RobotGuide
Math foundations/Force Closure and Grasp Metrics: ε Is a Linear Program

Foundations T25 / 27 · Contact and Convex Geometry

Force Closure and Grasp Metrics: ε Is a Linear Program

«Grips firmly» can be defined rigorously, and it can be computed into a number: the radius ε of the ball inscribed in the wrench set. How much a two-finger pinch falls short of a three-finger 120° grasp, how much ε drops when you move a finger by 40°, what a friction coefficient is worth in money — all of it is the output of this one LP.

About 50 min3 theorems8 proof steps4 self-test questions

One-line takeaway

Force closure = the image of the contact cones in wrench space contains a neighborhood of the origin; grasp quality ε = the inradius of that image, and the inner computation is a linear program.

The ranking metric of grasp datasets, why dexterous hands default to three-finger configurations, the marginal value of the friction coefficient for grasp quality — behind all of them is the language of this page. After reading it you can compute the ε of a grasp with your own hands instead of just watching demo videos.

Why this tool exists

Two fingers pinch a cup: stable. Move the fingers a few degrees and pinch again: still stable, but it feels «hollow». Move a bit more: suddenly it slips — the transition from stable to unstable is not gradual, so what happened in between? To answer that, you must first be able to write «stable» as a mathematical proposition.

The way to write it is to move the problem to a new home: instead of looking at what the contact forces look like in physical space, look at which directions their resultant force and torque (together, the wrench) can cover. The external disturbances the fingers can «withstand» are exactly the complement of what these wrenches can compose. Stable = no matter which direction the world pushes from, the wrench the fingers compose can push back.

In 1992 Ferrari and Canny turned this intuition into a number: look at the set the contact forces can compose in wrench space, and take the radius ε of the largest ball centered at the origin that fits inside it. Large ε = large margin in every direction. The advantage of this definition shows immediately: computing ε is a linear program — the friction-cone linearization from the previous page plugs in here exactly.

Definitions

Definition 3.1 Wrench and grasp matrix

In the planar case, a contact force f = (f_x, f_y) applied at position p = (p_x, p_y) exerts on the object the wrench w = (f_x, f_y, p_x·f_y − p_y·f_x) (force + torque about the origin). With k contacts and m cone rays per contact, the grasp matrix G ∈ R3×(km) packs the coefficients α of all ray forces into the total wrench: w = Gα.

Why define it this wayEach column of G is the wrench of «a unit force applied along this ray». All the static information of a grasp lives in G — contact positions in the torque column, contact directions in the force columns.

Definition 3.2 The wrench set

W = G·(C₁ × C₂ × … × C_k): the set of all total wrenches obtained by picking one force from each contact's friction cone (a polyhedral cone after linearization). A finite sum of affine images of convex cones is still a convex cone (each component is closed under positive combinations).

Why define it this wayW is «the space of directions the combined finger forces can cover». Object equilibrium requires the negative of the external wrench to land in W; force closure demands more — any disturbance in any direction can be withstood.

Definition 3.3 Force closure and the ε metric

Force closure: W contains a neighborhood of the origin (equivalently, 0 is an interior point of W). The ε metric: given a per-contact force bound (e.g. Σ of ray coefficients ≤ 1), W shrinks to a bounded set, and ε = max{r : the ball of radius r centered at the origin ⊂ W}.

Why define it this wayForce closure answers «can it hold», ε answers «how well». Bounding is mandatory: a cone is closed under scaling, so infinite forces give infinite wrenches — without a force bound the «inradius» is infinite and comparison becomes meaningless.

Theorems and proofs

Theorem 3.1 Criterion for force closure

W is a finitely generated closed convex cone. 0 ∈ int W ⟺ W = R³ (in the planar case).

Proof

(⟸) W = R³ clearly contains a neighborhood of the origin.

(⟹) Suppose B(0, r) ⊂ W. Take any x ∈ R³ and set y = (r/2)·x/||x|| ∈ B ⊂ W. Since W is a cone, λy ∈ W for all λ ≥ 0; take λ = 2||x||/r and get x = λy ∈ W. Hence R³ ⊆ W, and the reverse inclusion is trivial.∎

NoteThe proof uses only scaling closure of a cone — «neighborhood + cone = whole space». This is the mathematical content of «there is no middle ground between stable and unstable»: either W covers the whole space (force closure) or it has a hard boundary in some directions (push there and it collapses).

Theorem 3.2 The LP form of ε

Given a unit disturbance direction w (||w||=1), the maximum withstandable multiple along that direction, t*(w) = max{t : ∃α ≥ 0, Gα = t·w, Σ_j α_ij ≤ 1 ∀i}, is a linear program; ε = min_w t*(w).

Proof

Check the constraints one by one: Gα = t·w is a linear equality; α ≥ 0 is a linear inequality; the per-contact force bound Σ_j α_ij ≤ 1 is a linear inequality (here α_ij is the coefficient of contact i along ray j; since rays are unit vectors, |f_i| ≤ Σ_j α_ij, so the constraint Σα_ij ≤ 1 implies |f_i| ≤ 1 — it inscribes the true «force ball ∩ cone» constraint, consistent with the linearization philosophy).

The objective max t is linear. Together these form a linear program in standard form. t*(w) is finite (the force bound caps the wrench range) and an optimum exists (the feasible set is nonempty and bounded — α=0, t=0 is feasible).∎

The outer min is a continuous optimization over the direction sphere; exact solution needs the simplex or an SOCP dual. In engineering practice you discretize directions (N uniformly spaced directions on the sphere) and approximate. Discretization is the only approximation step in this page; the error shrinks as N grows but has no simple order — stated honestly, not passed off as exact.

NoteThis LP is the thing we ran 160 times in the experiment. Note «ε = min, not avg»: grasp quality is decided by the worst direction — precisely the thing a demo video cannot show you.

Theorem 3.3 The antipodal criterion for a two-finger planar grasp

Two fingertips squeeze a disk with the contact points separated by angle 2γ. The angle between the chord (the line joining the two contact points) and each contact normal is θ = 90° − γ. A two-finger grasp is force closed ⟺ the chord lies inside both friction cones ⟺ arctan(μ) ≥ θ.

Proof

Geometry, step one: contact points A, B and the center O form an isosceles triangle (AO=BO=R) with apex angle 2γ and base angles (180°−2γ)/2 = 90°−γ. The angle between chord AB and the radius AO = 90° − base angle = γ… in precise terms: the angle θ between the chord and «the normal at A pointing toward the center» satisfies θ = 90° − γ (for the antipodal case γ=90°, θ=0 and the chord lies along the normal ✓; for a 120° angular separation, γ=60° and θ=30° ✓ — matching the 30° plugged in in the experiment).

Full constructive proof for the antipodal case (θ=0): take three basis directions and construct the resisting wrench one by one (no force bound, pure closure). fx (along the chord): both fingers take normal force n, positive combinations give any multiple; fy (perpendicular to the chord): a symmetric pair of tangential forces, each c ≤ μn, either sign, and with n large enough any fy is reachable (μ>0 guarantees c can be nonzero); torque: a tangential couple c₁ = −c₂ (opposite directions) gives pure torque, likewise reachable at any magnitude. All three basis directions lie in the positive cone of W ⟹ W = R³ ⟹ force closure.∎

General case (θ > 0): the criterion «chord inside both friction cones ⟺ arctan μ ≥ θ» is Nguyen's (1988) constructive geometric test — draw the friction cones from each contact into the other; force closure is equivalent to the «conservative region» being nonempty, and the necessary and sufficient condition is exactly that the chord lies inside both cones. This page does not reproduce that construction line by line (boundary statement), but it verifies the recovery point numerically and exactly: at angular separation 120° (θ=30°), μ = 0.55 loses closure, μ = 0.5774 is critical, μ = 0.58 recovers — the recovery point matches tan(30°) = 0.5774 exactly (see the actual numbers).

Note«Which angle do you plug in?» is the most common hand-calculation trap in grasping: at angular separation 120° the chord–normal angle is 30°, not 60° (the base angle of the isosceles triangle is 60°, and the chord–normal angle is its complement). The angle in the criterion is the chord–normal angle.

Work it by hand

Two fingers antipodally squeeze a disk, μ = 0.4, per-finger total force bound 1. Estimate the maximum withstandable multiple along the pure shear direction (fy).

  1. A pure fy disturbance needs the two fingers' tangential forces adding in the same direction and symmetric normal forces (no spurious torque). Let each finger carry normal force n and tangential force c (same direction): c ≤ 0.4n (Coulomb), n² + c² ≤ 1 (force bound).
  2. Substitute c = 0.4n: n²(1 + 0.16) ≤ 1 ⟹ n = 1/√1.16 = 0.928, c = 0.371.
  3. fy capacity = 2c = 0.743.
  4. But ε = min over all directions: the experiment gives the worst direction as the force–torque mix (−0.545, +0.628, −0.556), with ε = 0.49 < 0.743. A hand calculation of one direction gives «local capability»; ε gives «the worst over all directions» — the two almost never agree, so never substitute single-direction intuition for ε.
Answer

Hand-computed fy capacity is 0.743; the true ε = 0.49 (worst case in a mixed direction). Lesson: ε is a min, not an avg — the demo-video feel that «every direction seems fine» systematically overestimates grasp quality.

Run the numbers

Where the numbers come from

Planar grasp ε (μ=0.4, 4 extreme rays, 160 disturbance directions on the sphere, one LP per direction, simplex with Bland's rule):

  • Two-finger antipodal (0°/180°): ε = 0.490, worst direction (−0.545, +0.628, −0.556)
  • Three fingers at 120°: ε = 0.895 — 83% above the two-finger case
  • Three fingers, one finger rotated along the arc by 5°/10°/20°/40°: ε = 0.875/0.856/0.821/0.696 — smooth degradation without losing closure (the normal still passes through the center)
  • Two fingers at 120°: closure lost at μ=0.4; lost at μ=0.55, critical at μ=0.5774, recovered at μ=0.58 — exactly matching the antipodal criterion μ=tan30°=0.5774
  • Three-finger μ sweep: μ=0.1→0.294, 0.2→0.548, 0.4→0.895, 0.6→1.018, 0.8→1.269, 1.0→1.449

Where this shows up

The object hereThe counterpart in RLUsed in
ε = min directional capacityThe mathematical identity of grasp-dataset ranking metrics; the gap between demo videos and ε rankingused in →
Force closure ⟺ 0 ∈ int WWhy dexterous hands use three fingers: the third finger covers the directions two cones cannotused in →
One LP per directionThe real cost structure behind «grasp optimization is expensive»: N directions × LPused in →
The marginal value of frictionFrom μ=0.4 to 1.0, ε rises 62%: material choice (coating/texture) is a first-class citizen of graspingused in →

Boundaries: where this article stops

Where this page stops: general 3-D grasping (6-D wrench, curved-surface contacts, the 4-DoF cone of soft fingers), form closure (frictionless criterion), dynamic grasping (inertia and centrifugal effects), and the error order of directional discretization of ε — none covered. Honest statement: ε is a static metric and excludes controller quality; a high-ε grasp with bad force control still drops the object — it ranks «geometric and frictional potential», not delivery capability.

Self-test: 4 questions

All hand-computation and «find the flaw in this proof» questions — nothing to memorize. Stuck? Go back to the section.

01What is the difference between force closure and form closure?

Form closure pins the object down by contact geometry alone (normals, frictionless); force closure is allowed to borrow friction. Two fingers pinching a cup is force closure (via friction); a fixture with four walls around the workpiece is form closure. Form closure is «harder», but usually needs more contact points.

02Why does computing ε require a bound on contact forces?

Without a bound, the cone is closed under scaling and t* is infinite for almost every direction — the inradius is infinite and grasps cannot be compared. The bound truncates the cone into a bounded set, and only then is ε a finite, comparable quantity. Different papers choose bounds differently (Σα≤1 or each f≤1); check the convention before comparing ε values.

03Three fingers at 120° beat a two-finger pinch by 83% in ε — what exactly is the extra finger worth?

The two-finger wrench set is short in force–torque mix directions (tangential capacity 2c is capped by the force bound); the third finger covers those directions. Experiment: the worst direction moves away from (−0.545,+0.628,−0.556) and the capacity rises. This is also why the human three-finger pinch (thumb + index + middle) is the default precision-grasp configuration.

04Moving a finger 40° along the arc drops ε by 22% but closure survives — why is «normal through the center» so critical?

Normals through the center guarantee that each contact's torque-arm structure and directional coverage leave no «holes»; sliding along the arc only changes torque-arm lengths (capacity degrades smoothly). Move the finger off the arc (normal no longer through the center) and pure-torque dead zones appear — closure can be lost outright. The character of a position error matters more than its size.

Up next

The full series · 27 articles

T1The Matrix Is a Ledger for Rotation≈ 25 minT2From Slope to Gradient: The First Line of Every Robot Learning Algorithm≈ 25 minT3Vectors and forces: the physical reading of the dot productAbout 20 minT4Two Rulers and a Better Rulerabout 25 minT5The Ledger of One Motor Revolution: Radians, Speed, and Powerabout 20 minT6Complex multiplication is rotation: the one-dimensional version of quaternionsAbout 20 minT7Bases and coordinates: why robots put everything in matrices45 minT8Matrix Multiplication: Chaining Transformations, Not a Table OperationAbout 40 minT9The Jacobian: Rotate a Joint One Degree — How Far Does the End Effector Move?About 45 minT10Eigenvectors: The Directions a Matrix Dare Not Touchabout 45 minT11SVD and the Pseudoinverse: When There Is No Exact Solution, Find the Nearest Oneabout 50 minT12Taylor expansion: putting a straight line to work in place of a curveAbout 40 minT13Probability spaces: why «random» needs axioms50 minT14Expectation: A Rigorous Construction of «Average»About 55 minT15Conditional Expectation: What the Vertical Bar Really Means≈ 60 minT16Four inequalities: the quantitative answer to «how many samples are enough»45 minT17Norms and fixed points: where «it always converges» comes from55 minT18Markov Chains: When the Time Average Equals the Ensemble Averageabout 55 minT19The Mathematical Construction of the MDP: What One Axiom Is Worthabout 50 minT20Dynamic Programming: The Accounting for Two Iterations≈ 50 minT21Optimization: why constraints cannot rely on penalties aloneAbout 55 minT22Stochastic Approximation: Every Term in the Learning-Rate Formulaabout 55 minT23The Mathematics of Contact: Hertz Theoryabout 45 minT24Friction and the Friction Cone: The Geometry of Constraintsabout 40 minT25Force Closure and Grasp Metrics: ε Is a Linear ProgramAbout 50 minT26Contact Kinematics: Why Rolling Is Not «Sliding Along»≈ 45 minT27Slip detection: from «I felt it» to «I computed it»About 50 min