Quartz 5

A Benchmark of Dexterity for Anthropomorphic Robotic Hands (POMDAR)

Properties2
descriptionPOMDAR: standardized 18-task dexterity benchmark for anthropomorphic robotic hands, grounded in human motor control taxonomy (Elliott & Connolly, Feix GRASP). Four categories: vertical, horizontal, continuous-rotation, grasping. Scores task throughput vs human baseline.
tagsclippings, robotics, dexterous-manipulation, benchmark

3 min read

A Benchmark of Dexterity for Anthropomorphic Robotic Hands (POMDAR)

Authors: Davide Liconti, Yuning Zhou, Yasunori Toshimitsu, Ronan Hinchet, Robert K. Katzschmann Submitted: 2026-04-10 Venue: arXiv 2604.09294

Abstract

Dexterity in robotic hand design lacks a consistent, quantitative definition — existing metrics focus on kinematic properties rather than actual task performance, making it difficult to compare designs or track progress toward human-level manipulation capability. POMDAR introduces a standardized dexterity benchmark built on human motor control taxonomies (Elliott & Connolly 13 patterns, Ma & Dollar 14 patterns, Feix GRASP 33 grasp types). It comprises 18 tasks across four manipulation categories. Mechanical constraints prevent compensatory strategies, ensuring tasks isolate intended motions. Performance is scored as throughput combining task correctness (80%) and execution speed relative to human baselines (20%).

Benchmark Structure: 18 Tasks

Manipulation Tasks (12 tasks)

Vertical Configuration (V1-V3) — in-hand adjustments with angular constraints:

  • V1 Wheel: coordinated angular in-hand rotation (±15°)
  • V2 Stick: progressive angular adjustment (±30°)
  • V3 Sphere: full angular freedom (±45°)

Horizontal Configuration (H1-H5) — manipulation along curved rails:

  • H1 Scissors: two-finger coordinated movement along curved path
  • H2 Chopsticks: pinch-and-slide along increasing curvature
  • H3 Palmar: palm-based horizontal manipulation
  • H4 Pinch: precision pinch along curved rail
  • H5 Squeeze: force-regulated squeeze manipulation

Continuous Rotation (C1-C4) — sustained rotational control via gravity-based clutch:

  • C1 Thread: slow precision rotation
  • C2 Stick: fast sustained rotation
  • C3 Wheel: large-object rotation
  • C4 Fidget: multi-axis rotation

Grasping Tasks (6 tasks, G1-G6)

Cylindrical, spherical, and disk-shaped objects at varying sizes, derived from Feix GRASP taxonomy (33 types consolidated into 6). Tests grasp robustness and quality without external scaffolding.

Scoring Methodology

Score = 0.8 × Correctness + 0.2 × Speed
  • Correctness: progress fraction for scaffolded tasks; discrete (0 / 0.5 / 1) for grasping
  • Speed: ratio of human baseline time to system execution time
  • Human baseline: 6 participants, 3 trials per task, wearing MoCap gloves

Taxonomy Grounding

  • Elliott & Connolly (1984): 13 manipulation coordination patterns
  • Ma & Dollar extension: 14 patterns (added Finger Pivoting and Finger Tracking)
  • Feix GRASP Taxonomy: 33 grasp types → 6 grasping benchmark tasks
  • 16 of 33 GRASP types already appear within manipulation patterns

Key Findings

ORCA hand variants tested (same operator, 1,140 total trajectories via teleoperation):

  • 2-finger (5 DoF): can only complete grasping tasks (G1-G6), fails on all manipulation tasks
  • 3-finger (no abduction): improvement on horizontal tasks, limited rotation
  • 5-finger without abduction: significant improvement but bottlenecked by abduction constraint
  • Full 5-finger (16 DoF): best overall, pronounced advantage in abduction-requiring tasks (V/H/C)

DoF boundary finding: Abduction (thumb opposition and finger splay) is the critical DoF separating grasping-only capability from full in-hand manipulation capability.

Open Source

CAD files, simulation assets, and evaluation videos publicly available.

Graph View

Backlinks