A Benchmark of Dexterity for Anthropomorphic Robotic Hands

Davide Liconti*, Yuning Zhou*, Yasunori Toshimitsu, Ronan Hinchet, Robert K. Katzschmann

*Equal contribution · Soft Robotics Lab, D-MAVT, ETH Zurich

PDF CAD (soon) Code (soon) Interactive sim BibTeX

POMDAR is a taxonomy-grounded dexterity benchmark for anthropomorphic hands. It formalizes dexterity as task throughput (correctness and speed) across vertical, horizontal, continuous-rotation, and pure-grasping configurations—implemented in the real world and in simulation.

Dexterity is central to anthropomorphic hands, but inconsistent definitions make comparisons across designs difficult. POMDAR is a benchmark that treats dexterity as task performance over manipulation and grasping motions grounded in established motor-control taxonomies. Mechanical scaffolding constrains motion, limits compensatory strategies, and yields unambiguous metrics. Scores combine task correctness and speed against a human baseline; the protocol is open and standardized for reproducible comparison of dexterous platforms.

POMDAR benchmark overview: four manipulation configurations and quantitative evaluation across ORCA embodiments

Fig. 1 (teaser). Four manipulation configurations and embodiment study overview.

Anthropomorphic hands are judged by kinematic proxies as often as by real manipulation outcomes. POMDAR instead follows a performance-based paradigm: the benchmark tasks are traceable to manipulation and grasp taxonomies, and scores reflect what the hand actually achieves on contact-rich interactions—not only workspace or DoF counts.

  • Representative tasks grounded in Elliott & Connolly, Ma & Dollar, and Feix et al. (GRASP).
  • Reproducible hardware: consumer 3D printing, shared components across tasks, compact footprint.
  • Observable motions via scaffolding that limits unintended arm/wrist compensation.
  • Throughput score: weighted combination of correctness (0.8) and speed vs. human baseline (0.2).

Benchmark construction

Tasks consolidate overlapping taxonomy entries into a practical set: 12 manipulation patterns (vertical, horizontal, and continuous rotation) and 6 pure grasping tasks. Vertical and horizontal setups use scaffolded rails and notches; continuous tasks use a gravity-based clutch; grasping tasks isolate pick-up and relocation without external supports.

Vertical (V1–V3)

Wheel, stick, and sphere: progress through discrete notches while maintaining in-hand control.

Horizontal (H1–H5)

Scissors, chopsticks, squeeze, palmar, pinch: translation along curved rails with increasing difficulty.

Continuous (C1–C4)

Thread, stick, wheel, fidget: sustained rotation using the clutch and geared mechanisms.

Grasping (G1–G6)

Wheel, sphere, disk, and three cylinder sizes—relocation in free space to test grasp quality.

POMDAR task configurations: vertical, horizontal, continuous rotation, and grasping panels

Benchmark apparatus and task configurations (overview figure).

Interactive simulation

Explore MuJoCo models of the benchmark objects in your browser: orbit the camera, pause, and drag bodies to apply forces. The first load downloads the physics engine and meshes and may take a few seconds.

Open the viewer full screen

Teleoperation Results

Real-world experiments teleoperate ORCA hand embodiments (2-, 3-, 5-finger without abduction, and full 16-DoF) on a Franka arm using Rokoko motion-capture gloves. Radar plots summarize per-task scores for each embodiment (manipulation and grasp panels). Click a task below to watch representative teleoperation clips.

POMDAR results: ORCA embodiments and radar plots for manipulation and grasp tasks
Manipulation tasks (Fig. 6B)
Grasp tasks (Fig. 6C)

Citation

@article{liconti2026pomdar,
  title={A Benchmark of Dexterity for Anthropomorphic Robotic Hands},
  author={Liconti, Davide and Zhou, Yuning and Toshimitsu, Yasunori and Hinchet, Ronan and Katzschmann, Robert K.},
  journal={arXiv preprint},
  year={2026}
}

Replace the BibTeX entry with the final venue information when available.