Week 10 — Grasping and Manipulation Day 1 of 7
Why This Matters
A grasping failure is rarely a software crash. It is usually a physical event: a part slips out of a gripper mid-transfer, a glass jar cracks under clamping force, or a pick-and-place robot drops a 3 kg casting onto a concrete floor. In automotive assembly lines, a single dropped powertrain component can stop a $60,000-per-hour production cell. In warehouse automation, a mis-grasped tote can jam a conveyor and cascade into hundreds of downstream errors.
The root cause is almost never the gripper hardware alone. It is a mismatch between the assumed contact model and reality. Engineers design a grasp assuming a clean point contact with infinite friction, then discover that a slightly oily surface reduces the friction coefficient from 0.6 to 0.15. A parallel-jaw gripper that “worked in simulation” fails on the line because the simulator modeled contacts as rigid and the real fingertips are rubber with 2 mm compliance.
This lesson builds the foundation you need to reason about these failures before they happen. You will learn how contact forces are modeled, why the friction cone is the single most important geometric object in grasping, and how to compute whether a grasp will hold. Every later topic this week — force closure, grasp planning, gripper selection — depends on the contact model you build today.
Core Concepts
Contact Models: From Point to Surface
When a robot finger touches an object, the physics of that interaction determines what forces the grasp can transmit. Engineers use three standard contact models, ordered by the number of force/moment components they can transmit.
| Model | DOF transmitted | Force components | Moment components | When to use |
|---|---|---|---|---|
| Frictionless point contact | 1 | Normal force only | None | Polished, lubricated surfaces; conservative bound |
| Point contact with friction | 3 | Normal + 2 tangential | None | Hard fingertips on hard objects (steel on steel) |
| Soft-finger contact | 4 | Normal + 2 tangential | 1 (torsion about normal) | Rubber/silicone fingertips, human-like grasp |
A frictionless point contact transmits only a compressive force along the surface normal. This is the most conservative model and gives a lower bound on grasp quality. A point contact with friction (also called a hard-finger contact) transmits a normal force plus two tangential friction forces. A soft-finger contact adds a torsional moment about the contact normal — this is the model you should use for any gripper with compliant pads, which covers most industrial parallel-jaw grippers.
The choice of model matters quantitatively. A two-finger grasp on a 100 mm cube using point contacts with friction gives 6 total contact DOF (3 per finger). The same grasp with soft-finger contacts gives 8. That extra torsional capability changes whether the object can resist a twisting disturbance, which is exactly the failure mode you see when a robot rotates a grasped part 90 degrees.
The key insight: a contact model is a constraint, not just a description. It tells you which force directions are admissible. Everything downstream — the friction cone, force closure, grasp metrics — flows from this constraint.
The Friction Cone
At any contact point with friction coefficient $\mu$, the set of admissible contact forces forms a cone around the surface normal. If the normal force is $f_n$ and the tangential force magnitude is $f_t$, Coulomb friction requires:
$$f_t \leq \mu , f_n$$
Geometrically, the contact force vector must lie inside a cone with half-angle $\alpha = \arctan(\mu)$ around the inward surface normal. For $\mu = 0.5$, that half-angle is about 26.6°. For $\mu = 0.1$ (oily steel on steel), it shrinks to 5.7° — almost a straight line.
This cone is why surface condition dominates grasp reliability. Consider a parallel-jaw gripper clamping a 2 kg aluminum block. The required vertical force to hold it against gravity is $mg \approx 19.6$ N. With two contacts sharing the load, each contact must supply 9.8 N of upward friction. If $\mu = 0.4$ (clean aluminum), each finger needs a normal force of at least $9.8 / 0.4 = 24.5$ N. If the block is oily and $\mu = 0.15$, each finger needs $9.8 / 0.15 = 65.3$ N — a 2.7× increase. A gripper sized for the clean case will drop the oily part.
The friction cone also defines the grasp wrench space. Each contact contributes a wrench (force + moment) to the object. The set of all wrenches a grasp can resist is the positive span of the contact wrenches, each scaled by its friction cone. Computing this span is the core of grasp quality analysis, and it is what the Ferrari-Canny metric quantifies: the radius of the largest ball, centered at the origin, that fits inside the grasp wrench space. A larger radius means the grasp resists disturbances in more directions with more margin.
Worked Example: Two-Finger Pinch Grasp
Take a 50 mm diameter steel cylinder, mass 1.5 kg, grasped by a parallel-jaw gripper with flat rubber pads ($\mu = 0.7$). The gripper applies 40 N per finger. Contact points are diametrically opposed.
The friction cone half-angle is $\arctan(0.7) = 35.0°$. Each contact can supply tangential force up to $0.7 \times 40 = 28$ N. Two contacts give 56 N of total vertical capacity against a 14.7 N weight — a safety factor of 3.8. But this assumes perfect alignment. If the cylinder is tilted 10°, the tangential component needed rises and the effective capacity drops. At a 30° tilt, the required tangential force per contact approaches the cone boundary, and the grasp is on the edge of slipping. This is why tilt tolerance is a real specification in gripper selection, not a nicety.
Mathematical Foundation
The friction cone constraint at contact $i$ with normal $\mathbf{n}_i$ and friction coefficient $\mu_i$ is:
$$ |\mathbf{f}{t,i}| \leq \mu_i , f{n,i}, \quad f_{n,i} \geq 0 $$
where $\mathbf{f}{t,i}$ is the tangential force vector, $f{n,i}$ is the normal force, and $\mathbf{n}_i$ is the unit inward normal. The non-negativity condition enforces that the finger pushes, never pulls.
The contact wrench in the object frame is:
$$ \mathbf{w}_i = \begin{bmatrix} \mathbf{f}_i \ \mathbf{r}_i \times \mathbf{f}_i \end{bmatrix} $$
where $\mathbf{f}i = f{n,i}\mathbf{n}i + \mathbf{f}{t,i}$ is the total contact force, $\mathbf{r}_i$ is the vector from the object’s center of mass to contact point $i$, and $\times$ is the cross product. The second block is the moment the contact force produces about the center of mass.
For a grasp to hold an external wrench $\mathbf{w}_{ext}$ (gravity, acceleration, payload), there must exist contact forces inside their friction cones such that:
$$ \sum_{i=1}^{k} \mathbf{w}i + \mathbf{w}{ext} = 0 $$
This is the equilibrium condition. It is a linear program when the friction cones are linearized (approximated as pyramids with $m$ edges, typically $m = 8$). The grasp wrench space is the convex hull of the contact wrenches scaled by their cone edges.
Worked numeric example. Two contacts on a 100 mm cube, center of mass at origin. Contact 1 at $(0.05, 0, 0)$ with normal $(-1, 0, 0)$; contact 2 at $(-0.05, 0, 0)$ with normal $(1, 0, 0)$. Friction coefficient $\mu = 0.5$. Apply 30 N normal force at each contact, plus tangential forces of $(0, 10, 0)$ N at contact 1 and $(0, -10, 0)$ N at contact 2 (a pure vertical shear pair).
Check cone: $|\mathbf{f}_t| = 10 \leq 0.5 \times 30 = 15$. Valid.
Wrench at contact 1: $\mathbf{f}_1 = (-30, 10, 0)$, moment $\mathbf{r}_1 \times \mathbf{f}_1 = (0.05, 0, 0) \times (-30, 10, 0) = (0, 0, 0.5)$. Wrench $= (-30, 10, 0, 0, 0, 0.5)$.
Wrench at contact 2: $\mathbf{f}_2 = (30, -10, 0)$, moment $= (-0.05, 0, 0) \times (30, -10, 0) = (0, 0, 0.5)$. Wrench $= (30, -10, 0, 0, 0, 0.5)$.
Sum: $(0, 0, 0, 0, 0, 1.0)$. The forces cancel but the moments add to 1.0 N·m about the z-axis. This grasp cannot resist a pure z-axis torque with this force assignment — it would need a different tangential force distribution. That is a concrete demonstration of why force closure analysis (tomorrow’s topic) requires checking all wrench directions, not just the obvious ones.
Practical Example
Scenario: A 6-DOF arm with a two-finger parallel gripper must pick a 250 g plastic bottle (diameter 65 mm, height 200 mm) from a conveyor moving at 0.3 m/s and place it in a case. The bottle surface is smooth HDPE with a measured friction coefficient of 0.25 against the gripper’s polyurethane pads. Gripper max force is 60 N per finger. Cycle time requires a 0.4 s transfer.
Step 1 — Identify the worst-case load. The bottle is accelerated laterally during the pick. Assume 2 m/s² lateral acceleration (conveyor tracking plus arm motion). Lateral force $= 0.25 \times 2 = 0.5$ N. Gravity force $= 0.25 \times 9.81 = 2.45$ N. Combined worst case $\approx 2.5$ N per contact pair, so 1.25 N per contact.
Step 2 — Compute minimum normal force. With $\mu = 0.25$: $f_n \geq 1.25 / 0.25 = 5$ N per finger. That seems trivial against a 60 N gripper.
Step 3 — Add a safety factor and check bottle deformation. Use a 4× safety factor: 20 N per finger. But HDPE bottles buckle. At 20 N per finger on a 65 mm diameter bottle with 0.8 mm wall thickness, the ovalization is small but measurable. At 40 N per finger, the bottle visibly deforms and the contact geometry shifts, which can reduce effective $\mu$ because the contact patch moves. Set the gripper command to 22 N per finger — enough margin, below the deformation threshold.
Step 4 — Verify with the friction cone. Half-angle $= \arctan(0.25) = 14.0°$. The resultant force direction (gravity plus acceleration) is about 11° from the tangential direction. This is close to the cone boundary. A wet bottle (condensation from a cold-fill line) drops $\mu$ to 0.15, half-angle 8.5°, and the grasp fails. The fix is not more force — it is a different pad material or a form-closure feature. This is exactly the kind of analysis that separates a working cell from one that fails every Monday morning when the line is cold.
Common Pitfalls
1. Assuming a single friction coefficient for all surfaces. What goes wrong: the grasp works in the lab on clean parts and fails on the line. Why: $\mu$ varies with surface finish, contamination, humidity, and contact pressure. A steel-on-steel contact can range from 0.1 (oiled) to 0.7 (dry, clean). Detect by measuring $\mu$ on actual production parts under actual conditions, not from textbook tables. Fix by adding contamination robustness (grooved pads, higher safety factor) or by measuring slip with a force/torque sensor.
2. Treating point contacts as sufficient for soft fingertips. What goes wrong: the grasp rotates about the contact normal unexpectedly. Why: rubber pads transmit torsional moment, and a point-contact model ignores it. Detect by comparing simulated and real behavior under twisting loads. Fix by using the soft-finger model (4 DOF per contact) whenever pad compliance exceeds about 1 mm.
3. Ignoring the moment components of the grasp wrench. What goes wrong: the object translates correctly but spins. Why: engineers check force balance and forget $\mathbf{r} \times \mathbf{f}$. Detect by computing the full 6D wrench sum in your grasp analysis. Fix by including moment equilibrium in the grasp quality check — the Ferrari-Canny metric does this automatically.
4. Sizing the gripper for the nominal payload only. What goes wrong: the grasp fails during acceleration or when a second part is added. Why: dynamic loads and combined payloads exceed the static weight. Detect by logging peak forces during real cycles. Fix by using a safety factor of at least 3–4× on the static load, and verify against the friction cone for the worst-case acceleration direction.
5. Assuming friction cones are symmetric in all directions. What goes wrong: the grasp holds against one shear direction but slips in the orthogonal direction. Why: the cone is symmetric about the normal, but the object geometry and contact placement are not. Detect by evaluating grasp quality across the full wrench space, not just along gravity. Fix by running the Ferrari-Canny or a linear-programming grasp check over a sampled set of disturbance directions.
Key Takeaways
-
The contact model is a constraint, not a description. Choosing frictionless, point-with-friction, or soft-finger determines how many force and moment components your grasp can transmit. Pick the model that matches your actual fingertip compliance.
-
The friction cone half-angle $\arctan(\mu)$ governs everything. At $\mu = 0.25$ the cone is only 14° wide. Small changes in surface condition cause large changes in grasp capacity — a drop from 0.4 to 0.15 nearly triples the required normal force.
-
Always check the full 6D wrench, including moments. Force balance alone is insufficient. The $\mathbf{r} \times \mathbf{f}$ term is where many grasps silently fail under rotation.
-
Size for the worst case, then verify with the cone. Use a 3–4× safety factor on static load, then confirm the resultant force direction stays well inside the friction cone for all expected accelerations and surface conditions.
-
Measure $\mu$ on real parts. Textbook friction coefficients are a starting point, not a specification. Production surface conditions, contamination, and temperature change the number that matters.
Tomorrow
Tomorrow we build on the friction cone to answer the central question of grasp analysis: does a set of contacts fully constrain an object against all external wrenches? We will formalize Force Closure and Form Closure, derive the rank conditions, and compute closure with the Ferrari-Canny metric on worked examples.