💻 computer science

KaRMA: A Kinematic Metric for Fine Manipulation Ability in Robotic Hands

This paper introduces KaRMA, a kinematic-only metric that quantifies the fine manipulation ability of robotic hands by measuring reachable in-hand translation and reorientation of a spherical object through feasible rolling motions, demonstrating its superiority over traditional static metrics in distinguishing dexterity and revealing tradeoffs invisible to existing baselines.

Original authors: Martin Peticco, Pulkit Agrawal

Published 2026-05-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Martin Peticco, Pulkit Agrawal

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to judge how "dexterous" (skillful with fingers) a robotic hand is. Traditionally, engineers have looked at simple stats, like "How many joints does it have?" or "How big is the area its fingers can reach?"

The authors of this paper argue that these stats are like judging a chef only by how many knives they own, rather than watching them actually chop an onion. Just because a hand has many joints doesn't mean it can smoothly move an object inside its grip without letting go.

To fix this, they created a new test called KaRMA (Kinematic Rolling Manipulation Ability). Here is how it works, explained simply:

The Core Idea: The "Rolling Pin" Test

Instead of asking a robot to pick up a cup or screw in a bolt (which depends on the robot's brain, sensors, and motors), KaRMA asks a very specific, physics-based question:

"If I pinch a small, smooth ball between the robot's thumb and index finger, how far can I roll that ball around inside the grip without ever letting go or changing my grip?"

Think of it like trying to roll a marble between your own thumb and forefinger. You want to move the marble from the tip of your fingers to the base, and spin it around, all while keeping it touching your skin. KaRMA measures exactly how much space the robot can cover doing this.

The Three Scores KaRMA Gives

The test doesn't just give one number; it gives three scores to tell a complete story:

KaRMA-T (The "Travel" Score): This measures how far the robot can slide the ball from side to side or up and down inside the grip. It's like asking, "How much of the room can this marble visit?"
KaRMA-R (The "Spin" Score): This measures how well the robot can rotate the ball. Can it spin the ball 360 degrees? Can it tilt it? Some hands are great at moving the ball around but bad at spinning it, and this score catches that.
KaRMA-S (The "Lucky Start" Score): This measures how picky the robot is about where it starts.
- A high score means the robot is robust: no matter where you start pinching the ball, it can still move it well.
- A low score means the robot is fragile: it can only move the ball well if you start in one very specific, perfect position. If you start slightly off, it gets stuck.

How They Tested It

The researchers took 16 different robotic hands (from simple ones with 3 fingers to complex ones with many) and ran this simulation. They didn't use real motors or cameras; they just used the mathematical "skeleton" (kinematics) of the hands.

They made sure the robot followed strict rules:

No letting go: The ball must stay touching the fingers the whole time.
No crashing: The ball can't hit the robot's wrist or other fingers.
No magic: The robot can't use force to push the ball through its own joints; it has to move naturally.

What They Found

The results were surprising and showed why the old "count the joints" method fails:

More Joints $\neq$ More Skill: Some hands with fewer joints performed better at rolling the ball than hands with many joints. The old metrics (like "workspace volume") often ranked these hands incorrectly.
Translation vs. Rotation are Different: A hand might be great at sliding a ball (high Travel score) but terrible at spinning it (low Spin score). The old metrics couldn't see this difference; they just gave one overall rank. KaRMA separates these skills, helping engineers choose the right hand for the right job.
The "Hidden" Limits: The study showed that even if a hand's fingers look like they can reach a lot of space, the robot's own joints often get in the way. When they added rules for joint limits and collisions, the "reachable space" for many hands shrank by 60% to 99%. This proves that just looking at the size of the hand isn't enough; you have to see how the joints actually move together.

The Bottom Line

KaRMA is a new "report card" for robotic hands. It ignores the robot's brain and sensors to focus purely on the mechanical ability to roll an object inside a pinch. It tells us not just if a hand can move things, but how well it can slide and spin them, and how forgiving it is if you don't start in the perfect position. This helps designers build better hands and helps buyers pick the right one for their needs.

Technical Summary: Kinematic Rolling Manipulation Ability (KaRMA)

Problem Statement

Traditional metrics for evaluating robotic hand dexterity rely on static properties such as workspace volume, manipulability indices (e.g., Yoshikawa), and grasp stability. While useful for local configurations, these metrics fail to quantify the standard definition of dexterity in robotic manipulation: the ability to continuously change an object's pose within the hand while maintaining contact from an initial grasp. Existing proxies, such as joint counts or fingertip workspace intersections, do not directly measure the reachable set of in-hand translations and reorientations under realistic constraints like rolling contact, joint limits, and collision avoidance. Furthermore, task-based benchmarks often depend heavily on controller design, sensing, and training, making them unsuitable for objective, model-based comparison of hand kinematics alone.

Methodology

The paper introduces Kinematic Rolling Manipulation Ability (KaRMA), a kinematics-only metric designed to quantify fine manipulation capabilities based on a hand's kinematic model (URDF).

Core Assumptions and Scope

Grasp Type: A two-finger precision pinch between the thumb and index finger.
Object: A small spherical test object (removes edge/corner effects).
Constraints: The metric enforces maintained rolling contact, joint limits, collision avoidance, and antipodal force feasibility. It explicitly excludes regrasping and finger gaiting.
Scaling: To ensure comparability across different hand sizes, all length parameters are non-dimensionalized using a characteristic length $L_{ref}$ derived from the URDF (combining maximum knuckle distance and median finger chain length).

Computation Pipeline

KaRMA computes the reachable set of object poses via a two-phase Breadth-First Search (BFS):

Seed Selection: Deterministically generates feasible initial pinch configurations using Sobol samples and inverse kinematics.
Phase 1 (Translation): Expands the search over a voxel grid aligned with the seed's manipulability ellipsoid. Each step solves a Quadratic Program (QP) to find a local joint displacement that achieves a target sphere translation while maintaining rolling contact constraints.
Phase 2 (Rotation): For each reachable voxel, the algorithm explores local rotational primitives around controllable axes to record reached orientation bins (discretized using HEALPix).

Reported Scores

KaRMA outputs three distinct scores:

KaRMA-T (Translational Ability): The dimensionless ratio of the volume of reachable voxels to a reference cube ( $L_{ref}^3$ ). It represents the fraction of the workspace reachable via rolling.
KaRMA-R (Rotational Ability): The mean fraction of orientation bins (out of 228) reachable within the top five best-performing translational voxels.
KaRMA-S (Initial Grasp Sensitivity): The ratio of the median translational score to the best translational score across feasible seeds. A value near 1 indicates robustness to the initial grasp; a value near 0 indicates high sensitivity.

Key Contributions

A Kinematics-Only Dexterity Metric: KaRMA is the first metric to quantify the reachable set of in-hand object pose changes under rolling contact in a two-finger pinch, strictly based on kinematic models without requiring controllers or dynamic parameters.
Practical Computation Procedure: The paper provides a standardized algorithm that evaluates translational ability, rotational ability, and seed robustness, accounting for coupled joints and complex constraints (collision, antipodal feasibility).
Comprehensive Evaluation: The metric is applied to 16 widely used robotic hands (ranging from 3 to 9 DOF), demonstrating its ability to distinguish between hands that rank identically under static proxies.

Results

The evaluation of 16 robotic hands yielded several critical findings:

Dynamic Range: KaRMA-T spans three orders of magnitude (0.0003 to 0.098), offering significantly better discrimination than DOF count or Global Conditioning Index (GCI).
Decoupling of Translation and Rotation: KaRMA reveals a trade-off between translational and rotational capabilities that static baselines miss. For example, the Allegro hand ranks 2nd in translation but 1st in rotation, while Sharpa ranks 4th in translation but 5th in rotation. No existing baseline predicts this divergence.
Limitations of Static Proxies:
- Workspace Intersection: While strongly correlated with KaRMA-T ( $\rho_s = 0.92$ ), it misranks 31% of hands by $\ge 2$ positions because it ignores collision and joint limit constraints.
- Jacobian-based Metrics: Metrics like Yoshikawa manipulability and GCI show no correlation with KaRMA scores. Low-DOF hands often have trivially well-conditioned Jacobians (high GCI) but poor in-hand manipulation ability.
Constraint Impact: Ablation studies show that joint limits are the dominant constraint, reducing unconstrained translational workspace by 61–99% for capable hands. Collision avoidance and antipodal feasibility further reduce the reachable set.
Robustness: The metric is invariant to world-frame translation and rotation. It is sensitive to joint limit specifications and sphere radius but robust to small variations in link geometry.

Significance and Claims

The authors position KaRMA as a standardized, objective tool for hand procurement and kinematic design iteration. Its significance lies in:

Objective Comparison: It allows for the comparison of different hand designs directly from their kinematic models (URDFs) without the variability introduced by control policies or task-specific training.
Revealing Hidden Trade-offs: It exposes that high translational reach does not guarantee high rotational dexterity (and vice versa), a distinction critical for task-specific hand selection (e.g., assembly vs. reorientation).
Consistency with Task Benchmarks: In cases where prior task-based benchmarks exist (e.g., Schunk SVH vs. Inspire), KaRMA aligns with published performance results where simple kinematic proxies (like GCI) fail or suggest the opposite ordering.

The paper explicitly states that KaRMA is not a general grasp planner or a predictor of task success on arbitrary objects. It is a "lower bound" metric for thumb-index rolling-pinch manipulation, intended to complement, rather than replace, full task benchmarks. Future work is identified as extending the metric to multi-finger grasps and non-spherical objects.