Gripping: holding an object so that it can be moved

Perception has finished. You know where the object is, roughly how big it is, and which way it is turned. The arm has not moved yet. Between those two facts sits a question with its own hardware, its own arithmetic and its own failure modes: what, physically, is going to hold this thing, and will it still be holding it when the arm stops.

This area is that question, in six documents. This one is the map: the six ways a robot can hold something, what your task actually needs, what a grip has to survive, and the three facts that decide everything downstream — that a rated payload is a static number, that the force you command is not the force the object gets, and that the sensor which says "holding" cannot say "holding the way you planned".

Who this is for, and what it is for#

This is written for someone who has read the object perception area or knows the equivalent, and now has to choose a gripper and a gripping strategy for a real project. You do not need to have used a force sensor. Every term is explained where it first appears.

Three boundaries keep this area from swallowing its neighbours.

Finding the object is not here. Everything about cameras, depth sensors, masks, poses and what they cost in millimetres belongs to object perception. This area starts with a measurement already in hand and asks what to do with it.

Moving the arm is not here either. Planning a path to the grasp pose, avoiding the table on the way in, and carrying the object afterwards belong to arm movement. Where the two meet — a grasp that is geometrically perfect and kinematically unreachable — is called out in section 7 and again in choosing a grip.

What is here is the interval between them. The fingers approach, they close, the object is held, the weight transfers to the wrist, and the grip survives everything the move does to it. That interval is where most pick-and-place failures actually happen, and it is the least written-about part of the pipeline.

Everything below carries the same three commitments the perception area makes. Every technique and every model family has five jobs it suits and five it cannot do, because almost everything in this field works on the demonstration its authors chose. Every licence was read from the project's own licence file, which in this area catches out more people than in any other, for reasons section 1 of licences and platforms sets out. And everything says whether it runs on an Apple Silicon Mac without an NVIDIA graphics card, because a large part of the grasp-model literature quietly assumes one.

Contents#

1. Where gripping starts and where it stops
2. The six ways to hold something
3. What your task actually needs
4. What a grip has to survive
5. Two payloads for one gripper
6. What you know before you close the fingers
7. The three things every tutorial leaves out
8. Knowing how much force was applied
9. The seven documents that follow

1. Where gripping starts and where it stops#

A pick-and-place cycle is usually drawn as a straight line: look, plan, move, close, lift, move, open. The line hides the fact that "close" is not one event and "lift" is not a separate one.

Written out at the level the hardware actually works at, a pick is seven steps.

1. The fingers are opened to a width chosen in advance, because opening them after the arm has arrived costs time and sometimes collides with the table.
2. The arm approaches along a direction chosen so that the fingers pass either side of the object rather than into it.
3. The fingers close, and stop either at a commanded position or on contact.
4. The gripper reports whether it stopped because it found something.
5. The arm lifts, and the wrist takes the weight.
6. The grip is checked against what was expected, because step 4 answers a weaker question than it appears to.
7. The arm moves, and the grip has to survive the acceleration.

Steps 1, 3, 4, 6 and 7 belong to this area. Step 2 and the second half of step 7 belong to arm movement. The measurement that decided the width in step 1 came from object perception.

The reason to draw the boundary here rather than anywhere else is that steps 3 to 7 share one instrument — the force the gripper applies and the force the wrist feels — and one arithmetic, which is the friction arithmetic in choosing a grip. Nothing before step 3 uses either.

2. The six ways to hold something#

There are six mechanisms in industrial use, and they are not variations on each other. Each exploits a different physical principle, and the principle decides what the object has to be like.

Parallel jaw and adaptive fingers hold by squeezing. Two or three fingers close on the object and friction between the pads and the surface resists gravity. This is the default, it works on almost anything rigid, and it needs two roughly opposed surfaces that the fingers can reach.

Suction holds by pressure difference. A cup is sealed against the object, the air inside is removed, and the atmosphere outside pushes the object onto the cup. It needs one reachable patch that is flat enough, smooth enough and airtight enough to seal against. It needs no access to the sides at all, which is why it dominates palletising and warehouse picking.

Magnetic holds by magnetic attraction. It needs the object to be ferromagnetic, which in practice means steel or iron and not aluminium, copper, brass or stainless steel of the austenitic grades. Where it applies it is extremely strong for its size and needs no seal and no fingers.

Soft and compliant grippers hold by wrapping. A pneumatic or elastomeric finger inflates around the object and conforms to whatever shape it meets. The contact pressure is low and spread out, which is why these handle food and fragile goods. They are slow compared with a jaw, and their grip strength is modest.

Multi-finger hands hold by arranging several independently controlled fingers around the object. They can do things no two-finger gripper can — hold a tool and then use it, change the grasp without putting the object down — and they cost an order of magnitude more, need far more control software, and are still rare outside research.

Custom tooling holds by being shaped like the thing. A hook, a mandrel that goes inside a bore, a shaped nest, a clamp that fits one part and nothing else. This is the cheapest and most reliable option when the part mix is fixed, and it is what a large fraction of working automation actually uses. It is also the one nobody writes tutorials about.

The table below is those six against the property of the object that decides whether each one can be used at all. Read the last column as the disqualifying condition: if it is true of your object, that row is out, whatever else is attractive about it.

3. What your task actually needs#

People choose a gripper by reading gripper reviews, which sorts them by how impressive the gripper is. Real cells choose the other way round: the task dictates a mechanism, usually within about a minute, and the interesting question is what that mechanism then costs you elsewhere.

The table is real task types against what they genuinely use. Read the last column carefully — it is where the effort usually goes that need not have.

Three things follow from that table, and each of them saves work.

Most working cells do not plan a grasp. They repeat one. When the part is fixtured, or the same carton arrives every time, the grip was decided by a person at commissioning and the robot only executes it. Grasp planning is what you reach for when fixturing has lost the argument, in exactly the way perception is what you reach for when the same has happened to the part's position.

The gripper choice decides the perception requirement, not the other way round. A suction cup needs to know where one flat patch is and its surface normal, which a depth image gives directly and cheaply. Two fingers need to know the object's width, its orientation and where its neighbours are, which is a much harder perception problem. Choosing suction can delete a week of perception work, and choosing fingers can create one. This is worth deciding in that order.

Changing the fingertip is cheaper than changing anything else. Almost every two-finger gripper takes custom fingertips, and a shaped fingertip turns a hard grasp problem into an easy one — a V-groove holds a cylinder without needing to know its diameter, a nest holds one part in one orientation. Before buying a better gripper or training a grasp model, work out what a machined piece of aluminium would do.

One last thing worth saying, because it cuts against the instinct. Task difficulty does not track gripper sophistication. The genuinely hard tasks — seating a connector, folding cloth, balancing a stone — are hard in force control and contact reasoning, and several of them are done with the plainest possible two-finger gripper. A task that needs a multi-finger hand is usually one where the object has to be re-oriented while held, which is a narrow and specific requirement rather than a general measure of difficulty.

3.1 When a grasp model makes things worse#

The instinct, when the object set is not fixed, is to reach for a grasp network. The perception area makes the general form of this argument in its own section 2.1, and gripping has a sharper version of it, because what a grasp network predicts is narrower than what people assume.

A grasp network predicts one thing: will the object slip out. It is trained on simulated or recorded attempts labelled by whether the object stayed in the hand. That is one of your constraints. It is almost never the only one. It has no way to know that a wine glass must be held by the stem, that a grip above half the object's height cannot be inverted afterwards, that the handle is where the fingers must not land, or that this part has to arrive at the next station the same way up every time. Those are task constraints and there is nowhere to type them in.

A rule can be argued with and a score cannot. When a grip fails, a rule tells you which sentence was wrong and you change the sentence. A network gives you a number that was high for a grip that did not work, and the only available response is more training data.

A grasp model is a candidate generator, and that is a real job. Used to propose fifty poses that something else then filters against reachability, task constraints and the gripper's own geometry, it earns its place immediately. Used as the decider, it silently drops every constraint you did not train it on. Models that grasp is written around that distinction.

4. What a grip has to survive#

A grip is usually judged by whether it lifts the object. That is the easiest of the four things it has to do, and the one the other three are measured against.

Gravity, while standing still. The friction at the pads has to exceed the weight. This is the calculation every datasheet quotes and it is the least demanding case.

Acceleration, while moving. The arm accelerates, and the object's inertia adds to the load on the pads. Robotiq's own manual for the 2F-85 works this through and the result is worth reproducing exactly, because it is more brutal than people expect. Their formula for the weight a friction grasp can hold is

W = (2 x F x Cf) / Sf

where F is the force the gripper applies, Cf is the coefficient of friction between pad and object, and Sf is a safety factor the integrator chooses. With 200 N of grip, a measured coefficient of 0.3 for silicone against lubricated steel, and a safety factor of 2.4, that gives 50 N — about 5 kg standing still. Their next sentence is the one that matters: at an acceleration of 2 g, that same 5 kg object exerts 98 N of inertial force on its own, and it is dropped. The rated payload was never a payload; it was a static holding capacity, and every move you make spends some of it.

The placement. Putting the object down loads the grip in a direction the pick never did. The object stops, the arm does not, and the resulting force is transmitted through the fingers. This is where objects get pushed out of the grasp, cracked, or wedged.

Disturbance. Something brushes the object, or it catches on the lip of a fixture, or the next item in a tote shifts against it. A grip that survives the first three and not the fourth looks completely reliable until the day the cell gets busy.

A useful consequence: the acceleration limit is often a better thing to tune than the grip force. Increasing the squeeze risks damaging the object, and increasing it is bounded by what the object can take. Slowing the move down costs cycle time and nothing else. If a grip fails only during transport, the first experiment is a slower move, not a harder squeeze.

5. Two payloads for one gripper#

The single most useful idea in this whole area is that the same gripper holding the same object has two different capacities, and which one applies is decided by how the fingers ended up around the object rather than by anything in the datasheet's headline.

Force fit, also called a friction grasp, holds the object only by squeezing it. Nothing stops it sliding out except friction. Form fit, also called a form-closure or encompassing grasp, has the fingers wrapped around or under the object so that sliding out would require the object to pass through the finger. Friction becomes a secondary concern.

OnRobot publish both numbers for the same product, which makes the point without any interpretation. On the RG2 the force-fit payload is 2 kg and the form-fit payload is 5 kg. On the 2FG7 they are 7 kg and 11 kg. Same motor, same fingers, same object — two and a half times the capacity, or half as much again, purely from the geometry of how the fingers arrived.

This is the argument for two things that otherwise look like fussiness. It is why choosing a grip spends a whole section on form closure rather than treating friction as the whole story, and it is why custom fingertips are usually the cheapest upgrade available: a fingertip with a V-groove or a lip converts a force fit into a form fit for the cost of a machined part.

It has a trap attached, and it is the best single example of the register this area is written in. Robotiq's 2-finger grippers are underactuated: one motor drives both fingers, and each finger folds at a knuckle. Whether the gripper produces a parallel fingertip grip or curls round into an encompassing one is not commanded. It is decided by where the object first touches the finger. Robotiq's manual names the boundary the equilibrium line and says plainly that grasping an object on that line is not recommended, because small variations in position will switch the grasp from parallel to encompassing and back. A cell that grips near the line gets 5 kg of capacity most of the time and 2 kg of it occasionally, with nothing in the robot program different between the two.

6. What you know before you close the fingers#

The best predictor of which approach you need is not how awkward the object looks. It is how much you know about it in advance, in the same way it is for perception — but the questions are different ones, because a gripper cares about different properties than a camera does.

The rows below go from knowing the most to knowing the least, and the work goes up as you go down.

Most projects that get into trouble here have reached for the fifth row when they were in the third. The fifth row is what papers are written about, so it is what people read about first.

7. The three things every tutorial leaves out#

Each of these is a case where the naive implementation appears to work, and keeps appearing to work until a specific condition arrives. They are the reason this area exists as more than a hardware catalogue, and each is covered properly later.

The force you command is not the force the object gets, and the difference depends on the object. Robotiq publish measured forces for the 2F-85 against payloads of six different hardnesses. Gripping a steel part it reaches 220 N. Gripping a 10 A durometer neoprene part — soft rubber — the same gripper at the same setting tops out at 115 N, barely half. The force setting on an underactuated gripper is a motor torque request, and how much of it arrives at the object is decided by how much the object gives way. Any threshold you calibrate on a metal part is wrong for a soft one, in the direction that makes a soft object look as though it was never gripped. Grippers and hardware, section 2 has the whole table.

"Object detected" is a weaker statement than it sounds. A two-finger gripper reports a successful grasp when the fingers stopped before their commanded position. That is true of a good grip, and it is equally true of a grip on the wrong part of the object, a grip on two objects at once, a grip on the edge of a fixture, and a finger that jammed. The gripper knows its fingers stopped. It does not know what stopped them. The check that distinguishes these is the wrist taking the weight, which is a different sensor answering a different question — the same pattern the perception area describes for the guarded move.

A grasp that is geometrically perfect can be kinematically impossible. A grasp pose is a full six-degree-of-freedom pose for the gripper, and the arm has to reach it with the whole gripper body clear of the table, the tote wall and the object's neighbours. A grasp search that scores candidates on quality alone and hands the best one to the planner will, for objects lying near a wall, propose excellent grasps the arm cannot make — and it will do so silently, because nothing in the quality score knows about the arm. The fix is to bound the search by the gripper's own body before scoring, which choosing a grip, section 7 works through. Putting the constraint at the end instead of the start is the single most common structural mistake in a grasp pipeline.

8. Knowing how much force was applied#

A question worth answering early, because the answer shapes what you can build.

ROS 2 defines the interface cleanly. control_msgs/GripperCommand carries a max_effort in its goal and an effort in its result, and the message's own comment specifies the units: "The current effort exerted (in Newtons)". So there is a standard, unambiguous way to ask for a force and read one back.

Whether anything fills that field in is a separate question, and for the most common gripper in the world the answer is no. The official Robotiq ROS 2 driver exports exactly two state interfaces, position and velocity; there is no effort interface, so nothing downstream can report the force being applied, and the maximum force is a static parameter read from the URDF rather than something you set per grasp. That is a reasonable decision — the hardware reports motor current, and converting that to fingertip force depends on which grasp mode the linkage settled into — but it means a great deal of published example code reads a number that was never measured.

The five routes that do work, and the one worth reaching for first, are set out in the two-finger gripper document. The short version is that you usually do not need the grip force itself. You need to know whether the object is held, and the wrist force sensor answers that directly.

9. The seven documents that follow#

If you are starting a project rather than reading through, the order that wastes least time is: this document, then grippers and hardware to fix the mechanism, then choosing a grip. Reach for models that grasp when the first three have run out, which for a table-top arm with a describable object family is later than most people expect.