IROS 2026 · Pittsburgh

Function-based Parametric
Co-Design Optimization of Dexterous Hands

Amin Mirzaee Harsh Gupta Wenzhen Yuan

University of Illinois Urbana-Champaign

Paper Code Video BibTeX

Every hand here is generated by our model. Turn it, flex it, switch designs.

Loading generated hand…

Bends every flexion joint toward its upper limit, within the exported joint ranges.

Fingers + thumb
–
Revolute joints
–
Palm size
–
Triangles
–

These nine hands are exact generation_v2 meshes, assembly transforms, joint axes, and joint limits. Joint motion is kinematic only, moving through a portion of the full range. Download this design’s configuration.

The full pipeline

HandCDO searches a single parametric design space covering palm shape, finger layout, kinematics, fingertips, and contact-surface curvature. Each candidate is scored in simulation, and the best designs are printed, assembled, and tested on real tool use.

Hover or tap a stage.

HandCDO framework: (A) hand optimizer parametric sampling, (B) hand generation, human-demonstrated grasp pose, (C) grasp pose sampling, (D) joint space sampling, (E) grasp initialization, (F) wrench-space test score; grasp optimization iterates and the grasp score returns to the optimizer

Abstract

Despite advances in dexterous hand manipulation, robotic hand design is still largely decoupled from task-driven evaluation and control, limiting systematic optimization. Existing robotic hand co-design approaches are often limited in scope, optimizing a small subset of design parameters. We introduce a comprehensive parametric framework for robotic hand generation that unifies palm structure, finger kinematics, fingertip geometry, and fine-scale surface curvatures within a single design space. Fine geometric features are introduced through parametric surface deformation kernels that directly influence contact interactions. We validate the framework on design optimization in grasp stability tasks in simulation and real-world dynamic scenarios. Our framework produces simulation- and fabrication-ready hand models, enabling rapid design iteration for dexterous hand co-design optimization and cross-embodiment policy training and control research.

How the co-design loop works

An outer optimizer proposes hands; an inner grasp search finds out how well each one can hold the task’s tool.

  1. A

    A tree-structured Parzen estimator (TPE) proposes the next hand: finger number, angles, offsets, link lengths, finger and thumb codes, fingertip size, and surface kernels. Past high and low scores shape where it samples next.

  2. B

    The parametric generator turns those numbers into meshes and a URDF, ready for simulation and for fabrication.

  3. Task

    Each task starts from a demonstrated power grasp and tool trajectory, recorded with the object’s 6-D pose. These define what the hand must hold and how the tool will move.

  4. C

    An inner search samples where the hand meets the tool, around the demonstrated grasp.

  5. D–E

    For each pose, joint configurations are sampled and the grasp is initialized in simulation.

  6. F

    The grasp is tested against disturbances along and about each axis. That score returns to the outer optimizer, which proposes the next hand.

Each iteration, TPE proposes 8 new designs (yellow), drawn mostly near past high scorers. Once scored, every past design is re-split into high (green) and low (red).
Candidate hands generated from sampled parameters.
A demonstrator’s hand (the MANO hand model) holding the hammer along its tracked 6-D hammering trajectory, with the tool-tip path and pose drawn.
Grasp poses are sampled within a region around the demonstrated grasp.
D: joint configurations are sampled over the finger abduction joints and the first two thumb joints. E: the hand closes around the handle into the grasp.
F: forces along and torques about X, Y and Z are applied in turn (the paper’s 12 tests); the tool moves in the grasp, and how well it is held sets the score.

One design space, from palm to fingertip

Each clip changes one family of parameters while the rest of the hand stays fixed. Rendered with the current generator.

Palm geometry

Shape, size, aspect ratio, orientation, and boundary smoothing define the palm outline.

Digit placement and count

Fingers and the thumb can sit anywhere along the palm outline, and where each one sits is itself a design parameter, as is the number of fingers.

Clips use the current generation_v2 generator. The paper’s optimization and experiments use iros_paper_codebase; both are in the repository.

What the optimizer found

The reported run searches 28 selected design parameters, scoring 8 sampled hands per iteration, and converges after roughly 200 iterations.

selected design parameters
28
hands sampled per iteration
8
iterations to converge
~200
designs fabricated
3
most important parameters
finger number, thumb normal offset
Mean and max grasp score per iteration rising from about 0.5 to about 0.8 over 250 iterations
A Optimization progress
Mean absolute SHAP value per parameter group; structural parameters and normal offset are largest
B Parameter-group SHAP values

Which parameters matter, and what they do

Each row is one design parameter, ordered by its impact on the simulated grasp score. Hover, tab to, or tap a row to see what it means, the range the optimizer searched, and a generated hand sweeping that range.

SHAP summary plot of 28 selected design parameters sorted by impact on grasp score; finger number and thumb normal offset have the largest impact

F₀ right finger (green, beside the thumb), F₁ middle (yellow), F₂ left (blue), T thumb; K₀ and K₁ are the two palm surface kernels.

Three designs, simulated and built

High-scoring design, simulation modelHigh-scoring design, fabricated
High score0.81
Mid-scoring design, simulation modelMid-scoring design, fabricated
Mid score0.60
Low-scoring design, simulation modelLow-scoring design, fabricated
Low score0.45

Left of each pair: simulation model. Right: fabricated hand. Score is grasp stability in simulation.

Real-world dynamic tool use

Hands mounted on a UR5e stir, hammer and cut. OptiTrack measures in-hand rotational slip of the tool; dashed lines are each hand’s mean. The plots draw in step with the cameras. Hover or tap a hand to highlight its curve.

Stirring

Relative in-hand rotational slip

Time (s)

Hammering

Relative in-hand rotational slip

Time (s)

Cutting

Relative in-hand rotational slip

Time (s)

Project video

Citation

@misc{mirzaee2026handcdo,
  title  = {Function-based Parametric Co-Design Optimization
            of Dexterous Hands},
  author = {Mirzaee, Mohammad Amin and Gupta, Harsh and Yuan, Wenzhen},
  year   = {2026},
  eprint = {2604.27557},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/2604.27557}
}

Open full resolution →