, the design principle is broadly reusable. Place rigid transmissions on axes where force and constrained geometry matter most; use tendons where flexible routing is necessary; and route those tendons through the instantaneous centers of unrelated rotational axes whenever possible. Tendon Pretension as a First-Class Design Variable. Tendon-driven hands often treat pretension as a workshop calibration detail. This paper elevates it into both a mechanical feature and an experimental variable. The REL Hand incorporates a compact preload mechanism inside the distal part of each digit. The mechanism consists of a grooved pulley attached to a screw held by embedded nut inserts. Rotating the screw turns the pulley, winds the tendon, and increases its initial tension. It can therefore compensate for installation variation, wire settling, or slack accumulated during use. The recommended setup procedure places the palm on a horizontal reference plane and commands the motors to their zero positions. The phalange sensors are then used to check the physical posture. If the measured deviation exceeds one degree, the tendon preload is adjusted until the error falls below that threshold. Mechanically, this is a simple solution. Operationally, it is significant. Slack creates a dead zone in which a motor rotates without producing joint motion. It also makes the mapping from actuator position to joint posture history-dependent. For position controllers and learned policies alike, removing that dead zone improves repeatability. The grasping experiments reinforce this point. With the preload mechanism engaged, the authors report performance improvements by factors of 1.9 for food objects, 1.7 for kitchen objects, and 4.7 for tools. The especially large tool result suggests that grasps requiring sustained distal closure or leverage are highly sensitive to tendon condition. Those ratios should not be interpreted without the corresponding absolute rates and trial counts: a large multiplicative improvement can begin from a low baseline. But the qualitative result is compelling. Mechanical calibration was not a small correction around an already solved grasp. For some object classes, it determined whether the hand functioned effectively at all. The unresolved issue is automation. A humanoid deployed for hours in a warehouse or home should not require frequent manual screw adjustment. The natural next step is to integrate tendon-tension measurement and either a motorized tensioner or a mechanically self-adjusting element. Pretension could then become an observed and regulated state rather than a maintenance assumption. Sensing and Embedded Control. The paper presents the REL Hand as an integrated mechatronic system rather than a mechanism connected to laboratory instrumentation. A central master board handles sensor acquisition, processing, and actuator commands. It communicates with a PC over UART, receives phalange measurements over an I²C bus, samples the force-sensitive resistors through analog channels, and sends commands to a separate motor-driver board. Each fingertip contains a thin-film force-sensitive resistor. These five sensors provide a basic indication of contact and normal loading. They are useful for confirming grasp closure and implementing threshold-based force regulation, but they do not provide contact position, distributed pressure, shear force, torsional slip, or local surface geometry. In modern dexterous-learning terminology, this is contact sensing, not rich tactile perception. The hand also includes ten custom modules that the paper calls inertial measurement units. Each uses an ADXL345 three-axis accelerometer, with two differently shaped circuit boards designed to fit recesses in the finger segments. The raw acceleration is filtered using a second-order Butterworth low-pass filter with a five-hertz cutoff, and joint or segment inclination is estimated from the direction of gravity. Strictly speaking, the described module is an accelerometer-based inclinometer rather than a complete IMU, since no gyroscope or magnetometer is identified. That distinction matters dynamically. When the hand is stationary or moving slowly, gravity dominates and provides a useful orientation reference. During fast motion, however, translational acceleration is mixed with gravity. The five-hertz filter suppresses gearbox vibration and sensor noise, but it also limits useful bandwidth. For quasistatic grasping, this may be entirely adequate. For throwing, catching, rapid reorientation, tool use, or contact-rich in-hand manipulation, it would be preferable to combine joint encoders, gyroscopes, motor current, tendon tension, and tactile feedback. The present sensors provide a practical foundation for closed-loop grasping, but not yet a high-bandwidth state estimator for dynamic dexterity. How the Hand Was Evaluated. The experimental program covers five complementary questions:. Can the digits generate useful contact force?. Can the hand reproduce a broad grasp taxonomy?. Does its motion resemble human hand motion?. Can it grasp standardized everyday objects?. Do pretensioning and embedded posture sensing materially improve operation?. That range is a strength. Many hand papers stop after reporting workspace, actuator torque, or a gallery of successful grasps. Here, the authors connect transmission design to force, posture, morphology, object interaction, and sensor operation. Fingertip Force. Across the digits, the reported average fingertip force is approximately 4.9 newtons. This is enough for many lightweight household objects and precision grasps, but it does not make REL Hand a high-force industrial end effector. It is also important not to multiply 4.9 newtons by five and call the result grip force. Object-level grasp strength depends on posture, contact normals, friction, moment arms, thumb placement, and how individual digits share the load. The force result should instead be read as confirmation that the compact tendon and gearbox arrangement transmits practically useful force to each fingertip. Future characterization should add force-position curves, force bandwidth, stall temperature, hysteresis, per-digit variance, and sustained-load tests. Grasp Taxonomy Coverage. The hand reproduced all 16 representative grasp configurations from the Cutkosky taxonomy. That includes a mixture of power and precision configurations and demonstrates that the available joint ranges, thumb opposition, and finger abduction are sufficient to form the canonical shapes. Taxonomy coverage is evidence of grasp versatility, but not by itself of manipulation dexterity. Forming a precision pinch does not establish that the hand can roll an object between fingertips, rotate it without arm motion, reposition contacts, recover from slip, or transition smoothly between grasps. The result is best classified as evidence of broad potential and grasp dexterity. Human-Motion Similarity. The authors compare REL Hand with a commercial OHand baseline using trajectory correlation and morphological similarity to human hand motion. REL Hand reportedly improves the two measures by 8 percent and 17.7 percent, respectively. This is one of the more strategically important results for embodied AI. A hand that follows human-like trajectories and arrives at human-like shapes should require less aggressive retargeting when demonstrations come from gloves, video, optical tracking, or human teleoperation. Retargeting can never be eliminated—human and robotic joint spaces are still different—but morphological similarity reduces one source of mismatch. The result nevertheless needs careful interpretation. The reported percentages are comparative improvements, and their significance depends on the absolute metric values, motion set, calibration, participant variation, and whether “percent” means a relative change or a difference in normalized score. They establish that REL Hand was closer under the authors’ protocol, not that its kinematics are interchangeable with a biological hand. YCB Object Grasping and the Preload Ablation. Object experiments used items from the Yale–CMU–Berkeley object set, covering common food packages, kitchen objects, and tools. The paper reports high grasping success for ordinary daily objects and, crucially, compares performance with and without effective tendon preload. Using recognizable benchmark objects is useful because it makes grasp scale and geometry easier to compare across systems. Still, hardware-level object grasping should not be conflated with autonomous end-to-end manipulation. Unless object pose, visual uncertainty, arm approach, disturbances, and recovery are randomized, the experiment primarily validates the hand’s ability to hold objects after an appropriate grasp has been selected. The preload comparison is arguably more informative than the absolute success gallery. It provides a causal link between one proposed mechanism and task performance. Rather than merely showing that the adjustment screw exists, the study demonstrates that tendon condition changes grasp outcomes substantially. Where REL Hand Sits Among Other Hybrid Designs. REL Hand is not the first robotic hand to combine tendons, linkages, and springs. The MCR-Hand family is a particularly relevant predecessor. MCR-Hand II combined linkage and tendon transmission in a compact anthropomorphic architecture, using kinematic modeling to characterize the hybrid mechanism. The REL paper describes it as a high-degree-of-freedom design whose transmission arrangement can induce coupled finger motion, requiring additional control compensation. MCR-Hand III extended that line toward an affordable, 3D-printed hand with linkage, spring, and tendon compliance. Its reported objectives included human-scale packaging, roughly seven-newton fingertip capability, sub-500-gram hand mass, and a total cost below about 800 US dollars. It also explored object-stiffness identification and adaptive grasping. A separate 2026 finger-unit study also combines tendon and linkage transmission in a three-degree-of-freedom index finger. That design emphasizes load-bearing capacity, robustness, modular manufacture, and a reported fingertip force of 25.3 newtons, but it is fundamentally a single-finger mechanism rather than a complete sensorized five-digit hand. REL Hand’s distinguishing contribution is therefore not the mere existence of hybrid transmission. It is the full-hand integration of 15 actuators, geometric cross-axis decoupling, fingertip pretension adjustment, force sensing, phalange posture sensing, and evaluation against both human motion and standardized object grasps. What the Paper Does Well. The first strength is mechanical coherence. The hybrid architecture is built around a specific transmission problem: obtaining high joint count in a compact hand without allowing the linkage-driven lateral axis to perturb tendon-driven flexion. The proposed cable routing addresses that problem at its physical source. The second is maintainability. Tendon systems are often demonstrated immediately after careful laboratory assembly, when tension is ideal. REL Hand anticipates that tendons loosen and provides an accessible means of restoring their operating point. Third, the design makes a sensible use of underactuation as a biomechanical prior. Independent finger-base flexion and abduction preserve the variables that most strongly determine grasp choice, while the distal joints follow a lower-dimensional closing synergy. That reduces actuator count without collapsing the hand into a simple one-motor adaptive gripper. Fourth, the work is unusually complete at the prototype level. The motors, transmissions, sensors, processing boards, printed structure, and control interface are integrated into one approximately human-scale device. The experimental program then progresses from isolated fingertip mechanics to grasp shape and object interaction. Finally, the preload ablation is good experimental practice. It tests whether a design feature contributes to the claimed task-level benefit instead of treating every successful grasp as validation of the entire system. What Remains Unproven. The largest gap is between grasping and dexterous manipulation. REL Hand demonstrates grasp repertoire and object retention, but the headline results do not establish sustained in-hand rotation, finger gaiting, controlled sliding, regrasping, or disturbance recovery. Passive coupling in the distal joints may become restrictive precisely in tasks that require those capabilities. The second concern is sensing. One force-sensitive resistor per fingertip is too sparse for detailed contact reasoning, while accelerometer-derived inclination is vulnerable to dynamic acceleration. The hand can observe whether it is approximately closed and touching something, but it cannot yet reconstruct a rich contact state. Third, the reported 4.9-newton average compresses potentially important variation. The thumb, little finger, and index finger have different geometry and manipulation roles. Per-digit force curves, hysteresis loops, repeatability, motor current, and thermal behavior would tell us more about how the hand performs under sustained use. Fourth, manual pretensioning is effective but operationally expensive. The requirement to check the zero posture and adjust tendons when the error exceeds one degree may be reasonable in a laboratory. It is less attractive for a fleet of general-purpose humanoids. The large performance gains from preload make automated condition monitoring more urgent, not less. Fifth, the prototype is printed in resin and uses small planetary gearmotors, springs, tendons, screws, and bonded pulleys. The paper’s headline evaluation does not establish long-cycle durability, tendon life, impact tolerance, lubricant sensitivity, or replacement time. These factors often separate an effective research hand from a deployable humanoid component. Finally, the kinematic comparison uses one commercial hand as the principal baseline. Hardware comparisons are difficult because controller tuning, calibration, motion scaling, and retargeting quality can influence trajectory metrics as much as morphology does. A stronger benchmark would compare several hands under the same human-motion dataset and optimization pipeline. Implications for Humanoid Robotics and Embodied AI. The most immediate implication is that mechanical decoupling can reduce learning complexity. A policy trained in simulation normally assumes a reasonably stable relationship between an action and the resulting joint motion. Tendon slack and cross-axis coupling violate that assumption. They create state-dependent delays and hysteresis that may differ between simulation and hardware or evolve over the robot’s lifetime. REL Hand’s routing strategy removes one predictable source of coupling, while its pretension mechanism constrains another. The likely consequence—an inference rather than a result directly tested in the paper—is a cleaner action-to-motion map and a smaller sim-to-real gap. The 15-active-coordinate structure is also well matched to human demonstration. It retains separate finger spreading, base flexion, and thumb opposition while encoding distal closure as a mechanical synergy. For imitation learning, this creates an action space that is more expressive than a conventional adaptive gripper but smaller than a fully actuated anatomical model. Morphological similarity could further improve teleoperation and video-based retargeting. If the robotic fingers trace trajectories closer to human fingers, the retargeting objective requires less compromise between fingertip position, joint limits, self-collision, and final grasp shape. The reported improvement over OHand points in that direction, although it should eventually be validated on full manipulation trajectories rather than isolated hand motions. The preload findings also have consequences for simulation. Tendon tension should not be treated as a fixed CAD parameter. It is a latent hardware state that can change grasp success by several-fold. A robust training pipeline should either measure it online or randomize slack, stiffness, friction, and dead-zone parameters during training. For whole-body humanoids, the 600-gram mass is encouraging, but integration will require more than a mechanical wrist adapter. Researchers will need to characterize electrical power, peak current during multi-finger closure, thermal accumulation inside the palm, wrist torque during impacts, communication latency, and fault behavior when one tendon or motor fails. There is also a useful modularity argument. A digit built as a replaceable actuator-sensor module is easier to service than a monolithic hand. For humanoid fleets collecting manipulation data continuously, repair time and calibration repeatability may become as important as peak benchmark performance. The Experiments That Should Come Next. A convincing next-generation REL Hand study would add six evaluations:. Empirical decoupling maps. Measure unintended motion and tension changes across every pair of active axes, both before and after prolonged use. Dynamic system identification. Report backlash, dead zone, rise time, bandwidth, force control accuracy, and hysteresis under different tendon pretensions. Durability testing. Run repeated grasp cycles and track force, zero-position drift, cable condition, and adjustment frequency. Randomized object benchmarks. Vary object pose, friction, mass, and approach error, and include external perturbations and recovery attempts. True in-hand manipulation. Test rolling, controlled sliding, reorientation, finger gaiting, and regrasping without relying entirely on arm motion. Richer proprioception and touch. Add joint encoders, gyroscopes, tendon-tension sensing, and distributed tactile arrays capable of detecting shear and incipient slip. These experiments would clarify whether the mechanism is primarily an excellent anthropomorphic grasping hand or a foundation for genuinely dynamic dexterous manipulation. Final Assessment. REL Hand is valuable because it treats hand design as a coupled mechanical, sensing, calibration, and control problem. Its tendon-and-linkage architecture is not simply a compromise halfway between two transmission families. It uses each transmission for a motion where it is appropriate and then employs geometric routing to keep their effects separate. The resulting hand has a credible human-like workspace, 15 independently commanded coordinates, a manageable 600-gram mass, basic force and posture sensing, and enough fingertip force to grasp a broad range of everyday objects. It forms all 16 representative Cutkosky grasps, follows human motion more closely than the evaluated commercial baseline, and shows that correct tendon pretension can transform object-level success. It is not yet evidence that the hard problem of robotic hand dexterity has been solved. The distal joints remain underactuated, the tactile state is sparse, posture sensing is primarily quasistatic, pretension adjustment is manual, and the evaluation centers on grasp formation rather than autonomous in-hand skill. But for humanoid robotics, that does not diminish the paper’s core message. Better policies do not eliminate the need for better mechanisms. By reducing physical coupling and exposing maintenance variables such as tendon tension, REL Hand makes the hardware dynamics more legible to both classical controllers and learned policies. And that may be the most transferable contribution: use morphology and transmission geometry to remove complexity before asking the learning system to cope with what remains.