Benchmarking Dexterity of Multifingered Robot Hands: A Review and Perspective

1. INTRODUCTION

• The HAND ERC links advanced dexterity to fuller independence for people with motor impairments, better manufacturing jobs, more resilient supply chains, and wider access to robotics.Benchmarks should measure progress and, ideally, provide actionable feedback to system designers.Source: §1 · p. 2
• Dexterity is the capacity to use contact efficiently and reliably to change or learn about the physical world; the review presents the HAND ERC’s evolving perspective because no universal definition exists.Manual dexterity involves hands, objects, and environmental constraints working together.Source: §1 · p. 2

1.1. The Relevance of In-Hand Manipulation to Dexterity

• In-hand manipulation moves grasped objects relative to the palm, using finger contact motions such as rolling, sliding, or twisting; multifingered hands offer versatility beyond fixed-grasp manipulation.The review focuses on multifingered hands and sensitive control of finger contact forces.Source: §1.1 · p. 2

1.2. Multifingered Hands

• A multifingered hand has articulated fingers branching from a palm, may use intrinsic or remote actuation, and can include wrist mobility, tactile sensing, and compliant surfaces; the review is not limited to anthropomorphic hands.Source: §1.2 · p. 3

1.3. Dexterity Benchmarking Levels

• System-level benchmarks test short-horizon atomic tasks such as jar-capping, but require integrated hands, arms, vision, control, and other system components.The review notes that most previous dexterity benchmarking has been at the system or long-horizon level.Source: §1.3 · p. 4
• Because full-system sensitivities are difficult to establish, the review proposes tracing design effects through component- and hand-level benchmarks, while noting that continuous sensitivity analysis misses qualitative strategy shifts.The proposed chain relates design-to-component, component-to-hand, and hand-to-system performance sensitivities.Source: §1.3 · p. 4

2.1. Taxonomy-Inspired Benchmarks

• Taxonomy-inspired benchmarks range from human movement classes to contact-based frameworks; examples include 50 tasks combining grasping and in-hand manipulation, POMDAR’s six grasps and twelve atomic tasks, and normalized rotation and translation measures.The cited frameworks include Elliott and Connolly’s movement classes, Bullock et al.’s contact relationships, Coulson et al.’s benchmark, and Liconti et al.’s POMDAR.Source: §2.1 · p. 5

2.2. Task-Centric Benchmarks

• Human activity assessments such as Jebsen–Taylor and SHAP emphasize functional task completion, while atomic suites represent dexterity with tasks such as relocating a sphere, manipulating a pen, opening a door, and hammering.The atomic suite selects one representative task for each of four broad manipulation classes.Source: §2.2 · pp. 5, 7
• Domain-specific suites target tasks such as assembly, in-hand manipulation, and deformable-object interaction, using metrics including completion time, reliability, final pose error, and deformation outcomes.Examples include NIST assembly taskboards, YCB-object manipulation, and deformable manipulation benchmarks.Source: §2.2 · p. 6
• Large-scale simulation benchmarks cover varied manipulanda, skills, and interaction horizons; Great March 100 uses dataset-derived primitives, language-model expansion, and expert review to produce 100 representative tasks.Some benchmarks are embodiment-specific, while ManiSkill3 and RoboMIND seek to generalize across embodiments and domains.Source: §2.2 · pp. 6–7
• Simulation may miss physical sensing noise, contact dynamics, hardware variability, and environmental uncertainty; ManipulationNet addresses this with reproducible hardware kits, a unified client, and five real-world tasks.The five tasks span Physical Skills and Embodied Reasoning tracks.Source: §2.2 · p. 7
• DexBench organizes 56 cases across 18 categories using six object-state-complexity axes, spanning atomic and long-horizon tasks.Source: §2.2 · p. 7

2.3. Perspective

• The HAND ERC selected a compact suite of fewer than 20 atomic tasks for its DexNex testbed, spanning varied characteristics and difficulty; tasks are expected to benefit from in-hand manipulation over fixed-grasp strategies.Task selection considered direct versus indirect manipulation, object and environment properties, control demands, constraints, and task dynamics; base tasks include success criteria, metrics, rationale, and difficulty variations.Source: §2.3 · pp. 7–8

3.1. Common Benchmarks for the Hand, Wrist, and Fingers

• Robot-hand datasheets commonly report size, workspace, repeatability, speed, and strength, while durability lacks standard measures and has received little attention.ORCA, RUKA, and LEAP use continuous-operation tests for repeatability and robustness; grasp endurance measures how long a hand holds a hanging weight.Source: §3 · pp. 8–9, §3.1 · p. 10

3.2. Hand Benchmarks

• Hand benchmarks cover grasp taxonomies, grasp capacity and resilience, anthropomorphic motion, thumb mobility and interaction, and strength, speed, and grasp dynamics.Examples include AHAP’s scored daily-living grasps, the Kapandji test, KaRMA’s pinch-object translation and rotation scores, and grasp cycle time.Source: §3.2 · pp. 10–11

3.3. Finger Benchmarks

• Finger benchmarks assess effort, motion, and dynamics through measures including fingertip strength, force tracking, speed, repeatability, and trajectory deviation.Finger response to external excitation has received little attention; maximum speed may be reported as fingertip speed, joint speed, or full-range frequency.Source: §3.3 · pp. 11–13

3.4. Wrist Benchmarks

• Wrist benchmarks score mobility, stability, speed and torque, load capacity, flexibility, and mechanical responsiveness.Stability is quantified by deviation from the desired pose under external loading; load capacity concerns structural failure under bending loads.Source: §3.4 · p. 13

3.5. Perspective

• Current hand-level benchmarks assess force production or motion individually, while dynamic performance is rarely measured.Existing datasheet metrics largely follow industrial-arm measures designed for precise free-space positioning rather than sensitive interaction.Source: §3.5.1 · p. 14

3.5.1. Gaps in benchmarking.

• Impedance control suits low desired impedance, while admittance control suits high desired impedance; robot fingers’ contact-force regulation makes impedance control promising.The review notes that feedback control cannot arbitrarily reshape plant dynamics because stability and passivity impose limits.Source: §3.5.2 · pp. 14–15

3.5.2. Addressing the gaps.

• Mechanical transparency requires faithful bidirectional force transmission, but transmission compliance, inertia, friction, backlash, cogging, and torque ripple distort force and motion.These transmission effects complicate sensitive interaction and motivate evaluating transmission behavior rather than only force or motion capabilities.Source: §3.5.1 · p. 14, §3.5.2 · p. 15
• The review proposes forward- and backward-direction measures of breakaway force, force bandwidth, and force accuracy to characterize mechanical transparency.Force bandwidth reflects response to changing force commands or environmental displacement; fingertip impedance measurements provide a local linear approximation and do not capture configuration variation or transmission nonlinearities. Transparency is presented as a promising route to high-quality impedance control and reduced impact forces and internal wear.Source: §3.5.2 · pp. 15, 17

4.1. Actuation Benchmarks

• Actuator benchmarks span normalized force or torque, stroke, maximum output, and dynamic response, while mass-accounting differences complicate comparisons and static measures miss dynamics.Volume-normalized metrics also matter under humanoid-hand size constraints.Source: §4.1 · p. 17
• Step response, bandwidth, force-velocity and speed-torque behavior, and force smoothness, linearity, and hysteresis capture actuator dynamics and output regulation.Dynamometers assess bandwidth under varying loads; stroke limits, soft-actuator hysteresis, and gearbox backlash can impair force regulation.Source: §4.1 · pp. 17–18

4.2. Mechanical Transmission Benchmarks

• Mechanical advantage, friction, backlash, hysteresis, degree-of-freedom coupling, and hardware complexity shape transmission transparency; compliance and reflected inertia also affect control bandwidth and impedance.Mechanical advantage may vary with configuration, while high tendon compliance can reduce position-control bandwidth.Source: §4.2 · p. 18
• Tendon benchmarks assess material strength, lifetime, stiffness, creep, strength degradation, routing friction and hysteresis, and termination strength.Life-cycle tests use expected routing elements, and terminations are often the limiting failure point; tendons are a primary source of failure in many robot hands.Source: §4.2 · pp. 18–19

4.3. Soft, Durable Skin Benchmarks

• Skin softness can enlarge contact patches at low forces, stabilize grasps, and simplify roll-slide control, but softness and durability tend to be inversely correlated.Skin thickness, modulus, and surface texture influence how contact forces reach an underlying tactile layer.Source: §4.3 · p. 20
• Common skin metrics include Young’s modulus, tensile strength, elongation at break, and Shore hardness, while viscoelastic adhesion can produce stick-slip sliding.Reported skin materials include rubber, PDMS, platinum-cure silicones, TPU, and foams.Source: §4.3 · p. 20
• Skin friction is commonly measured by reciprocating sliding against a defined counter surface or estimated with contact-mechanics models, and can be tuned through stiffness, coatings, or surface patterns.Texture effects on friction depend on pattern and lubrication; some studies report stiffness or contact area without friction.Source: §4.3 · p. 20
• Wear mechanisms for skin elastomers are not well studied, and durability is reported only qualitatively.Source: §4.3 · p. 20

4.4. Sensing Benchmarks

• Sensor benchmarks use operating range, resolution, accuracy, and sensing frequency to characterize measurement span, smallest detectable change, ground-truth deviation, and acquisition rate.Source: §4.4 · p. 20
• Kinesthetic sensing benchmarks are relatively limited, and sensor fit within hand geometry motivates volume-specific resolution to capture performance relative to consumed space.Joint positions can be ground-truthed with external cameras, while joint-loading measurements can be checked against known loads or external measurements.Source: §4.4 · pp. 20–21
• Tactile benchmarks cover sensor-specific measures such as spatial resolution and robustness, alongside sensor-agnostic spatial acuity evaluated through accuracy, stability, and generalizability.The psychophysics-inspired framework also guides hardware optimization.Source: §4.4 · p. 21
• Gripper stiffness and actuator angular velocity significantly affect tactile-sensor performance, indicating that element-level metrics may not capture behavior under integration conditions.Source: §4.4 · p. 21

Transmission

• Transmission evaluation includes hardware complexity and the effects of compliance on position-control bandwidth.Source: Transmission · p. 22
• Hysteresis, backlash, and inertia are listed as transmission properties relevant to benchmarking.Source: Transmission · p. 22
• Transmission friction can be characterized through efficiency.Source: Transmission · p. 22

• The passage lists tendon diameter under tendons, with bibliography markers (57, 60).Source: Transmission · p. 22
• The passage lists routing friction with references 109, 110, and 125, without specifying its role or classifying it as a transmission characteristic.Source: Transmission · p. 22
• The passage mentions termination strength and gives markers (121, 122); it does not describe it as a transmission characteristic or state any other classification.Source: Transmission · p. 22
• Material strength is also listed (60, 120–122).Source: Transmission · p. 22

• The cited passage identifies spatial resolution and associates it with references 144–147, 149, 151–153, 156–161, and 163–166.Source: Sensors · p. 22
• The cited passage names stretchability and gives references 148 and 150; it does not establish from this excerpt that stretchability is classified as a sensor characteristic.Source: Sensors · p. 22
• Calibration error is also listed (66, 164).Source: Sensors · p. 22
• Table 4 groups existing component-level metrics by subsystem.Source: Table 4 · p. 23

4.5. Perspective

• Component metrics developed in isolation do not capture all requirements, constraints, and tradeoffs that emerge when components are integrated into a dexterous hand.Source: §4.5.1 · p. 23

4.5.1. Gaps in benchmarking.

• Transmission friction and efficiency lack standardized metrics for their combined contribution to hand transparency, and tendon routing comparisons are complicated by configuration dependence and diverse routing elements.Source: §4.5.1 · p. 23
• Tactile sensing lacks standardized representations and metrics for comparing technologies and their suitability for whole-hand coverage.Source: §4.5.1 · p. 23
• Soft-skin benchmarks lack standardized measures of how contact area evolves during interaction, despite skin’s roles in stabilizing compliance and conformal contact.Source: §4.5.1 · p. 23

4.5.2. Addressing the gaps.

• The HAND ERC is developing standardized component-level benchmarks and equipment, including testbeds for tendon properties, tactile sensing, and contact-area evolution.The testbeds address tendon transmission efficiency and strength/lifetime, tactile accuracy, resolution and sensitivity, and contact-area changes under different loading conditions.Source: §4.5.2 · pp. 23–24, Figure 6 · p. 24
• Actuator output inertia, transmission friction, and transmission stiffness are identified as performance measures that influence hand mechanical transparency.Higher actuator inertia reduces responsiveness; transmission friction is approximated by force transmission efficiency and breakaway force, while low transmission stiffness reduces apparent finger impedance but also force-control bandwidth.Source: §4.5.2 · pp. 24–25

• Component- and hand-level benchmarks are proposed to help bridge the attribution gap between low-level design decisions and system- and application-level performance.The HAND ERC is developing test equipment for these benchmarks.Source: §5 · p. 25
• Establishing how design decisions affect system performance remains unresolved, so the paper envisions feedback loops combining many simulated hand designs with fewer physical-hand experiments.These loops would build a database of component-, hand-, and system-level benchmark performance and sensitivities.Source: §5 · p. 25, Figure 7 · p. 26
• Accurate simulation remains a major challenge for contact-rich interactions, soft skins, tendon transmissions, tactile sensors, and simultaneous hard and soft contacts.Domain randomization is sometimes used, but the paper calls for efficient simulators that better model hands in real environments.Source: §5 · p. 25
• Rapid-prototyping materials can introduce friction, compliance, hysteresis, creep, and poor strength or durability, negatively affecting mechanical transparency.The paper calls for manufacturing techniques that accelerate the physical design-build-test cycle of high-transparency hands.Source: §5 · p. 25
• The field is moving faster than ever, partly due to unprecedented private investment, and benchmarking is expected to help measure and accelerate progress.Source: §5 · pp. 25–26

• The authors report no affiliations or memberships they believe might affect the review’s objectivity.Source: DISCLOSURE STATEMENT · p. 26
• They report no funding they believe might affect the review’s objectivity.Source: DISCLOSURE STATEMENT · p. 26
• They report no financial holdings they believe might affect the review’s objectivity.Source: DISCLOSURE STATEMENT · p. 26