PaperlayerView exact revision on arXiv

Source-linked AI summary

Morphology and actuation as inductive biases in robotic hand manipulation

Zalán Tari, Eszter Birtalan, Péter Polcz, Miklós Koller

arXiv:2609.05206v1cs.ROeess.SY

TL;DR

Robotic hands differ in morphology and actuation, and how these structures shape coordination and control is the central question. The paper analyzes kinematics and actuation across four morphological aspects and finds that structural benefits are nonuniform across fingers and control levels.

  • Problem

    The paper asks how contrasting robotic-hand morphology and actuation structures shape kinematic conditioning and coordination, which affect control difficulty.

    Source: §2.3 · p. 7

  • Method

    A unified framework evaluates conditioning in the task Jacobian, actuation matrix, and their product across four morphological aspects, then tests predictions in reinforcement-learning experiments.

    Source: §3 · p. 15, §5 · p. 21

  • Results

    Structural benefits were nonuniform: axis geometry can improve thumb conditioning but not long-finger conditioning, while networked coupling favors effective mapping at long fingers.

    Source: §4 · p. 18, §5 · p. 21

  • Takeaways & Limitations

    Anatomical fidelity invests selectively across morphological aspects, while simplified designs trade expressiveness and redundancy for uniform actuator authority and a smaller policy space.

    Source: §5 · p. 21

    Only two hands were considered, so applying the framework to soft-bodied and underactuated grippers is needed to test its generalizability.

    Source: §5 · p. 21

Abstract

from arXiv · show

Robotic hands vary widely in anatomical fidelity and mechanical complexity, and these structural choices influence the coordination of joint motions and the difficulty of controlling the system. A unified framework is presented in which the kinematic and actuation stages are analysed separately and in composition, through the conditioning of the task Jacobian, the actuation matrix, and their product. It is applied to two hands representing opposing design philosophies, the Shadow Dexterous Hand and the Anatomically Correct, Biomechatronic Hand, along four morphological aspects: joint axis geometry, actuator-to-DOF ratio, coupling architecture, and authority distribution. All parameters are derived from the hands' canonical digital representations. Anatomical fidelity carries no uniform advantage: oblique axes improve thumb conditioning but leave the long fingers worse conditioned than the orthogonal-axis design, while the branching tendon network improves the effective control mapping at every long finger and worsens it significantly at the thumb, where actuator authority is concentrated on thumb opposition. Predictions derived from these metrics are evaluated against reinforcement learning experiments using PPO, DDPG+HER, and TQC+HER, across three different tasks.

Nomenclature

The nomenclature defines notation for hand kinematics, actuation, and conditioning analyses, alongside abbreviations for the hands, joints, and reinforcement-learning methods.

Kinematic notation

  • n, m, k, and r denote kinematic DOF, independent actuators, task-space dimensionality, and redundancy, while x, T(q), T0, and ξ describe task space and kinematic geometry.Redundancy is r = n − k; T(q) is the forward-kinematics transform in SE(3), T0 the home pose, and ξ a joint screw axis in R^6.Source: Nomenclature · p. 2

Actuation notation

  • A(q) maps actuator input u to joint velocity, B(q) = J(q)A(q) is the effective control mapping, and Rm(θ) and Lm describe the tendon network.Rm(θ) is the moment-arm Jacobian and Lm the vector of muscle-level tendon lengths.Source: Nomenclature · p. 2

Analysis metrics

  • σ denotes singular value, κ(·) the condition number σmax/σmin, ω(q) manipulability, and qref the mid-range reference configuration.Source: Nomenclature · p. 2

Abbreviations

  • SDH and ACBH abbreviate the Shadow Dexterous Hand and Anatomically Correct, Biomechatronic Hand; RL methods are PPO, DDPG, TQC, and HER.DOF means degrees of freedom; PoE, D-H, and URDF denote kinematic representation formats, while CMC/TCMC, MCP, PIP, DIP, and IP denote joint names.Source: Nomenclature · p. 2

1. Introduction

The introduction frames hand morphology as an inductive bias on manipulation control and learning, and proposes comparing structural metrics with reinforcement-learning outcomes across two contrasting hands.

Introduction

  • Morphology influences which manipulation behaviours can be controlled and how efficiently learning discovers them, while adaptability across tasks remains underexplored in performance-focused design studies.Structural design choices therefore act as inductive biases on robotic-hand control and learning.Source: §1 · p. 3
  • The study compares SDH’s orthogonal axes, partial underactuation, sparse coupling, and uniform authority with ACBH’s oblique axes, overactuation, branching tendons, and specialized authority.These hands represent contrasting design philosophies across all four aspects of the proposed taxonomy: joint axis geometry, actuator-to-DOF ratio, coupling architecture, and authority distribution.Source: §1 · p. 3
  • Canonical digital models supply each hand’s parameters; a common structural model yields task Jacobian J, actuation matrix A, and effective control mapping B=JA.The SDH parameters come from its official URDF; the ACBH parameters come from the MuJoCo model developed by Polcz et al., with some differences from the physical hand.Source: §1 · p. 3, Figure 2 · p. 4
  • The study asks which morphological aspects have measurable consequences and which design philosophy suits each, then checks framework predictions against RL training outcomes.The comparison covers two hands, two tasks with variations yielding three environments, and three RL algorithms; findings are specific to these hands, though the framework can generalize when parameters are extractable.Source: §1 · p. 3

2. Methods

The methods analyze hand geometry and actuation separately and in combination as control mappings, comparing structural metrics with reinforcement-learning training data. They pair task-matched algorithms with three manipulation environments spanning fingertip reaching to block rotation.

2.1. Mathematical background

  • The coordinate-invariant Product of Exponentials represents joints by screw axes and composes their motions for structural comparison.Unlike Denavit–Hartenberg parameters, it describes arbitrarily oriented axes without link-attached coordinate frames.Source: §2.1 · p. 5

2.2. Robotic hands as systems

  • The framework models kinematic structure and actuation separately to analyse hands’ physical coordination properties independently of control strategies.The formulation applies across hands; differences between the ACBH and SDH lie in their structures, not the task definition.Source: §2.2 · p. 6

2.3. Kinematic structure

  • The SDH has fewer DOFs per finger and largely orthogonal aligned axes, whereas the ACBH has more DOFs and anatomically derived oblique, non-parallel axes.The paper uses these differences to analyse kinematic conditioning and redundancy.Source: §2.3 · p. 7

2.4. Actuation structure

  • The SDH uses fixed actuator-to-joint mappings, including coupled PIP–DIP joints, while the ACBH’s branching tendon network yields configuration-dependent actuation mappings.For the SDH, the index, middle, and ring fingers have three actuators for four joints, while the thumb is fully actuated; the ACBH mapping is derived from tendon moment arms and a pseudoinverse.Source: §2.4 · pp. 7–8

2.6. Coordination under bounded actuator effort

  • Under bounded actuator effort, A and B map the same input set to joint- and task-space velocity ellipsoids whose axis ratios encode conditioning.Directions with smaller singular values require greater effort and can become mechanically disfavoured coordination patterns.Source: §2.6 · p. 8

2.7. Structural metrics

  • Structural metrics cover Jacobian conditioning, task-specific redundancy, and per-finger manipulability, with metrics evaluated at the midpoint of each joint’s range unless stated otherwise.Manipulability complements κ(J) and κ(B); it is evaluated per finger for multi-finger tasks.Source: §2.7.1 · p. 9, §2.7.2 · p. 11, §2.7.3 · p. 11

2.8. Effective control-mapping conditioning

  • The condition numbers of A, J, and B = JA isolate actuation, kinematic, and combined actuator-to-task conditioning, respectively.Comparing them identifies whether coordination bottlenecks arise from geometry, actuation, or their interaction.Source: §2.8 · pp. 11, 13

2.9. RL algorithms used for empirical evaluation

  • PPO is on-policy with clipped updates, whereas DDPG+HER is off-policy and relabels failed episodes with alternative goals for sparse-reward learning.These algorithms generate empirical training data for comparison with the structural metrics.Source: §2.9 · p. 13
  • TQC pools quantile predictions from multiple critics, drops the highest values to limit overestimation, and averages the remainder into a stable actor-update target.Each critic models a distribution of expected future returns rather than a single average.Source: §2.9 · p. 13

2.10. RL training environments

  • PPO was used for Reach, off-policy methods with HER for sparse-reward block tasks, and DDPG+HER and TQC+HER for the hardest task, BlockRotateXYZ.The two algorithms on BlockRotateXYZ compare a baseline off-policy method with a more capable one.Source: §2.10 · pp. 13–14
  • In Reach, the thumb and a randomly selected long finger must reach two nearby 3D points, forming a pinching gesture.Source: §2.10 · p. 14
  • BlockRotateZ requires stable fingertip contact and coordinated force to rotate a cube to a target orientation, with each episode shuffling it only along Z.Source: §2.10 · p. 14
  • BlockRotateXYZ rotates a block along all three spatial dimensions and used a modified success threshold of 0.4 radians instead of the canonical 0.1.The study identifies it as its most demanding environment.Source: §2.10 · p. 14

3. Results

The section evaluates morphological conditioning at the reference configuration, from individual fingers through pinch and power-grasp mappings, then compares these structural predictions with reinforcement-learning training outcomes.

3.1. Per-finger metrics

  • Joint-axis geometry slightly favored SDH long-finger conditioning but favored ACBH at the thumb; ACBH also used more actuators per DOF and improved long-finger coupling.SDH long fingers had κ(J) = 3.84, below the ACBH long-finger range of 5.77 to 9.55, while ACBH thumb κ(J) = 4.76 versus SDH 9.45. SDH was underactuated (m/n < 1), whereas ACBH was overactuated (m/n > 1), with 7 to 11 muscles per finger driving 5 or 6 DOFs; ACBH’s branching tendons yielded κ(B) < κ(J) on all four long fingers, unlike the slight worsening from SDH’s J0 coupling.Source: §3.1 · pp. 15–16
  • Actuation-stage conditioning explains why SDH’s effective mapping tracks its kinematics, while ACBH’s thumb actuation is a bottleneck despite favorable thumb kinematics.SDH κ(A) never exceeds 2 ≈1.41, whereas ACBH thumb κ(A) = 30.74 and κ(B) = 20.03 despite κ(J) = 4.76; ACBH long-finger benefits arise from directional alignment of actuation authority with kinematics, not κ(A) magnitude alone.Source: §3.1 · p. 16

3.2. Task 1 – Precision pinch

  • For precision pinch, nearly identical kinematic conditioning accompanied more than twofold higher effective actuation conditioning for ACBH.ACBH used 11 joints and 18 muscles (r = 5), with κ(J) = 9.55 and κ(B) = 20.03; SDH used 9 joints and 8 actuators (r = 3), with κ(J) = 9.45 and κ(B) = 9.45.Source: §3.2 · p. 16

3.3. Task 2 – Power grasp

  • For power grasp, the pinch pattern persisted: ACBH had higher effective conditioning, as thumb specialization set the combined-system bottleneck.ACBH engaged 26 joints and 39 muscles (r = 11), with κ(J) = 9.55 and κ(B) = 20.03; SDH engaged 22 joints and 18 actuators (r = 7), with κ(J) = 9.70 and κ(B) = 10.01.Source: §3.3 · p. 16

3.4. Authority distribution in the ACBH thumb

  • The ACBH thumb’s high κ(B) reflects anisotropic actuator authority, not rank deficiency, and is consistent with the thenar muscles’ opposition specialization.The 11 × 6 thumb moment-arm Jacobian was full-rank, but σmin(Rm,T1) was roughly 30× smaller than σmax; the IP joint relies on only two of eleven thumb muscles. This authority-distribution effect is invisible to kinematic analysis alone and emerges when A(q) is constructed, illustrating morphology as a structural inductive bias.Source: §3.4 · p. 17

3.5. Empirical RL training results

  • Across PPO and DDPG+HER/TQC+HER, outcomes ranged from SDH-leading reach and XYZ rotation to comparable Z rotation and failure for both hands under DDPG+HER.For reach, the SDH achieved 0.75 success at approximately 11 million timesteps (≈6 h), versus 0.65 for the ACBH at approximately 25 million timesteps (≈19.5 h); for BlockRotateZ, the ACBH reached 0.692 and the SDH 0.665 at approximately 16 million timesteps. Both remained below 0.1 on BlockRotateXYZ with DDPG+HER; with TQC+HER, the SDH reached ∼0.76 after ∼23 million timesteps (≈34 h), versus ∼0.25 for the ACBH (≈73 h).Source: §3.5 · p. 17

4. Discussion

The discussion finds that morphological advantages are non-uniform: their effects depend on the hand, task, and whether kinematic or actuation conditioning is considered. It connects these structural patterns to task performance and reinforcement-learning behavior, with training results partially supporting the predictions.

4.1. Structural patterns across the morphological aspects

  • Anatomical axes improve thumb kinematic conditioning, but long fingers favor SDH; branching tendons improve ACBH long-finger actuation mapping, while its thumb authority is highly anisotropic.ACBH thumb κ(B) = 20.03 contrasts with κ(J) = 4.76; this actuation outlier’s magnitude depends on simulated tendon routing and rest-length calibration, though its direction reflects documented anatomical specialization.Source: §4.1 · p. 18

4.2. Implications for manipulation tasks

  • The hands have approximately equal system-level κ(J) on the pinch task (9.55 versus 9.45), as ACBH’s thumb advantage is offset by SDH’s long-finger advantage.SDH conditioning is structurally homogeneous, whereas ACBH combines the study’s best-conditioned kinematic thumb with its worst-conditioned actuation thumb, affecting any task involving the thumb.Source: §4.2 · pp. 18–19

4.3. Implications for control and learning

  • Morphology biases control and learning at different scales: ACBH’s branching tendons benefit long-finger coordination, while its thumb’s κ(B) = 20.03 makes policy-gradient exploration ill-conditioned along MCP and IP flexion directions.This calls for targeted exploration schedules, action-space regularization, or curriculum strategies that decouple opposition from MCP and IP control; these adaptations are not required for the SDH thumb.Source: §4.3 · p. 19

4.4. Structural predictors of reinforcement learning efficiency

  • The structural predictors link actuator count to sample complexity, κ(B) to exploration isotropy, and redundancy to stability, with the best tradeoff depending on task demands.SDH’s 18 versus ACBH’s 39 actuators favors lower-dimensional learning, especially under PPO; ACBH’s per-finger redundancy (r = 2–3) can support internal posture adjustment, while SDH’s r = 1–2 may aid early convergence on straightforward positioning tasks.Source: §4.4 · pp. 19–20

4.5. Interpretation of RL training outcomes

  • Training partially supported structural predictions: PPO Reach favored SDH, TQC+HER BlockRotateXYZ exposed the ACBH thumb bottleneck, while DDPG+HER BlockRotateZ did not.Both hands failed BlockRotateXYZ under DDPG+HER, which was uninformative about their structural differences; ACBH’s smoother success-rate curve may reflect the stabilizing effect of higher redundancy.Source: §4.5 · p. 20

5. Conclusion

The paper introduces a four-aspect framework for analyzing hand morphology and actuation, then evaluates its structural predictions against reinforcement-learning experiments. The findings show selective trade-offs rather than a uniform advantage from anatomical fidelity.

  • The framework analyzes conditioning of J, A, and their product B to attribute coordination bottlenecks to geometry, actuation, or their interaction, using parameters derived from canonical digital representations.Source: §5 · p. 21
  • Across PPO, DDPG+HER, and TQC+HER, the SDH converges faster and more reliably under straightforward training; the ACBH may benefit from reward shaping or curricula.Source: §5 · p. 21
  • The SDH’s smaller action space (m=18) predicts on-policy sample-complexity gains, while ACBH’s high κ(B) predicts anisotropic exploration and slower convergence when tasks require consistent thumb participation.Source: §5 · p. 21
  • Networked coupling gives ACBH κ(B)<κ(J) at all four long fingers, but its thenar group shows a 21× imbalance in joint-level actuator authority with manipulation consequences.Source: §5 · p. 21
  • Joint-axis geometry offers no kinematic advantage at long fingers but can improve thumb conditioning; actuator-to-DOF ratio provides proportional per-finger redundancy without directly improving conditioning.Source: §5 · p. 21
  • Anatomical fidelity selectively invests across four aspects, while simplified designs trade expressiveness and redundancy for uniform actuator authority and smaller policy spaces; testing softbodied and underactuated hands remains future work.Source: §5 · p. 21

CRediT authorship contribution statement · A. Kinematic parameters of the SDH and ACBH

The contribution statement assigns Zalán Tari primary analytical and drafting roles, with software, resource, editorial, supervisory, and funding contributions shared among the authors.

CRediT authorship contribution statement

A.1. SDH Denavit–Hartenberg parameters

This section documents the Shadow Dexterous Hand’s Craig modified Denavit–Hartenberg parameters, extracted from its URDF model at zero configuration.

A.1. SDH Denavit–Hartenberg parameters

A.2. SDH PoE screw axes

This subsection documents the Shadow Dexterous Hand’s PoE screw axes at zero configuration.

A.2. SDH PoE screw axes

  • The axes are represented as ξ_i = [ω_i; p_i × ω_i] in the palm-fixed space frame and were derived from the official URDF.The axes are specified at zero configuration.Source: Table 4 · p. 27

A.3. ACBH Denavit–Hartenberg parameters · A.4. ACBH PoE screw axes

These appendices document the ACBH’s modified D-H parameters and PoE screw axes by finger. The D-H description absorbs each distal joint into the end-effector transform, while small residual α values reflect oblique anatomical joint axes.

A.3. ACBH Denavit–Hartenberg parameters

  • The ACBH D-H tables report parameters for the index and middle fingers, with each distal joint absorbed into end-effector transform T0.The distal joints are the DIP joints for long fingers and the IP joint for the thumb.Source: §A.3 · p. 27, Table 5 · p. 27, Table 6 · p. 27
  • The ACBH D-H tables report parameters for the ring and little fingers, with each distal joint absorbed into end-effector transform T0.The distal joints of long fingers are DIP joints.Source: §A.3 · p. 27, Table 7 · p. 28, Table 8 · p. 28
  • The ACBH thumb D-H table reports parameters with the distal IP joint absorbed into end-effector transform T0.Parameters were extracted from the ACBH MuJoCo model at zero configuration.Source: §A.3 · p. 27, Table 9 · p. 28
  • Small residual α values in the ACBH D-H parameters reflect oblique anatomical joint axes, not numerical inconsistencies.The parameters are extracted at zero configuration from the ACBH MuJoCo model.Source: §A.3 · p. 27

A.4. ACBH PoE screw axes

B. Reinforcement learning training details … B.3. Hyperparameters

The block presents success-rate curves for the two hands across three tasks and training algorithms, alongside the computational environment and selected training hyperparameters.

B.1. Success-rate curves

  • Supplementary Figure 1 plots PPO success rates over time for the SDH and ACBH in the Reach task.Values are smoothed with an exponential moving average of weight α=0.6.Source: Figure 1 · p. 30
  • Supplementary Figures 2 and 3 plot DDPG+HER success rates over time for the SDH and ACBH in BlockRotateZ and BlockRotateXYZ.Both use exponential moving-average smoothing with weight α=0.6.Source: Figure 2 · p. 31, Figure 3 · p. 31
  • Supplementary Figure 4 plots TQC+HER success rates over time for the SDH and ACBH in BlockRotateXYZ with rotation_threshold set to 0.4 radians.Values are smoothed with an exponential moving average of weight α=0.6.Source: Figure 4 · p. 32

B.2. Computational environment

  • Experiments used Stable-Baselines3 2.7.0 with MuJoCo 3.10.0, running 16 parallel Reach instances and 19 instances for each block-manipulation variant.Runs were performed locally on a workstation with an NVIDIA GeForce RTX 5060 Ti (16 GB), Intel Core i5-14400F, and 32 GB DDR4 memory.Source: §B.2 · p. 32

B.3. Hyperparameters

  • The DDPG+HER hyperparameters specify MultiInputPolicy, four future-sampled goals, learning_starts = 250 * 19 * 10, and a [256, 256, 256] network with two critics.The passage defines learning_starts as max_episode_steps * num_envs * 10.Source: §B.3 · p. 33
  • A separate hyperparameter listing specifies MultiInputPolicy, four future-sampled goals, learning_starts = 100 * 19 * 10, and a [256, 256] network with 25 quantiles.The passage defines learning_starts as max_episode_steps * num_envs * 10.Source: §B.3 · p. 34

Related papers

Loading 2609.05206v1…