PaperlayerView exact revision on arXiv

Source-linked AI summary

A fully articulated humanoid dexterous hand with human-level joint structure enabled by linkage-based actuation

Jihao Li, Jianguo Wang, Haotian Guo, Yanzhe Wang, Guodong Lu, Huixu Dong

arXiv:2610.36789v1cs.RO

TL;DR

Existing humanoid hands face trade-offs among anthropomorphic dexterity, precision, and load capacity. Torch-Hand combines a 21-DOF hand with palm-integrated screw-linkage actuation and mechanically decoupled joints; the paper reports precise fingertip positioning alongside coordinated grasping and manipulation capabilities.

  • Problem

    Existing compact anthropomorphic hand architectures have not reconciled dexterity, precision, and payload capacity across their different actuation and transmission strategies.

    Source: Main Text: INTRODUCTION · p. 2

  • Method

    Torch-Hand uses a compact 21-DOF anthropomorphic hand with palm-integrated, linkage-screw actuation and independently addressable, mechanically decoupled joints.

    Source: Main Text: INTRODUCTION · p. 3

  • Results

    Below 0.03 mm repeatable fingertip positioning accuracy is reported, alongside a 33-grasp repertoire and task-matrix demonstrations of coordinated manipulation.

    Source: Main Text: INTRODUCTION · p. 3, DISCUSSION · p. 19

  • Takeaways & Limitations

    Torch-Hand provides a compact hardware route for precise, coordinatable, high-load manipulation within humanoid robotic systems.

    Source: DISCUSSION · p. 20

Abstract

from arXiv · show

Humanoid dexterous hands are a core component of embodied intelligence, enabling robots to translate high-level intent into rich and adaptive physical interaction. However, existing designs remain constrained by a persistent trade-off among anthropomorphic kinematic complexity, mechanical robustness, load capacity and control simplicity, limiting dexterous manipulation in unstructured environments. Here we present Torch-Hand, a fully articulated humanoid dexterous hand that replicates the joint structure and motion hierarchy of the human hand with 21 degrees of freedom (DOF), comprising a highly articulated five-DOF thumb and four DOF in each of the remaining fingers. The differentiated thumb integrates multi-axis CMC motion to provide broad opposition and flexible contact orientation relative to the four fingers. Unlike tendon-driven or hybrid actuation approaches, Torch-Hand adopts an entirely linkage-based transmission architecture. By integrating screw-linkage actuation, the hand achieves high grasping force and substantial payload capacity while maintaining a compact anthropomorphic form. Its actuation architecture is optimized for control tractability, with all actuators integrated within the palm and arranged in a serial-robot-like configuration that establishes a direct correspondence between motor motion and joint kinematics. This architecture mechanically decouples functional DOF to reduce unintended inter-joint motion propagation, while allowing independently addressable DOF to be actively coordinated into coupled multi-joint motion patterns during complex manipulation. Comprehensive experiments assess reachable workspace, coordinated fingertip trajectories, thumb opposition, 33-grasp taxonomy coverage, grasping force, payload capacity and task-level manipulation. These results establish Torch-Hand as a robust hardware foundation for dexterous embodied manipulation.

Main Text: INTRODUCTION

Torch-Hand is presented as a compact linkage-driven hand intended to combine anthropomorphic dexterity, motion precision, and load capacity. Its 21-DOF design and validation framework target capabilities that existing actuation strategies have struggled to combine.

Main Text: INTRODUCTION

  • Existing compact anthropomorphic hands have not reconciled dexterity, precision, and payload capacity, reflecting trade-offs across motor-gear, pneumatic, tendon-driven, and hybrid architectures.The cited constraints include distal bulk and thermal limits, compliance and low stiffness, transmission friction and elasticity, and added interfaces.Source: Main Text: INTRODUCTION · p. 2
  • Torch-Hand uses a palm-integrated 21-DOF architecture with a five-DOF opposable thumb and four four-DOF fingers.The design follows the human hand’s functional motion hierarchy and keeps actuators within the palm.Source: Main Text: INTRODUCTION · p. 3
  • Its independently addressable, mechanically decoupled DOFs support precise fingertip control while remaining available for coordinated multi-joint motion.The architecture is reported to achieve repeatable fingertip positioning accuracy below 0.03 mm.Source: Main Text: INTRODUCTION · p. 3
  • Rigid linkage pathways and screw transmission are presented as the means to combine force output, repeatable transmission, and structural robustness in a compact platform.The authors describe screw-driven sliders as providing mechanical advantage, positional stability, and self-locking capability.Source: Main Text: INTRODUCTION · p. 3
  • The evaluation spans fingertip workspace, thumb opposition, the 33-grasp taxonomy, trajectory and repositioning tests, payload experiments, and task-level manipulation.The task matrix combines fine fingertip operation, object handling, and forceful tool use.Source: Main Text: INTRODUCTION · pp. 3–4

RESULTS

  • Torch-Hand combines human-hand-like mobility, repeatable fingertip precision, and high-force load-bearing in one fully linkage-driven system.The authors describe this as overcoming the dexterity–precision–payload trilemma within an anthropomorphic platform.Source: RESULTS · p. 5

THEORY MODEL OF THE DEXTEROUS FINGER

Torch-Hand assigns functional joint motions to local linkage pathways while differentiating the four-finger and thumb modules. The resulting serial-chain model maps joint motion to fingertip pose, preserving independent control while allowing coordinated movement.

THEORY MODEL OF THE DEXTEROUS FINGER

  • Each of the four fingers has four DOFs, with individual slider-crank units providing independently regulated motion and transparent input-output mapping.The DOFs are MCP abduction/adduction, MCP flexion/extension, PIP flexion/extension, and DIP flexion/extension.Source: THEORY MODEL OF THE DEXTEROUS FINGER · p. 5, Figure 2 · p. 6
  • A shared-link four-bar mechanism coordinates PIP and DIP flexion during unloaded motion, while a compressible link permits independent DIP movement under contact or loading.The design makes the coordination condition-dependent rather than a fixed mechanical constraint.Source: THEORY MODEL OF THE DEXTEROUS FINGER · p. 7
  • The five-DOF thumb has three orthogonal CMC axes that expand its workspace and support opposition, with MCP-IP coordination designed for contact-dependent decoupling.Its motions are CMC rotation, abduction/adduction, flexion/extension, MCP flexion/extension, and IP flexion/extension.Source: THEORY MODEL OF THE DEXTEROUS FINGER · p. 7
  • The differentiated architecture preserves natural joint coordination during free motion and enables adaptive decoupling during object interaction.The authors connect functional differentiation of the thumb and four fingers with opposition workspace, grasp stability, and dexterous manipulation.Source: THEORY MODEL OF THE DEXTEROUS FINGER · p. 7

KINEMATIC ANALYSIS OF THE DECOUPLED MOTION FINGER

  • The serial-chain model maps anatomical joint variables to fingertip pose while independently addressable DOFs can move simultaneously for coordinated, human-like motion.The model represents four-finger and thumb chains, with the thumb’s orthogonal CMC axes forming a serial base for opposition.Source: Figure 3 · p. 8, KINEMATIC ANALYSIS OF THE DECOUPLED… · p. 9

MECHANICAL STRUCTURE OF TORCH-HAND

Torch-Hand implements its anthropomorphic hand through palm-centered actuation and independent linkage pathways. Screw-driven sliders and differentiated finger and thumb mechanisms support precise joint regulation, coordinated motion, and load-bearing.

MECHANICAL STRUCTURE OF TORCH-HAND

  • All drive units are concentrated in the palm, keeping the digits anthropomorphic and lightweight while accommodating independently actuated DOFs within the hand-scale form factor.The palm layout also simplifies wiring and leaves internal digit space for linkage transmission and potential sensor integration.Source: MECHANICAL STRUCTURE OF TORCH-HAND · p. 10
  • Screw-driven transmissions convert actuator rotation into guided slider motion, which drives joint rotation through linkage units.Rail-slider guidance and bearing-supported revolute joints support displacement accuracy and reduce parasitic friction; screw self-locking helps sustain loads within structural limits.Source: MECHANICAL STRUCTURE OF TORCH-HAND · p. 10
  • Spatial linkages and aligned joint axes let finger DOFs operate independently or in coordinated combination without transmission interference between PIP and MCP motions.The MCP flexion/extension motion uses a spatial four-bar mechanism, while PIP power passes through a universal joint.Source: MECHANICAL STRUCTURE OF TORCH-HAND · p. 12
  • The thumb’s orthogonal CMC axes provide independently controlled rotation, abduction/adduction, and flexion/extension for opposition and contact orientation.Local MCP and IP linkages preserve independently addressable joint motion across an expanded thumb workspace.Source: Figure 4 · p. 11, MECHANICAL STRUCTURE OF TORCH-HAND · p. 12

KINEMATIC AND MECHANICAL PERFORMANCE OF TORCH-HAND

Torch-Hand combines broad kinematic capability with repeatable positioning and high-load grasping. Evaluations cover joint workspaces, coordinated fingertip motion, thumb opposition, grasp diversity, force, and payload capacity.

KINEMATIC AND MECHANICAL PERFORMANCE OF TORCH-HAND

TASK-LEVEL MANIPULATION DEMONSTRATIONS OF TORCH-HAND

Torch-Hand is evaluated across tasks ranging from fine fingertip manipulation to forceful tool use, with results highlighting coordinated control across those demands. The discussion frames the platform as a broadly applicable hardware basis while identifying durability and reliability as areas for future evaluation.

TASK-LEVEL MANIPULATION DEMONSTRATIONS OF TORCH-HAND

  • Torch-Hand threads a 0.35 mm nylon line through a 0.5 mm needle eye, regulates pipette droplets, twists lipstick, and extracts a nail with a claw hammer.The demonstrations span tight geometric constraints, precise positioning and force control, opposition, and high-force tool interaction.Source: TASK-LEVEL MANIPULATION DEMONSTRATIONS… · p. 17
  • Across the 3×3 task matrix, mechanically decoupled joint control combines with coordinated multi-joint action, supporting transitions from delicate fingertip tasks to high-load tool use.The matrix spans increasing manipulation demand and interaction load; the demonstrated transitions retain thumb-finger opposition, motion coordination, and stable forceful grasping.Source: TASK-LEVEL MANIPULATION DEMONSTRATIONS… · pp. 17–18, Figure 8 · p. 18

DISCUSSION

  • The task demonstrations span household, service, laboratory, and industrial contexts without changing the hand’s mechanical configuration.The tasks require distinct combinations of precision, opposition, coordination, grasp stability, and force output.Source: DISCUSSION · pp. 19–20
  • Torch-Hand is presented as an extensible hardware substrate for future sensing and feedback integration, with long-term durability, fatigue resistance, and reliability under sustained high-load interaction still requiring systematic evaluation.The discussion identifies visual perception, tactile sensing, and force feedback as potential additions for more adaptive operation.Source: DISCUSSION · p. 20

MECHANICAL DESIGN AND FABRICATION

Torch-Hand combines modular palm-integrated construction with screw-driven linkage actuation and a serial kinematic model. The design and assembly procedures support independent module inspection and traceable actuator-to-fingertip motion.

MECHANICAL DESIGN AND FABRICATION

  • The hand uses replaceable linkage modules, a titanium-alloy palm frame and cover, aluminum-alloy phalanges and links, rubber fingertip pads, and protective ABS shells.Finger and thumb modules connect to the palm through modular mechanical interfaces.Source: MECHANICAL DESIGN AND FABRICATION · p. 21
  • Palm-integrated motors drive screw-and-slider transmissions whose outputs connect to linkages that convert linear motion into joint rotation.Ball screws serve finger MCP flexion/extension, while trapezoidal screws serve the remaining joints; guide rails and slider blocks support the transmissions.Source: MECHANICAL DESIGN AND FABRICATION · p. 21

MODULAR ASSEMBY PROCEDURE

  • Finger, thumb, and palm modules are assembled separately, inspected, and then integrated, with joint axes aligned to minimize transmission interference among finger DOFs.Before testing, each DOF is actuated independently and in selected combinations to check direction, range, and smooth motion.Source: MODULAR ASSEMBY PROCEDURE · pp. 21–22

ACTUATION, LINKAGE TRASMISSION AND SERIAL KINEMATIC MODELING

  • Each motor’s rotation maps through screw-driven slider displacement and a linkage-specific geometric closure to joint angle, after which serial-chain kinematics calculate fingertip pose relative to the palm.The finger chains are modeled as open serial chains, with fingertip position extracted from the final homogeneous transformation matrix.Source: ACTUATION, LINKAGE TRASMISSION AND… · p. 22

EXPERIMENT SETUP FOR KINEMATIC PERFORMANCE EVALUATION

Kinematic performance is assessed through sampled fingertip workspaces, commanded trajectory tracking, thumb opposition, and execution of a 33-grasp taxonomy. The protocols define sampling, command, and success criteria for each evaluation.

EXPERIMENT SETUP FOR KINEMATIC PERFORMANCE EVALUATION

  • Reachable fingertip workspaces are estimated by sampling 50,000 configurations per finger model and 100,000 for the thumb within measured joint ranges.Sampled fingertip positions form a point cloud whose boundary points define the workspace envelope.Source: EXPERIMENT SETUP FOR KINEMATIC… · p. 22
  • Trajectory tracking uses six planar paths commanded at 20 Hz and two spatial paths at 30 Hz, with success requiring uninterrupted motion inside a 2 mm path width.The planar paths test three flexion/extension joints; spatial paths also coordinate MCP abduction/adduction.Source: EXPERIMENT SETUP FOR KINEMATIC… · p. 23
  • The Kapandji test executes calibrated thumb-and-finger postures, counting a target successful when the thumb tip contacts its prescribed region without assistance or hand repositioning.Small joint-angle offsets may compensate for assembly tolerance and contact compliance.Source: EXPERIMENT SETUP FOR KINEMATIC… · p. 23
  • For each of 33 grasp types, human glove-recorded joint angles are mapped to Torch-Hand, and success requires holding the scaled object stably for 10 s without support or visible slip.Small command offsets are used only when needed for object contact, assembly tolerance, and compliance.Source: EXPERIMENT SETUP FOR KINEMATIC… · p. 23

EXPERIMENT SETUP FOR MECHANICAL PERFORMANCE EVALUATION

Mechanical performance protocols measure repeat positioning, fingertip force, whole-hand grip force, destructive grasping, and payload lifting. They specify instrument setup, actuation conditions, and success criteria.

EXPERIMENT SETUP FOR MECHANICAL PERFORMANCE EVALUATION

  • Repeat positioning is measured against a 3.000 mm reference over 50 trials per joint, with mean error, standard deviation, and maximum bounded error calculated for each finger.A dial indicator records the stabilized reading after the finger moves from neutral to terminal posture.Source: EXPERIMENT SETUP FOR MECHANICAL… · p. 23
  • Fingertip force is recorded after stabilization while the actuator runs in current-loop mode at its preset current limit, with each joint-angle condition tested for 10 cycles.Single-joint output is tested, as is coordinated MCP-PIP flexion for the four fingers.Source: EXPERIMENT SETUP FOR MECHANICAL… · pp. 23–24
  • Whole-hand force is measured while all five digits compress a sensor-equipped grip ball, and payload success requires lifting a 40 kg dumbbell without visible slip, structural failure, or loss of grasp posture.A palm-mounted wedge stopper prevents the ball from slipping; aluminum-can crushing is assessed qualitatively.Source: EXPERIMENT SETUP FOR MECHANICAL… · p. 24

TASK-LEVEL MANIPULATION DEMONSTRATIONS

The demonstrations retargeted human hand motions to Torch-Hand and replayed the resulting joint sequence alongside a pre-recorded arm pose. Success required completing the intended task without assistance after initial placement while maintaining grasp or contact without visible slip, posture loss, or interruption.

TASK-LEVEL MANIPULATION DEMONSTRATIONS

  • Human hand joint sequences were recorded at 20 Hz, converted to fingertip trajectories, and retargeted to Torch-Hand joint angles through inverse kinematics.Arm end-effector poses were obtained by teleoperation and replayed during demonstrations.Source: TASK-LEVEL MANIPULATION DEMONSTRATIONS · p. 24
  • A task counted as successful only if Torch-Hand completed its intended outcome without assistance after initial placement and maintained contact without visible slip, posture loss, or interruption.The arm first moved the hand to the recorded task pose; the hand sequence was then sent to the controller at 20 Hz.Source: TASK-LEVEL MANIPULATION DEMONSTRATIONS · p. 24

Supplementary Materials

The supplementary materials comprise methods, figures, tables, and movies.

Supplementary Materials

Supplementary Materials

The supplementary materials comprise methods, figures, tables, and movies.

Supplementary Materials

Supplementary Methods

The supplementary methods derive local slider-to-joint mappings and use serial-chain kinematics to represent fingertip motion. They also describe thumb opposition axes and the retractable-link mechanism for DIP motion.

Supplementary Methods

  • Slider position determines finger-joint bending through crank-slider geometry, with the resulting joint variables expressed in DH coordinates.The actuator input is converted to slider displacement, then to joint rotation through the slider-crank linkage's geometric closure.Source: Supplementary Methods · p. 30
  • The atan2 function determines a polar angle over (−π, π] using both input signs, resolving quadrant ambiguity and remaining defined when x=0.Source: Supplementary Methods · p. 30
  • With the PIP at its target bend, the retractable link still permits DIP bending, and its actual length together with the PIP angle determines the DIP angle.The passage distinguishes the retractable link's original length from its length after compression.Source: Supplementary Methods · p. 31
  • Finger chains are modeled with DH transformations from a palm-fixed frame to the fingertip, while the thumb is a five-DOF serial chain.The thumb's first three transformations correspond to orthogonal CMC axes that support opposition.Source: Supplementary Methods · p. 32
  • Supplementary figures show the slider-to-joint geometric derivation and the assembly workflow for the finger modules, thumb module, and palm integration.Source: Figure S1 · p. 34, Figure S2 · p. 35
  • Supplementary tables cover Torch-Hand comparisons, kinematic-pair definitions, finger and thumb DH parameters, benchmark-object specifications, and repeat-positioning accuracy.Source: Table S1 · p. 36, Table S2 · p. 37, Table S3 · p. 37
Loading 2610.36789v1…