RobotGuide
Math foundations/Slip detection: from «I felt it» to «I computed it»

Foundations T27 / 27 · Contact & convex geometry

Slip detection: from «I felt it» to «I computed it»

Slip is not an instantaneous event: the tangential force «pre-slips» at the edge of the contact first, and the approach of the ratio |f_t|/f_n toward μ is a measurable precursor. Turning the precursor into a detector is a hypothesis-testing problem — and the optimal detector is given by Neyman-Pearson.

About 50 min3 theorems13 proof steps4 self-test questions

One-line takeaway

The stick-slip precursor is more separable in «amplitude» than in «derivative»: the ratio detector scores AUC 1.000, the rate-of-change detector only 0.549. The key to grip force regulation is not «feeling the slip»; it is seeing it in the ratio 80 milliseconds before the slip happens.

The unified tactile representation studies three quantities — contact position, force magnitude, and temporal evolution — and the observability of the latter two originates in this article. Why tactile sensing needs kHz sampling and why the grip force regulation loop should be event-triggered: you will have both answers by the end.

Why this tool exists

Carrying a full glass of water across the room: did you feel the instant the glass started to slip? You did not — you felt the precursor of your fingertips «getting tighter» over a short window beforehand, and you added force early. If you wait until it actually slips, the water is already spilled. The human hand is an event-triggered feedforward controller whose input is the slip precursor.

For a robot to reproduce this behavior, two things are missing: the physics of the precursor — what happens on the contact surface before slip, and which quantity changes monotonically; and the mathematics of the decision — when that quantity becomes a binary «add force / don't add force» decision, how do you weigh the cost of false alarms (crushing the egg) against misses (dropping the glass). These map respectively to the Cattaneo-Mindlin theory of contact mechanics and the Neyman-Pearson lemma of statistics.

One widely repeated misconception needs correcting first: many people believe the slip precursor lives in the «rate of change» of the signal (the derivative), because «stick-slip is oscillation». Real runs say the opposite: the amplitude ratio detector scores AUC 1.000, the rate-of-change detector 0.549 — close to blind guessing. The lesson is in the numerical experiment section.

Definitions

Definition 5.1 Partial slip

As the tangential load grows from zero, the normal pressure is smallest at the edge of the contact region (the Hertz distribution vanishes at the edge, T11), so the Coulomb condition |f_t|/p is violated locally first — the edge slips while the center still sticks. As the tangential force grows, the slipping annulus contracts inward, and the overall |f_t|/f_n rises monotonically toward μ. Macroscopic slip happens only after the annulus swallows the center.

Why define it this wayThis is the physical substance of the «precursor»: the monotonic climb of |f_t|/f_n during the partial slip stage. It is not noise; it is a deterministic mechanical process — and therefore detectable.

Definition 5.2 Stick-slip oscillation

When the friction coefficient drops with relative velocity (static μ_s > kinetic μ_k, the Stribeck effect), the sliding phase under elastic drive can go unstable: a limit-cycle oscillation alternating between sticking and sliding. The «squeak» a tactile sensor hears is this limit cycle.

Why define it this wayStick-slip happens after slip has already occurred — detecting it matters not for «anticipation» but for «confirmation». Precursor detection targets partial slip (Definition 5.1), not stick-slip. The two are routinely conflated.

Definition 5.3 Binary hypothesis testing

H₀: stable hold (no action needed) versus H₁: slip imminent (add force). Detection statistic T (e.g., the windowed maximum of |f_t|/f_n), threshold θ: alarm when T ≥ θ. Detection rate P(T ≥ θ | H₁), false alarm rate P(T ≥ θ | H₀). The ROC curve is the locus of detection rate versus false alarm rate as θ varies.

Why define it this wayThis turns «feel» into an optimizable quantity. The cost of a false alarm = crushing the object (wasted force, fragile parts scrapped); the cost of a miss = the object falls. The ratio of the two decides the optimal threshold — this is not a feeling problem, it is a pricing problem.

Theorems and proofs

Theorem 5.1 The edge slips first (the edge argument of the Cattaneo-Mindlin mechanism)

For a circular contact region (Hertz pressure distribution) under a gradually increasing tangential load F_t, local slip appears first at the edge of the contact region, and the slipping annulus contracts inward as F_t grows; before macroscopic slip, the overall ratio |F_t|/F_n rises monotonically from 0 toward μ.

Proof

Local Coulomb condition: not slipping at radius r inside the contact region requires |local tangential stress q(r)| ≤ μ·p(r). T11 proved that p(r) = p₀√(1 − r²/a²) vanishes at the edge r → a (with a diverging gradient).

Edge behavior: p(a) = 0, while for any nonzero tangential load the tangential stress component at the edge is nonzero (continuity of the elastic stress field: the edge is a geometric singular point, where the tangential stress tends to a finite value or vanishes more slowly — rigorously, the Cattaneo-Mindlin solution gives q(r) the same-order edge behavior ∝ √(1−r²/a²), and its ratio to p(r) tends at the edge to the limit q₀/p₀, which grows in step with the load ratio |F_t|/F_n).

So the ratio |q/p| reaches μ at the edge first: the slipping annulus grows inward from r = a, with |q| = μp (saturated) inside the annulus and |q/p| < μ (sticking) outside. F_t grows → the annulus moves inward. The center slips last: when the annulus swallows the center r = 0, |F_t|/F_n = μ and macroscopic slip begins.

Monotonicity: F_t = ∫|q| dA = μ·(integral of p over the annulus area) + (contribution of the sticking region) is a continuous concave function of F_t (saturated-region contribution linear, sticking-region contribution sublinear), and its ratio to F_n rises monotonically from 0 toward μ. ∎ (The full closed-form solution is the work of Cattaneo 1938 and Mindlin 1949: the slipping annulus radius c satisfies an explicit algebraic relation; this article gives the structural argument, with the closed form deferred to the bounds statement.)

NoteThe edge singularity was planted back in T11: p(r) vanishing at the edge plus a diverging gradient doom the edge to break first. Partial slip is not «approximate slip»; it is the asymptotic expansion of slip — both are governed by the same Coulomb condition, at different spatial stages.

Theorem 5.2 Condition for stick-slip oscillation

A spring-slider system (stiffness k, mass m, belt speed v_b, regularized friction μ(v) = μ_k + (μ_s − μ_k)e−|v|/v_s) loses stability at the sliding-phase equilibrium point v = v_b ⟺ the belt speed falls in the «negative damping» interval of the Stribeck descending branch; after the instability the system enters a stick-slip limit cycle.

Proof

Sliding-phase equation: m·ẍ = k(x₀ + v_b t − x) − F_f(v), where F_f(v) = μ(v)·N·sign(v), v = ẋ. Use sliding-phase moving coordinates y = x₀ + v_b t − x (spring extension), v = ẋ = v_b − ẏ.

The equilibrium y* = μ(v_b)N/k has v = v_b. Linearize: small-perturbation analysis of δÿ. Differentiating in y: m·δÿ = −k·δy − F_f'(v_b)·δv, δv = −δẏ, so m·δÿ + F_f'(v_b)·(−δẏ)·(−1)? Term by term: d(−F_f)/dẏ = −F_f'(v_b)·dv/dẏ = −F_f'(v_b)·(−1) = F_f'(v_b). Equation: m·δÿ − F_f'(v_b)·δẏ·(−1)…… written out carefully: m δÿ = −k δy + F_f'(v_b)·δẏ.

Characteristic equation: mλ² − F_f'(v_b)λ + k = 0.

On the Stribeck branch at v = v_b, μ'(v_b) < 0 (more speed, less friction), so F_f'(v_b) = μ'(v_b)N < 0. For the characteristic equation mλ² + |F_f'|λ + k = 0, when the discriminant = F_f'² − 4mk > 0 both roots are positive (two real roots: λ = (F_f' ± √(F_f'²−4mk))/(2m), with numerator F_f' < 0: one root is positive — divergent); when the discriminant < 0 the complex roots have real part = F_f'/(2m) < 0? Recompute: λ = (F_f' ± √(F_f'² − 4mk))/(2m); when F_f' < 0 and |F_f'| > 2√(mk), both roots are positive real → instability.

Conclusion: when |F_f'(v_b)| > 2√(mk), i.e., the Stribeck slope is steep enough (or k, m small enough), the sliding phase destabilizes and the trajectory falls into the stick phase (at v = 0 the spring force is held up by static friction); the spring force climbs again to μ_s·N and breaks loose once more — a limit cycle forms. ∎

NoteThis «negative-damping instability» structure shares a root with every friction-induced squeal (brake squeal, creaking door hinges). The oscillation frequency is set jointly by k, m and the climb time of the stick phase — in the experiment the toy parameters give a 1 Hz oscillation whose frequency does not matter; what matters is the waveform shape: a slow climb in the stick phase (spring charging) + a fast drop in the sliding phase (discharge), the sawtooth you see in a tactile sensor.

Theorem 5.3 Neyman-Pearson lemma

Testing H₀ (distribution p₀) against H₁ (distribution p₁), given a false alarm rate cap α, the likelihood ratio test (reject H₀ when p₁(x)/p₀(x) > η) has the largest detection rate among all tests with false alarm rate ≤ α.

Proof

Let φ be any test satisfying E₀[φ] ≤ α (φ: sample → [0,1]), and φ* the likelihood ratio threshold test (with η tuned so that E₀[φ*] = α). We show E₁[φ] ≤ E₁[φ*].

Pointwise comparison: φ*(x) − φ(x) is nonnegative at points where the likelihood ratio > η and nonpositive where < η (on the equality region anything goes). Hence (φ* − φ)(x)·(p₁(x) − ηp₀(x)) ≥ 0 for all x (likelihood ratio > η ⟹ p₁ − ηp₀ > 0 and φ*−φ ≥ 0; symmetrically otherwise).

Integrate: E₁[φ*] − E₁[φ] − η(E₀[φ*] − E₀[φ]) = ∫(φ*−φ)(p₁ − ηp₀)dx ≥ 0.

By construction E₀[φ*] = E₀[φ] (both tuned to false alarm rate α; if φ has false alarm rate < α, then η is taken smaller, the argument goes through the same way with E₀[φ] ≤ E₀[φ*], and the inequality still favors the conclusion), so E₁[φ*] ≥ E₁[φ]. ∎

NoteThe proof is one integration-by-parts-style pointwise completion of the square. Mechanics translation: the likelihood ratio = «the probability of this observation in the 'about to slip' world / the probability in the 'holding steady' world» — a threshold test is just queuing by that ratio. Any hand-picked threshold (e.g., «add force when the ratio exceeds 0.9μ») deserves one question: where is your η?

Work it by hand

You hold an object under a 2 N normal force; the current tangential force is 0.6 N and μ = 0.5. (1) What is the slip margin? (2) If the cost ratio of a false alarm (extra force crushing the object) to a miss (dropping it) is 1:10, and the detection statistic is approximately Gaussian — N(0.45, 0.01) while holding, N(0.55, 0.01) when about to slip — where is the optimal threshold?

  1. (1) Margin = μF_n − F_t = 0.5×2 − 0.6 = 0.4 N. In ratio terms: the current |F_t|/F_n = 0.3, a ratio margin of 0.2 below μ = 0.5.
  2. (2) The likelihood ratio threshold η = cost ratio = false alarm cost / miss cost = 1/10 = 0.1. With two equal-variance Gaussians the likelihood ratio threshold reduces to a shift from the midpoint of the means: threshold θ = (μ₀+μ₁)/2 + σ²·ln(η)/(μ₁−μ₀) = 0.5 + 0.01×ln(0.1)/0.1 = 0.5 − 0.23 = 0.27? — this number is below both means, which means with a cost ratio of 1:10 you almost always alarm (a miss is too expensive) and the threshold collapses to a very low value.
  3. Check the semantics: a miss costs 10× more ⟹ the threshold should lean toward «prefer false alarms» ⟹ the threshold drops ✓ (0.27 < 0.45 = the holding mean, meaning most holding samples also alarm — unless the costs were understood the other way around. In engineering the cost ratio is often reversed: crushing is expensive ⟹ the threshold rises to 0.5 + 0.23 = 0.73, acting almost only when slip is imminent). You must be able to compute both directions — the key is to ask first which side is more expensive.
Answer

Margin 0.4 N (0.2 in ratio terms); NP threshold = midpoint ± σ²·ln(cost ratio)/(mean gap), with the sign decided by «which side is more expensive». The real point of this problem: price first, then talk thresholds.

Run the numbers

Where the numbers come from

(a) Spring-slider with regularized friction (m=50 g, k=20 N/m, v_b=20 mm/s, μs=0.5, μk=0.3); (b) ROC of slip precursor detection (2000 episodes, detection window = 80 ms before slip, 1 kHz sampling, noise σ=0.01):

  • Stick-slip oscillation: the spring force climbs to 0.50 (μs·N) at the end of the stick phase and drops to 0.10 in the sliding phase — a sawtooth waveform; this is the stick-slip signature
  • Ratio detector (windowed maximum of |f_t|/f_n): AUC = 1.000; best operating point detects 95.4% / false alarms 0.0% @ threshold 0.505
  • Rate-of-change detector (windowed maximum of dr/dt): AUC = 0.549 — close to blind guessing
  • Reversed conclusion: the precursor is in the amplitude (the monotonic climb of the ratio), not in the derivative — «stick-slip is oscillation, so differentiate» is a common misconception

Where this shows up

The object hereThe counterpart in RLUsed in
The monotonic climb of the ratio |f_t|/f_nWhere the unified tactile representation gets the observability of «force magnitude + temporal evolution»used in →
The kHz time scale of partial slipWhy tactile sampling needs kHz while 30 Hz vision suffices: the partial slip window is tens of millisecondsused in →
The NP lemma and cost ratiosThe design basis for «event-triggered» grip force regulation: price first, set the threshold secondused in →
Stick-slip = an unstable limit cycleThe mathematical identity of «contact noise» in force control: not noise, but Stribeck negative dampingused in →

Boundaries: where this article stops

Where this article stops: the full closed-form Cattaneo-Mindlin solution (an explicit algebraic relation for the slipping annulus radius), corrections for soft fingers with large contact areas, spectral detection of high-frequency micro-vibrations (used in real engineering but invisible in toy models), and learned slip detectors (replacing the NP statistic with a network output) — none are covered. Honest disclaimer: the experiment is a toy setup — the 80 ms pre-slip window, the noise level, and the cost ratio are all set values; on real hardware the numbers change with the sensor and the object surface. What does not change is the structure: ratio climb + hypothesis-testing pricing.

Self-test: 4 questions

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

01Are partial slip and stick-slip the same thing?

No. Partial slip is local edge sliding before macroscopic slip (the precursor stage of the ratio climb); stick-slip is the unstable oscillation after sliding has already begun (a limit cycle). The detector targets partial slip — it buys you lead time; stick-slip can only «confirm that slipping happened». Conflating the two is pitfall number one in slip detector design.

02Why does partial slip start at the contact edge?

The Hertz pressure distribution vanishes at the edge (with a diverging gradient, T11), so the Coulomb condition |q| ≤ μp breaks at the edge first. As the tangential load grows, the slipping annulus contracts inward from the edge and the center slips last. The edge singularity is the setup T11 planted.

03Should the detection statistic be the amplitude of the ratio |f_t|/f_n or its rate of change? What does the experiment say?

Amplitude. Measured AUC: ratio detector 1.000, rate detector 0.549 (guessing level). «Stick-slip is oscillation, so differentiate» fails at the partial slip stage — partial slip is a monotonic climb, not an oscillation. The derivative only amplifies noise.

04How do the false alarm and miss costs enter the threshold?

NP lemma: the optimal threshold is set by the likelihood ratio, and the cost ratio decides where it sits — expensive misses push the threshold left (more alarms), expensive false alarms push it right (fewer alarms). Formula: θ = midpoint + σ²·ln(cost ratio)/(mean gap). Price first, set the threshold second — the order cannot be swapped.

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