Spatially distributed biomimetic compliance enables robust anthropomorphic robotic manipulation

Abstract

The impressive capabilities of humans to robustly perform manipulation stems from compliant interactions, enabled by the structure and materials distributed in the hands. We propose that mimicking this spatially distributed compliance in an anthropomorphic robotic hand enhances open-loop manipulation robustness and leads to human-like behaviors. Here we introduce the ADAPT Hand, equipped with configurable compliant elements on the skin, fingers, and wrist. After quantifying the effect of compliance on individual components against a rigid configuration, we experimentally analyze the performance of the full hand. Through automated pick-and-place tests, we show the grasping robustness mirrors the estimated geometric theoretical limit, while stress-testing the robot to perform 800+ grasps. Finally, 24 items with varying geometries are grasped in a constrained environment with a 93% success rate. We demonstrate that the hand-object self-organization behavior, driven by passive adaptation, underpins this robustness. The hand exhibits different grasp types based on object geometries, with a 68% similarity to natural human grasps.

Introduction

The human ability to robustly pick, place, and manipulate objects in uncertain environments is common place, and we perform it with ease¹. Whilst our planning and sensory-motor capabilities play a key role, these compliant and robust interactions are also heavily influenced by our hand morphology, kinematics, and dynamics^2,3. From softness in the skin creating stable contact, to compliant muscle synergies⁴ forming diverse ranges of grasp types⁵, the physical structure can enable robust interactions through emergent self-organized behaviors in uncertain environments⁶ without the need of an active controller. This serves as inspiration for advancing the baseline open-loop capabilities of anthropomorphic robotic manipulators. Such physical intelligence will enable robotic hands with fluidic, human like dynamic behaviors and is essential for fully leveraging the potential of rapidly advancing control and learning strategies^7,8.

Compliance based approaches, whether embodied⁹ or physically intelligent¹⁰, improve robotic hand versatility. Fully compliant hands like RBO Hand I/II/III^11,12,13, BCL-26¹⁴, and Shorthose Hand¹⁵–typically pneumatically actuated–show robustness to object geometry. For instance, RBO Hand-III manipulates a Rubik’s cube using purely open-loop actions¹⁶. However, fully soft hands lack force application, repeatability, and agility. Tendon-actuated soft-rigid hands mimic human musculoskeletal structures¹⁷, utilizing rolling contact joints and compliant ligaments^18,19,20,21 for impact resistance and low friction^22,23,24,25.

Embedding compliance in actuation also enhances robustness. Underactuated tendon-routing enables multi-finger grippers to grasp varied objects with a single actuator^26,27,28. Compliance in actuation extends to theoretical frameworks, from active joint torque control²⁹ to adaptive synergies that enable diverse behavior with minimal actuation^30,31.

The requirements for compliance extends beyond the hand, with the wrist motion contributing to coordinating the behavior of passive, i.e. non-actuated hands³². This combination of wrist motion and bio-inspired passivity in the structure has been demonstrated to enable tasks such as dynamic and varied piano playing³³ and grasping²¹.

Compliance enhances open-loop robustness across various approaches. However, systematic assessment methods remain lacking, both for individual elements and overall hand performance concerning compliance³⁴. Localized hand elements (e.g., skin, joints, fingers) are often evaluated using force-displacement graphs to measure stiffness variability^17,35 and finger workspace for kinematics^14,36. Yet, no assessment metric spans different manipulation scenarios, nor is there a unified standard. While human studies analyze directional stiffness³⁷ or tapping force^38,39, they lack the range of interaction-based tasks needed for robotic manipulation. For robotic hands, existing benchmarks are often high-level task metrics influenced by control strategies^40,41,42, object-focused with sparse fail/success criteria⁴³, or limited to static capabilities like grasp taxonomies⁵, range of motion⁴⁴, and hardware features⁴⁵. Moreover, assessments rarely examine both distinct spatial regions and their combined effect on the full hand.

We propose that a spatial distribution of compliance, matching human magnitudes, is essential for robust, fluid interactions in anthropomorphic hands. By integrating distributed compliance in the skin, fingers, and wrist, human-like behavior can emerge with minimal open-loop control. Stiffness matching to a human has shown benefits for individual components such as skin^46,47,48 and arm⁴⁹. Compliance distribution has also been explored in locomotion^50,51, enabling stable, self-organized gait patterns⁵². Applying this to robotic hands allows exploration of hand-environment self-organization, facilitating unconscious selection of robust grasp configurations (Fig. 1B). In addition, if this compliance is variable, its effect across different manipulation tasks can be studied.

Fig. 1: An anthropomorphic robot hand designed with biomimetic spatial distribution of stiffness leads to a emergence of robust and self-organizing behavior.

A The ADAPT Hand is compliant in its skin, finger, and wrist, which operate in different lengths scales, 0.1 cm, 1 cm, and 10 cm respectively. B Through a small number of identical open-loop waypoints, the hand can grasp objects of largely varying geometry successfully. The distributed stiffness allows for the hand to be robust upon unknown environmental interactions, and a self-organization to take place between the hand and objects, resulting in different grasps. C A diagram to illustrate how the different stiffness elements affect environmental interaction in different ways which complement each other.

Achieving this requires a robotic hand with spatially distributed compliant elements that has stiffness that can be varied and multiple actuated degrees of freedom for diverse tasks. We introduce the ADAPT (Adaptive Dexterous Anthropomorphic Programmable sTiffness) Hand (Fig. 1A), featuring a soft skin covering the contact surface, series elastic actuated fingers/thumb, and an impedance-controlled wrist, all configurable to match their biological counterpart. The compliant elements in the ADAPT Hand interact at different length scales: skin (${{\mathcal{O}}}\left(0.1{{\rm{cm}}}\right)$), fingers (${{\mathcal{O}}}\left(1{{\rm{cm}}}\right)$), and wrist (${{\mathcal{O}}}\left(10{{\rm{cm}}}\right)$) (Fig. 1A). While the displacement of each element is not strictly linear (e.g.: the finger has a torsional stiffness at the joints), the length scales describe how they affect the physical interactions, and allow for easier comparison when analyzing the full hand. The effect of compliant elements combine to produce interaction with manipulated objects that is robust to uncertainty, compared to the non-human-matched rigid configuration, shown in Fig. 1B. Although qualitative, Fig. 1C presents a rationale to how the deformation of each compliant element during a grasping motion. The wrist motion conditions how the hand interacts with the environment globally, and its stiffness allows for the full hand to safely adapt to the 10 cm scale disturbance. The compliance in the base joints allow the finger motion to adapt to the 1cm scale. Finally, at the interface of the interaction, the skin deformation creates a larger contact area increasing the stability.

Although robotic hands are abundant, hands which implement compliant elements are limited. Table S1 compares state of the art robotic hands, focusing on designs which deliberately include spatially distributed compliance. Among these hands, the ADAPT Hand is unique as it is the only design which a) include compliant elements in the skin, finger, and wrist, and b) those elements are configurable in its stiffness. Within other hands, notably the DLR hand⁵³ also features stiffness elements that are configurable through its actuation in the fingers and wrist; however lacks the inclusion of the skin, and realized by complex actuation and control. On the other side of the spectrum, the Gilday hand²¹ is highly anthropomorphic with compliant skin and fingers; however lacks a wrist and cannot be actuated for dexterous motion.

In this work, through an experimental study, we first show the benefits of compliance on the skin and fingers is quantified through three or more interaction tasks involving single and multi-finger tasks. Extending to the full hand, and using a robotic setup to autonomously perform pick-and-place tasks, we quantify the open-loop robustness for grasping approaches with respect to an estimated theoretical limit based on object and hand geometry. Using the same setup, we also evaluate the robustness of the ADAPT Hand design through an extensive autonomous experiment with over 500 grasps across 9+ hours of uninterrupted pick-and-place operation. The hand showed a success rate of 97%. Finally, by introducing a compliant wrist we show how the distributed compliance leads to self-organization of emergent grasp types, which matches that of a human (direct similarity of 68%), with a success rate of 93% to handle 24 items spanning a few to 100s of millimeters. We attribute this robust performance of the full hand as a result of the contributions of the compliant elements at the component level.

Using a custom robotic hand with spatially distributed compliance that can be configured and an extensive experimental evaluation, spanning both component level interactions and full hand tasks, we demonstrate the human matched stiffness in an anthropomorphic hand results in robustness and human-like behaviors in manipulation tasks.

Results

ADAPT hand

Robotic platform

The ADAPT Hand is a bio-inspired compliant anthropomorphic robotic hand platform, with joint kinematics that reflect an human adult’s hand (detailed in Section “ADAPT Hand Hardware”) and the stiffness reflecting a relaxed human adult’s hand. The robotic system is designed to have independently configurable stiffness in the skin, finger, and wrist.

The hand itself, shown in Fig. 2A, is custom-built and features 12 actuators controlling 20 joints via tendon-driven actuation. A key design feature is the finger (Fig. 2B), where two actuators independently control the flexion-extension motion. Two antagonistic tendons (blue and green cables) actuate the metacarpophalangeal (MCP) joint, while a single tendon (red cable) flexes the proximal interphalangeal (PIP) and distal interphalangeal (DIP) joints; where these motions are independent. The abduction-adduction motion is coupled (controlled by one actuator) and is series elastic. The thumb follows a similar design but includes two antagonistic actuators at the carpometacarpal (CMC) joint. Despite the underactuation, the hand achieves all 33 grasp taxonomies⁵ and all 10 positions of the Kapandji score⁴⁴ (see Supplementary Fig. S1). The full ADAPT hand system consists of the robotic hand mounted on a Franka Research 3 robotic arm, which controls the global position and orientation for manipulation tasks. Here, the wrist refers to the dynamics at the base of the hand, driven by the robotic arm’s movement.

Fig. 2: The ADAPT Hand with its salient features of the finger design and the human matched stiffness profiles.

A The ADAPT hand. B Finger design with two independent actuation and a series spring on the metacarpophalangeal(MCP) flexor. C The force-displacement measurement curve for the skin, finger, and wrist for the robot and four human subjects. For the skin and finger, the rigid configuration of the ADAPT Hand is also shown for comparison.

Stiffness configuration and human matching

The ADAPT hand is designed with configurable stiffness across its components. The skin has a modular design, allowing different materials to be used while maintaining the same geometry (Fig. 7D). Finger compliance is achieved via a series spring on the MCP flexion tendon (Fig. 2B left), while wrist stiffness is adjusted through the robotic arm’s impedance controller. More advanced stiffness variation methods exist: FEA-modeled geometry³⁵ and jamming-based mechanisms⁵⁴ for the skin, variable stiffness mechanisms⁵⁵ and joint-level stiffness tuning⁵⁶ for the fingers, and cable-driven designs for dynamically adjusting wrist stiffness with low inertia⁵⁷. While the ADAPT hand’s stiffness configurability is simplified for integration, this trade-off enables all three elements to function within a single system. While the skin and finger stiffness are are only manually adjustable offline, it is sufficient for this work as only two settings (compliant/human-matched and rigid) are used. The wrist stiffness is tunable online but remains constant throughout experiments.

All experiments use two stiffness settings: a human-matched compliant mode and a rigid mode. Fig. 2C shows force-displacement curves for each compliance element, comparing the robot to four human subjects (detailed in Section “Measuring human stiffness”). The ADAPT hand’s stiffness is configured so its force-displacement curve falls within human trial bounds. Skin stiffness is adjusted by selecting materials to cover the rigid structure. The chosen material (EcoFlex20) closely matches human skin, whereas a fully rigid covering is 10–40 times stiffer. Finger stiffness is tuned by swapping the MCP joint’s series spring. A rigid finger (without the series spring) is 30 times stiffer than the human-matched version. The plateauing force profile of the tuned finger, similar to the human hand, enables near-constant fingertip force under large displacements. Wrist stiffness is controlled via the robot arm’s impedance settings, which are tuned to match human wrist force profiles in the vertical direction (the other two directions are set to be identical).

Motion programming

The actuation signals are manually programmed open-loop signals. That is, an operator will record key waypoints for both the finger/thumb joints and the wrist 6dof pose (detailed in Section “ADAPT Hand motion programming”). During execution, the recorded waypoints are linearly interpolated and directly replayed. For every task in the work, the motion is programmed to best mimic how a human would perform them.

Skin: contact stability

The skin is crucial in robot-environment interactions, serving as the direct interface. In humans, cutaneous stiffness enhances local deformation, improving contact stability and shear force generation. To assess the impact of skin stiffness, we compare a human-matched soft skin to a rigid PLA counterpart, coated with EcoFlex20 to maintain consistent surface friction and isolate compliance effects.

The primary effect of skin stiffness is increased shear force for a given motion. Shear force was analyzed during two sliding motions (forward and backward) on a plate (Supplementary Fig. S2) at varying finger-surface displacements. The top right of Fig. 3A shows shear force measurements via a load cell at the midpoint of the motion. Across different motions and surface conditions, the soft skin consistently generates higher shear force than the rigid case. Fig. 3A (bottom) presents representative shear force profiles for both motions, showing similar temporal patterns but nearly double the force magnitude with soft skin.

Fig. 3: Experimental results of the effect of a rigid and a soft skin for three different manipulation tasks.

A Results from the finger sliding experiment. Schematic (top-left), box plot and raw values for the shear force of rigid and soft skins measured at the midpoint of the interaction (top-right), example time series of the shear force of rigid and soft skins for the two sliding motions (bottom). B Results from the knob turning experiment. Schematic (top) and turn angles for the soft and rigid skins as the diameter d and resisting torque τ is varied (bottom). C Results from the finger gaiting experiment. Schematic (top) and completed gaits for soft and rigid skins as the held block width w is varied (bottom).

Skin stiffness on task performance

The soft skin’s ability to generate higher shear force enhances contact stability. To measure its impact, we test a cylindrical knob-turning task using the middle finger and thumb with predefined finger motion. Performance is measured by turn angle θ, while the environment varies by adjusting knob diameter d and resistance torque τ (Fig. 3B top). Fig. 3B (bottom) shows that across all conditions, the soft skin outperforms the rigid one. Notably, as τ increases, performance drops more for the rigid skin (37deg) than for the soft skin (18deg), demonstrating greater robustness.

The benefits of soft skin extend to tasks requiring contact stability, such as finger gaiting when holding an object. Performance is measured by counting completed gaits before failure while varying block width w. Results (Fig. 3C) show that soft skin improves stability and performance. While the rigid skin benefits from increased w due to higher holding forces, the soft skin consistently outperforms it.

In both experiments, the soft skin creates a larger contact area, enhancing stability through higher shear forces. This results in consistently better task performance and, in some cases, increased robustness to environmental uncertainty.

Finger: pose adaptation

The finger’s flexion-extension motion originates from the series elastically actuated MCP joint. In this section we show how the series elastic MCP joint can enable centimeter scale pose adaptation of the finger, resulting in robust behavior.

When external forces are applied, the entire finger passively conforms to the environment due to the MCP joint’s position at the finger’s base. Combined with PIP/DIP actuation, this allows the fingertip to maintain consistent contact along a surface. Fig. 4A illustrates two sliding motions (slide front and slide back), achieved by flexing or extending the PIP/DIP joints through three waypoint transitions (detailed in Supplementary Fig. S2). The compliant MCP mechanism also enables pseudo force control via overdriving the joint. If contact occurs at q°, setting the position demand to $q+{\Delta }_{{{\rm{MCP}}}}^{\circ }$ maintains stable contact at a similar force magnitude, even on unknown surface shapes. Fig. 4A (right) shows normal and shear force profiles for slide back motion under Δ_MCP = 0, 7.5, 15mm for two trials. Force magnitude increases almost linearly with Δ_MCP, and repeated actions show high consistency (average RMSE: 36 ± 43mN across all experiments).

Fig. 4: Compilation of the experimental results regarding the finger focusing on the usage of a soft metacarpophalangeal(MCP) joint.

A Key frames from two finger sliding motions(left). By overdriving the metacarpophalangeal(MCP) joint, pseudo force control is possible with good repeatability shown by the force profile over time (right). B Trajectories of relative changes in the three joint angles for the sliding motions executed by a human, soft robot finger, and a rigid robot finger. C Schematic for the finger sliding experiment (top). Maximum forces recorded as the sliding plate is displaced in the z and θ directions for the two motions/soft and rigid fingers (bottom). D Schematic for the knob turning experiment (top). Tendon waypoints required for the motion(mid) and turn angles as the environment is varied (bottom) for the soft and rigid fingers. E Schematic for the finger gaiting experiment (top). Completed gaits (mid) and average holding forces (bottom) for soft and rigid fingers as the held block width w is varied. F Success rates for the three cubes rotated in-hand (top). Pictorial sequence of the in-hand cube rotation sequence (bottom).

Human motion comparison

Through the compliant MCP joint, the ADAPT Hand finger motion shows a kinematic resemblance to that a human finger at the joint displacement level. For the slide back and front motions, a similar motion was performed by a human whilst visually recording its pose.

Figure 4B shows the joint evolution of the finger during the two sliding motions for a human, the robot with a soft MCP, and the robot with a rigid MCP. In both motions, the soft finger is most similar to the human motion. For the slide back motion, the difference is most notable in the relative motion of the MCP and PIP joints. The two joint angles are near consistent on the rigid finger, implying that all the motion is through the DIP joint (the image above the plot shows this in effect). For the slide front motion, the soft finger shows even higher similarity with the human while the rigid finger barely moves. The slide front motion inherently relies on the MCP joint to flex as the PIP/DIP joints are extending, which the rigid finger cannot achieve without extra waypoints.

To quantify these differences, RMSE error against human motion is computed (Supplementary Fig. S3). The robot(soft) has an error of 9.2 ± 5. 5° compared to 14.5 ± 7. 8° for the robot(rigid).

While this comparison only indicates similarity between human and robot motion, it highlights the necessity of a compliant MCP joint for natural PIP/DIP motion while maintaining contact.

Measuring robustness through finger stiffness

In Fig. 4A, B we demonstrate human-like sliding motions with minimal planning (only three waypoints) by leveraging the compliant MCP joint. The robustness (behavior invariance to changes in the environment) is measured by comparing task performances between the soft and rigid fingers while varying the environment.

First, we evaluate the consistency of forces generated during sliding motions. While displacing the finger and surface in the z and θ directions (Fig. 4C top), we measure force variability for both finger types.

The scatter plots in Fig. 4C compare the normal and shear forces (F_V, F_H) for the two sling motions between the soft and rigid fingers for combinations of Δ_z and Δ_θ. The MCP overdrive (Δ_MCP) on the soft finger is chosen to match the maximum force applied to the rigid finger when Δ_z, Δ_θ = 0 as shown by the opaque scatter point.

The error bars show the standard deviation of maximum force recorded, where the soft finger is on average 2.4 times lower. The lower variability in interaction forces implies the soft finger is less influenced (i.e.: more robust) by position variability.

The knob-turning task further evaluates robustness during finger actuation. Turn angle is measured while varying knob size and robot displacement (Fig. 4D top). For the soft finger, the motion is achieved using just three waypoints, reusing the slide back and front motions for the thumb and middle finger, respectively. Executing the same waypoints on the rigid finger causes damage due to excessive forces. Instead, the rigid finger requires precise motion planning with 15 waypoints to follow the knob’s contour. These waypoints show in in Fig. 4D illustrates how compliance simplifies control.

The performance of this task is summarized in the scatter plot (Fig. 4D bottom) for three settings: default (when the motion was programmed), reduced knob diameter, and a 1 cm knob displacement. The soft finger consistently outperforms the rigid one. Since the rigid finger relies on precise waypoints, force variations lead to inconsistent contact and lower turn angles. For example, with a smaller knob, the soft finger’s turn angle increases by 10°, while the rigid finger’s drops by 8°. When the knob is displaced, the rigid finger is damaged due to high reaction forces.

In the finger gaiting task, the number of completed gaits until failure and the average holding force F_Hold is recorded while the width of the block w is varied (Fig. 4E top). Similar to knob-turning, a separate trajectory for the rigid finger was programmed which carefully tracks the contact with the block. Unlike the knob turning, the number of waypoints remain identical, but a high precision is still required for the MCP joint angle to follow the width of the block. The raw measurements for the completed gaits and F_Hold for the three w settings are shown in Fig. 4E as two performance metrics. For both metrics, soft finger has a higher consistency as w changes in comparison to the rigid finger. This is a direct extension of the result in Fig. 4C, as the interaction forces are consistent under uncertainty in displacement, resulting in a more robust behavior.

In-hand cube re-orientation

Throughout Fig. 4A–E, while quantifiable benefits for robust interaction were demonstrated through the compliant MCP joint, the actions themselves are simple and its scalability to more complex manipulation actions are unclear. To show the simple waypoint planning methodology can be extended to complex motion, a cube re-orientation task (see Supplementary Video S1) is performed, where cubes of varying dimensions are rotated continuously in an open-loop action comprised of multiple waypoints.

Figure 4F captures key frames of the robot using its fingers, thumb, and palm to continuously re-orient a cube with a total of 12 waypoints (number of waypoints marked at every frame). While the entire motion is complex, it can be formed by combining the simpler motions shown in Fig. 4C–E such as the slide front used to passively re-orient the cube. The same programmed motion is repeated 20 times for cubes with three different sizes, with 100% success rate of the small and medium cubes and 90% for the large cube, highlighting the robustness to cube dimension uncertainty. This motion is unique to the soft finger, as the rigid configuration would damage the robot when applying force on the cube.

Systematic pick-and-place robustness assessment of the ADAPT Hand

By introducing compliance in the robot’s skin and fingers, we demonstrated improved robustness and stability in both single and multi-finger interactions. Now, extending to all five digits, we explore its impact on pick-and-place, a fundamental manipulation task. We assess hardware robustness and open-loop repeatability over hundreds of picks, quantifying performance. The robot’s picking success is evaluated as an object is incrementally displaced along the x and y axes, compared against a theoretical geometric limit (Supplementary Fig. S4).

To systematically evaluate robustness at scale, an automated robotic system is developed for continuous pick-and-place tasks. It includes a secondary arm with a movable plate and an overhead camera (Fig. 5A). The process begins with Robot 1 introducing a manual offset Δ_M by first identifying the object’s displacement Δ_O and adjusting the plate by Δ_M − Δ_O to position the object accordingly. Robot 2, equipped with the ADAPT Hand, then executes an open-loop pick-and-place motion. This cycle, outlined in Fig. 5A, runs for hours with minimal intervention.

Fig. 5: Pick and place experiment results showcasing the setup, object wise evaluation, and a timeseries demonstrating extensive testing.

A Robotic setup to conduct large quantities of automatic pick-and-place experiments while controlling the displacement of the object. B Measured limits on object displacement (orange axis), estimated geometric limits based on object size and hand closure motion. C Success and failed grasps throughout the two experiments: robustness assessment and uninterrupted pick-and-place totaling 845 grasps. Each row corresponds to a different object, while the red lines indicate failed grasp attempts.

Measuring open-loop robustness

Using this setup, we measure the ADAPT Hand’s robustness to displacement against the estimated theoretical upper limit. Five objects with distinct geometries were selected, each programmed with a human-like grasp. Five additional unseen objects with similar geometries were then introduced (see Fig. 5B). For each object, Δ_M was incrementally varied in horizontal and vertical directions until two consecutive grasp failures occurred.

Figure 5B displays ten pre-grasp hand poses overlaid with measured and theoretical displacement limits. The blue shaded area represents the estimated geometric limit, where an object’s geometrical center can be positioned within the fingers’ range without fingertip collisions at the start of the grasp (see Supplementary Fig. S4). Measured limits in the vertical and horizontal directions are shown as orange error bars.

Surprisingly, the hand meets or exceeds theoretical geometric limits for open-loop pick-and-place, despite using approximate grasps and including unseen objects. At millimeter-scale displacements, skin stiffness maintains grasp stability. At centimeter-scale displacements, finger pose adaptation plays a larger role, allowing different fingers to hold the same object stably (Supplementary Fig. S5). Excluding the outlier (Empty coke can), the measured-to-theoretical limit ratio is 1.2 ± 0.2 (vertical) and 0.8 ± 0.2 (horizontal). Vertical robustness exceeds the theoretical limit as fingertip collisions still result in successful grasps. Horizontal robustness falls short since large deviations focus forces on one side, pushing the object out. For the coke can, measured limits far exceed theoretical predictions due to its round shape (allowing fingertip grasps) and low weight (enabling grasping at the ends).

Extended period of operation

Using the continuous testing capabilities of the robotic system (Fig. 5A), we evaluate the ADAPT Hand’s robustness and repeatability in an extended, uninterrupted trial. The system completed 500 grasps at Δ_M = 0 with a 97% success rate, running without any hardware or software intervention. Most failures (15/16) occurred when grasping the tape, which was already near the robustness limit (as seen in Fig. 5B).

Figure 5C presents a time series of all experiments. The first phase (left of the blue dotted line) evaluates robustness (Fig. 5B), while the second phase represents uninterrupted operation. Across both experiments, the ADAPT Hand completed 845 grasps over 16 h without hardware/software modifications. The only visible wear was minor dirt accumulation on the silicone fingertips (Supplementary Fig. S6), which had no impact on performance during or after testing. Supplementary Video S2 provides a sped-up version of the trial, with raw footage details in Supplementary Note 1.

Self-organizing grasps

The previous section focused on skin and finger stiffness for local interaction stability. A compliant wrist extends these interactions, enabling tasks like grasping while sliding an object off a table (Fig. 6A) - a natural human action but challenging for robots requiring making and breaking contact. Using open-loop hand and wrist trajectories mimicking human table grasping, we assess robustness across different objects and observe emergent grasp types.

Fig. 6: Experimental results from a contact rich grasping task off a tabletop surface, showcasing the motion, successes, and categorization of grasp types against the object geometry and between the robot and a human.

A Key frames overlayed with wrist motion (blue arrow), finger motion (red arrow), and new contact regions (pink shade) for the human and robot grasping an object off the table. The number of waypoints for the wrist or hand illustrated underneath the frames. B Objects used for this experiment with failed trials indicated by red circles above the image. C Five grasps identified between the human and robot. D Object geometry plotted with the color indicating the grasp type observed. E Direct comparison between robot and human grasps for every object.

When humans perform this task, three invariant behaviors emerge (Fig. 6A, top row, showing key frames of grasping a lemon): Firstly, the approaching motion is always the hand moving downwards until making contact, either with the table or the object. Secondly, after the initial contact is made the fingers continue to hold its contact during the motion, either with the table or the object. Thirdly, the grasping motion shown by the “Flex fingers”, “Wrist up”, and “Retract wrist” frames is a single continuous motion where the fingers and thumb curls to form a grasp as the wrist moves away from the table.

By incorporating a wrist with human-matched stiffness alongside the compliant skin and fingers, the robot reliably replicates human grasping behaviors (Fig. 6A, bottom two rows). As the wrist presses into the table, compliance enables safe, controlled contact with the fingers or palm (first two frames). The grasping motion (last three frames) mirrors human behavior, with the fingers and thumb flexing as the wrist lifts while maintaining continuous contact. This interaction is shaped by distributed compliance, particularly at the wrist, which conditions the fingers’ response to the environment. During the grasping sequence, only the wrist is actively actuated (Fig. 6A, bottom). The fingers move only before the final wrist motion to form a grasp, constrained by the table. As the wrist lifts, it gradually removes this constraint, allowing the fingers to self-organize and secure the object while maintaining contact through compliance.

Using the same open-loop motion, 24 objects of varying geometry were tested, from thin items like a pencil to bulkier ones like an apple (Fig. 6B). Each object was grasped three times with varied placements along the robot’s motion path, achieving a 93% success rate (70/75 trials). Failures occurred only with the phone, bolt, and lemon. Supplementary Video S3 presents the full grasping experiment from two angles.

Emergence of discrete grasp types

The robustness of the robot to grasp a wide variety of objects through identical commands can be explained through self-organization. Consider the two robot motions given in Fig. 6A. In the top row, the lemon is a large object where the robot makes multiple contacts throughout the motion (see pink highlights). On the bottom row, the robot only makes contact with a pencil just when the grasp happens. The resultant grasp of the two objects are also different: the lemon is held with a power grasp while the pencil is held only with the fingertips, which is appropriate if a human were to grasp the two objects.

Grasp motions were analyzed for both the robot and human and categorized into five distinct grasp types (Fig. 6C), classified by hand posture and contact points, aligning with grasp taxonomies (see Section “Grasp type categorization”). Supplementary Video S3 illustrates the four grasp types observed in the robot.

The scatter plot in Fig. 6C clusters objects by geometry (length, width, height) and corresponding robot grasp types. The robot self-organizes into discrete grasp types based on object dimensions–fingertip grasps for small/flat objects (blue), extends to use its thumb as object height increases (green), and full power grasps for bulkier items (yellow, red).

Like in Fig. 6A, robot grasp types closely match human behavior. Humans naturally adjust their grasp based on object geometry⁵⁸, which was also observed when subjects grasped the same objects.

The matrix in Fig. 6C directly compares robot and human grasps. Humans exhibit one additional grasp type–the finger surface grasp, where fingers remain straight, which the robot cannot replicate due to programmed PIP/DIP curling. Despite this, 68% (17/24) of grasps match exactly, with the rest in adjacent categories. Assuming humans choose the most stable grasp through experience, the robot’s biomimetic kinematics and distributed stiffness enable it to self-organize and select similar grasps.

Discussion

Although the proposed concept realized by the ADAPT Hand show remarkable performance, there are clear limitations to be addressed in the future. At the heart of this work is the distributed compliance, which can negatively affect certain tasks. For instance the soft skin and finger limits the ability to exert high forces in one direction. This means strongly pinching an object or button pressing tasks are challenging. When using the full hand for manipulation, the passive adaptation is simply a result of force equilibrium working in a desired way. This means, when the external forces are too large (e.g.: an object that is too heavy such as the phone in Fig. 6B), the execution fails. The same is true for balancing forces between multiple fingers and a held object. Since the fingers will passively adapt, tasks such as controlled object re-orientation is difficult.

Sensor feedback and control are natural next steps to address these limitations. This includes sensing joint torques, tendon forces, and tactile feedback across the skin surface. However, applying sensor feedback to this robot is not straightforward, as its manipulation capabilities emerge from interactions between its compliance and the environment. Although actuation signals use position commands, each waypoint is programmed with a high-level intention (e.g., “apply downward force” or “move roughly in this direction”) rather than targeting specific positions or forces. Sensory-motor control should therefore not enforce explicit trajectories but instead complement emergent behavior. One approach could combine high-level intention through open-loop waypoints with a controller to regulate contact forces, similar to a shared-control scheme used in ref. ⁵⁹ used to stabilize a noisy EMG signal. Alternatively, bio-inspired sensory-motor control, such as central pattern generators used in salamander locomotion⁶⁰, could foster emergent, efficient behaviors. While central pattern generators may not directly apply to manipulation, a sensory-motor coordination framework that drives emergent behaviors⁹ is essential for sustaining and enhancing robust interactions.

The ADAPT Hand as a hardware platform itself also has room for improvement. One direction is to increase the level of biomimicry of the human hand. The current hand although is bio-inspired, clear differences are present especially in the joint placement and the skin/flesh distribution. For example, there is no material covering the MCP joints. Consequentially, multiple grasp formations do not translate well from a human to the robot. Another limitation is the lack of a wrist just below the hand. While the 7dof robot arm can orient the hand in any angle, in practice the arm must move a large displacement to resemble small wrist motions seen in a human. Likewise to the hand itself, increasing the biomimicry in the arm kinematics can enable more complex human motions to be realized by the robot. Not only could this simplify manual programming or teleoperation by a human, but one could imagine mapping motions learnt from the abundance of videos of humans to be applied to this biomimetic robotic hand-arm structure.

At a more high level, methodologies to evaluate and understand complexity in manipulation must be discussed. The interactions with a robotic hand (especially one where deformable interactions take place) is so complex, which makes it difficult to draw generalizable conclusions and meaningful theories which can be translated to other scenarios, while being so context dependent. This work takes an experimental approach, showing that robust behaviors observed at the component level (skin, finger) directly relate to robustness at the full-hand level. In this work, we use a large quantity of experimental data and variety of tests that span different manipulation scenarios and interactions. However, there can always be more metrics, preventing hard conclusions to be drawn under this experimental metric driven approach. Theoretical contributions face similar challenges. Simple models fail to capture manipulation complexity, while complex models approach physics based simulations, offering little generalization (like real world experiments). Despite its limitations, experimental benchmarking better reflects reality than theoretical models. Quantifying robustness is another challenge. This work automates long-duration experiments to go beyond single-use robotic demonstrations. However, operational hours remain insufficient to assess hardware maturity fully. Speed is another limitation. Humans perform the same number of grasps over ten times faster⁶¹–but this too is context-dependent. Overall, more structured methods are needed to link low-level functionalities to high-level behaviors without requiring massive experimental validation or excessive task simplification.

Conclusion

In this work, we present an anthropomorphic robot hand, the ADAPT Hand, designed with a biomimetic distribution of compliance across different lengths scales in the skin, finger, and wrist. Starting from low level interactions of the skin and finger levels, we show the presence of a compliant skin and MCP joint on the finger leads to higher performance and robust to environmental uncertainty across three of tasks. Through the ADAPT Hand’s configurable stiffness, a direct comparison between a compliant and rigid robotic hardware is made. Expanding the task to include all five digits, the robustness within the hand is measured to lie close to an estimated theoretical limit for a pick-and-place task. Using the same measurement setup, the robustness and repeatability of the hardware platform itself is shown through a damage-less execution of over 800 grasps and 15 h of operation. Finally, a compliant wrist motion is introduced to grasp 24 objects through an interaction-rich motion across a tabletop surface. This motion which fully combines the distribution of compliance across the robot body, where a wrist motion which conditions the passive pose adaptation of the fingers while contact stability from the skin enhances the grasp. The resultant discrete grasp types are self-organized based on the object geometry, similar to that of a human. Not only, a comparison between human and robot grasp shows 68% of grasps are directly matching. Overall, we demonstrate an physical intelligence approach towards anthropomorphic robot design which considers the interaction and motion at all the lengths scales of robotic manipulation. The robustness given simplest form of actuation control input culminating in the self-organizing grasps.

Methods

ADAPT hand hardware

The ADAPT Hand is a custom designed and fabricated anthropomorphic robot hand, with its four fingers and thumb having dimensions similar to an adult human. The entire hand is fabricated from commercially available 3D printed Polylactic acid(PLA) and Thermoplastic Polyurethane(TPU) with shore hardness 98A/65D, and uses tendons (cables) for its actuation. Figure 7A shows the ADAPT Hand without the skin, showing the joint kinematics and actuation scheme. A total of 12 servo motors (Dynamixel XM430-W210-R) located beneath the hand actuates the 20 joints (four joints per finger/thumb). Having less actuators than joints means certain joints are underactuated (shown by subscripts a, b, c in Fig. 7A) or passive (“P” in Fig. 7A). Details on the CAD model is presented in section Supplementary Note 1.

Fig. 7: Detailed description of the ADAPT hand hardware alongside the motion programming scheme.

A Actuated, underactuated, and passive joints for the ADAPT Hand. B Finger design and tendon routing. C Series elastic actuation design for the the MCP joint. D Soft and rigid skin designs. E Abduction/adduction linkage mechanism. F Waypoint recording and replay method used to program the robot motion.

Finger design

The finger design shown in Fig. 7B is a key feature of the ADAPT Hand, where the same design is used across the five digits. One actuator controls the pin-jointed MCP joint with antagonistic tendons (the thumb has an extra antagonistic joint, but rotate 90 degrees to make the 2DoF CMC joint). The PIP/DIP flexure joints which are coupled by a single flexor tendon is actuated by another motor. The pin joint in the MCP joint allows the routing of the PIP/DIP flexor to pass through the center of rotation, fully decoupling the two axis of actuation. The extension forces for the PIP/DIP come from the combined effect of the TPU flexure joint and the elastic thread, placed on the backside of the finger (shown by a black line in Fig. 7B).

The compliance at the finger level is generated by a series elastic MCP joint which is achieved by routing the MCP flexor tendon around a pulley connected to an extension spring (shown in Fig. 7C). Replacing or removing the series spring can change the finger stiffness (as in Fig. 2B).

MCP abduction-adduction motion

A notable feature of the mechanical design of the ADAPT Hand is the series-elastic linkage mechanism driving the abduction-adduction axis of the MCP joints for the index, ring, and little fingers. The mechanism and its open and closed states are shown in Fig. 7E. The linkage mechanism connects to the MCP pin joint with a series elastic TPU material, making the MCP joint compliant in 2 axes. Having an actuated spread axis of the MCP joints increases the workspace (thus the capability of the hand) such as holding large objects.

Dexterity of the ADAPT Hand

Combining all the mechanisms of the fingers, the ADAPT Hand can be actuated to produce dexterous motions. Starting from the zero position (all fingers straight), the hand can be actuated to produce all 33 grasp taxonomies⁵. Likewise, the hand is able to complete all 10 postures on the Kapandji test. The results for both tests are shown in Supplementary Fig. S1.

Skin design

A modular skin fully covers one side of the ADAPT Hand. Fig. 7D shows the index and middle finger equipped with a rigid and soft skin respectively. The skins are identical in their geometry and is approximately 3 mm offset from the “bones”. The soft skin is fabricated from cast EcoFlex20 and the rigid skin is 3D printed PLA. A thin (≈0.5 mm) layer of EcoFlex20 is glued on the surface of the rigid skin to maintain the surface friction property.

ADAPT Hand motion programming

In all experiments, the ADAPT Hand (including the robotic arm) operates through manually programmed open-loop motions. For both the hand and arm, a series of manually determined key waypoints are recorded to then be played back (see Fig. 7F). The source code details are presented in section Supplementary Note 1, with the software system integration and hardware interfacing which allows the programming and replay of waypoints on the hand and arm described in Supplementary Fig. S7.

Programming the hand

The ADAPT Hand is controlled by directly commanding the tendon displacements for each actuator (which is proportional to the motor angle). To simplify the procedure to manually record waypoints, the hand (fingers) is operated using a custom built “signal mixer box” with 12 linear potentiometers which map to the position of each actuator. As illustrated in the top left of Fig. 7F, two waypoints can be simply defined by varying the linear potentiometer positions. Being an interactive device, the mixer box allows for quickly programming motions while having full control over the tenon positions.

Once a set of one or more waypoints (slider position) are recorded, the robot can smoothly move between the waypoints as in the top right of Fig. 7F.

Programming the arm

The Franka Research 3 robot arm is controlled using a gravity compensated impedance control introduced in ref. ⁶², where the end effector 6 dof pose and corresponding stiffness can be commanded. To program the arm, the end effector stiffness is set to zero which allows the arm to be manually moved around. Likewise to the hand, after recording few key poses (see bottom left of Fig. 7F), the 6 dof waypoints are interpolated and replayed.

For motions which involve both the hand and arm waypoints are replayed sequentially, meaning the hand and arm are not actively actuated at the same time.

Experimental setup and procedure

Measuring human stiffness

The force displacement characteristics of the human and ADAPT hand presnted in Fig. 2 is measured by recording the reaction force on a loadcell and its Cartesian position as it is moved by a 6-axis robot arm (UR5). For measuring the wrist reaction force on the ADAPT hand, the force estimation of the Franka Research 3 is used while the hand was displaced manually.

The measurement setup and movement directions for the skin, finger and wrist is shown in Supplementary Fig. S8. For all human measurements, four subjects (two male two female) were selected and were blindfolded and instructued to relax while the robot arm makes contact, to remove as much bias and effect of conscious reaction from the human.

The forces are recorded for three repeats to capture any variations.

Measuring low-level interactions

Three tasks: finger sliding, knob turning, and finger gaiting, were conducted to characterize the low level interactions (such as contact forces and kinematics) of the ADAPT Hand skin/finger with the environment. The experimental setup are shown in Fig. 8A–C for the three tasks respectively.

Fig. 8: Experimental setup for measuring hand-environment interactions at the skin and finger level.

Independent and dependent variables for every experiment are shown in blue and orange respectivley. A Finger sliding experiment. B Knob turning experiment. C Finger gaiting experiment.

In the finger sliding task, a single finger interacts with a wooden plate through two sliding motions generated by combining a flexing motion of the MCP joint and a flexing or extending motion of the PIP/DIP joints (see Supplementary Fig. S2). Figure 8A shows the experimental setup where the ADAPT Hand (rigid held by a UR5 arm) interacts with the wooden plate mounted above two load cells measuring the vertical (F_V) and horizontal (F_H) forces. The finger joint angle data used in Fig. 4B were extracted by recording the April tag markers throughout the motion. For the two motions, five independent variables were combinatorialy tested with two repeats which are: pose offset Δθ (±10deg), Δz (±10 mm), soft and rigid configurations for the skin and finger, and overdrive of the MCP tendon Δ_MCP (±7.5 mm) (only for the soft finger).

In the knob turning task, the middle finger and thumb was used to turn a knob shown in Fig. 8B. The knob turn angle θ is used to assess the performance measured by a position encoder (AMS AS5048B). The environment was varied in three ways: the x and y position offset (±10 mm each), the diameter of the knob d(45 ± 15 mm), and the reaction torque of the knob τ (low: 3.3 ± 0.6 Nmm, High: 18.8 ± 3.9 Nm). The reaction force is modulated by varying the vertical forces applied on the knob which rests on a plastic surface. When the finger is in the soft configuration, the motion is near-identical to the sliding motions introduced in the finger sliding experiment. In the rigid finger configuration, a secondary motion is programmed to replicate the same motion to ensure the robot doesn’t damage itself.

In the finger gaiting task, a plastic block is held between the thumb and four fingers shown in Fig. 8C. Starting from all four fingers contacting the block, a finger gaiting pattern is executed (shown by 1, 2, 3, 4) in repeat until eventually the block is dropped. Only the width w of the block is varied during this experiment (±5 mm), while the number of completed gaits and the holding force F_grip is measured by an inbuilt load cell.

Grasp type categorization

The grasp types shown in Fig. 6C for both the robot and human are categorized based on which part of the hand is used to hold/interact with the object and its posture. Each grasp is also related with the grasp taxonomies in ref. ⁵. The “Finger surface” grasp (only present for the human) is achieved by keeping the four fingers straight, and using that as a surface to push the object against by the thumb tip (corresponds with #22:Parallel extension taxonomy). The “Fingertip” grasp uses only the tips of the fingers and thumb to hold the object (corresponds with #6-8:Prismatic 2-4 finger taxonomy). The “Tip and thumb” grasp uses the fingertips and the middle and/or proximal phalanges (corresponds with #10:Power disk taxonomy). The “Power (small)” and “Power (large)” grasps are both power grasps where one or more phalanges of the fingers/thumb and the palm is used, distinguished based on the diameter of the grasp (corresponds to #2:Small diameter and #1:Large diameter taxonomies).

Although only 24 objects are used for the experiment, 25 grasps are recorded because the paper tape generated two distinct grasp types.