Grasp Quality Metric

• Also called: grasp quality measure, grasp score
• Category: Planning & manipulation › Manipulation
• Tags: grasping, grasp-analysis, wrench-space, manipulation, datasets

A score for how good a grasp is, for example how large a disturbance wrench its contacts can resist, used to rank candidate grasps, guide grasp optimization and label grasp datasets.

Definition

A grasp quality metric assigns a scalar score to a grasp so that candidates can be compared and the best one chosen. Analytic metrics work from a model of the contacts between hand and object: their positions, surface normals and friction coefficients. The most widely used is the epsilon metric of Ferrari and Canny (1992). Each contact's friction cone is approximated by a set of edge forces, each force is converted into a wrench about a reference point, usually the object's center of mass, and the convex hull of these primitive wrenches forms the grasp wrench space. The epsilon quality is the radius of the largest ball centered at the origin that fits inside that hull: the largest disturbance wrench magnitude the grasp can resist in every direction when the sum of contact normal forces is limited to one. A variant that limits each contact's force separately builds the hull from the Minkowski sum of the contact wrench sets, and a related volume metric uses the volume of the hull instead of the ball.

The epsilon metric is positive exactly when the origin lies strictly inside the grasp wrench space, which is the condition for force closure, so it extends the yes-or-no force-closure test into a graded score. It has well-known sensitivities: forces and torques have different units, so the value depends on how torque is scaled, often by a characteristic object radius, and on the chosen reference point, and it assumes exact rigid contacts and a known friction coefficient. Other analytic metrics measure properties of the grasp matrix, such as its smallest singular value, or of the hand configuration, such as distance from joint limits or singularities.

Because analytic scores do not always predict real outcomes, many systems use empirical or learned metrics. A grasp can be scored by its success rate in simulation under perturbed object pose, friction and shape, or by its robustness to such uncertainty, as in the Dex-Net line of work, and neural networks can be trained to predict such a score directly from images or point clouds. Large grasp datasets filter or label candidates with one of these metrics, so the choice of metric shapes what a learned grasp planner later treats as a good grasp.

Why it matters

Grasp planning is a search, and a search needs an objective. The quality metric is that objective: it decides which grasps a planner executes and which grasps a dataset counts as positive. Knowing what a metric measures and what it ignores helps explain why a grasp that scores well in analysis can still slip on the real hand.

Key applications

• Ranking candidates in analytic grasp planners such as GraspIt!
• Labeling large synthetic grasp datasets
• Objective or loss term in grasp optimization
• Training targets for learned grasp quality networks
• Comparing robot hand designs by the grasps they can achieve
• Choosing regrasps during in-hand manipulation

In practice

A team generating a dexterous grasp dataset samples hand poses around thousands of object meshes and, for each candidate, finds the contact points between the fingers and the surface. They approximate each contact's friction cone with an eight-sided pyramid for an assumed friction coefficient, compute the epsilon metric about the object's center of mass with torques divided by the object's bounding radius, and discard grasps whose score is near zero. Because the analytic score assumes perfect contacts, they then test the remaining grasps in a physics simulator, applying gravity along several directions in turn, and keep only grasps that hold the object in every test. The dataset records both the analytic score and the simulated outcome, so users can see where the two disagree.

Common misconceptions

• A high epsilon score guarantees a successful grasp — it assumes exact contact locations, rigid bodies and known friction, and ignores whether the hand can actually apply the required forces
• A zero score means the object will fall — zero means no force closure, but a grasp can still hold an object against a known load such as gravity
• The epsilon value is a physical property of the grasp — it depends on torque scaling, the reference point and the friction cone approximation, so values are comparable only under the same conventions
• A grasp quality metric is the same as force closure — force closure is a binary condition; a quality metric grades grasps, and the epsilon metric is positive exactly for force-closure grasps

Who should know this

• Grasp planning researchers
• Dexterous manipulation engineers
• Teams building synthetic grasp datasets
• Robot hand designers

Related technologies

• GraspIt!
• Dex-Net
• Grasp Wrench Space
• Qhull
• Grasp Matrix
• Physics Simulators

Related terms

• Force Closure
• Grasp Planning
• Friction Cone
• Wrench
• Antipodal Grasp
• Eigengrasp

Related projects

• LeRobot: HuggingFace ML library for real-world robotics in PyTorch
• Isaac Lab: NVIDIA robot learning framework: one task API across Newton, PhysX and Isaac Sim
• NVIDIA Isaac GR00T: NVIDIA open foundation model for humanoid robot reasoning and skills
• OpenVLA: 7B-parameter Vision-Language-Action model for robotic manipulation
• Drake: MIT robotics toolbox for model-based dynamics, optimization, and control system design
• OpenArm: 7DOF humanoid arm for physical AI research in contact-rich environments