Paper that folds itself

A sheet that folds itself is an appealing object and a well-funded one. Print a crease pattern, embed something in it that contracts when heated or wetted or exposed to light, apply the stimulus, and the sheet becomes a shape without anybody touching it.

The demonstrations work, and they work less often than they should. The usual diagnosis is that the actuator is too weak. It is usually the wrong diagnosis.

An actuator has two jobs #

Set out what a self-folding crease has to accomplish, and it separates cleanly.

The first job is to supply a fold angle — to move the sheet along the one degree of freedom its vertices have. Something has to bend the sheet along the line, against whatever stiffness the material has, far enough that the fold is established. That is a force problem and it is the one everybody designs for: choose an actuator with enough moment, put enough of it along the crease, apply enough stimulus.

The second job is to determine which configuration the sheet ends up in. A crease pattern with a valid assignment usually has more than one folded state consistent with it, and a sheet that has been bent along all the right lines can still be the wrong object.

The second job has no force attached to it. It is a selection problem, and adding torque does not select.

Where the choice comes from #

The source of the ambiguity is the vertex, and it is visible in the kinematics.

A degree-four vertex is a spherical linkage, and solving its closure requires intersecting two circles. Two circles meet twice, so at every point of the motion there are two configurations — two branches — and the vertex is on one or the other.

The branches are not small perturbations of one another. Take a vertex with sectors of sixty, ninety, a hundred and twenty and ninety degrees. One branch gives the assignment mountain-valley-mountain-mountain; the other gives mountain-valley-valley-valley. Folded flat, those are two different objects, and a model made of many such vertices has a combinatorial number of possible outcomes.

At the flat state the two branches meet. That is the point where the choice is made, and it is made at the moment the sheet leaves the plane.

Both branches are downhill #

Now the part that makes it a hard problem rather than a tuning problem.

Give every crease a spring that wants it fully folded, and compute the stored energy along each branch. The energy is highest at the flat state — every crease is as far from its target as it can be — and falls to zero at the flat-folded state, on both branches.

So both are downhill, all the way, from the same starting point. There is no barrier between them to be overcome and no minimum on one side that is deeper than the other. The energy landscape is a ridge with the flat state on top and the two folded states at the bottom of either side, and the sheet rolls off whichever way it happens to lean.

Which is a genuinely awkward result. A stronger actuator makes the sheet roll faster. It does not make it roll the right way.

The choice is made at second order #

The energy curves say more than both are downhill, and the extra says why the problem is fragile rather than merely unsolved.

The two branches meet at the flat state, and they meet with the same energy — every crease is equally far from its target there, on either branch, because at the flat state there is only one configuration. So the two curves agree in value at the branch point, and they agree in slope too, since the slope is set by the same set of crease angles.

The first place they can differ is the second derivative. Whichever branch curves away faster is the one a quasi-static descent takes, and the quantity deciding the sheet’s shape is therefore a difference of curvatures at a single point.

That is a small quantity by construction. It is not a barrier to be overcome, not a depth to be compared, and not anything an actuator’s strength appears in — which is the precise sense in which more torque cannot help.

It also says what does. A pre-folded crease, a slightly stiffer hinge, a panel biased a degree out of plane: each of those breaks the agreement at first order, which dominates a second-order difference outright. The remedy for a second-order selection is a first-order bias, and it can be arbitrarily small and still decisive — which is why the fix in practice is a nudge rather than a stronger motor.

Which theorem was checked, and how #

Two things are computed here and one is asserted.

The kinematics is solved rather than posed: the vertex’s four crease directions are constructed under the constraint that the sector angles are fixed, and the fourth is solved for by intersecting two cones. The fold angles are then measured off the resulting configuration, which is where the branch structure comes from.

The energy is a sum over the creases of the squared distance from a full fold, evaluated along each branch, and the plot is that function.

The assertion is the one that matters: before drawing, the generator checks that the two branches actually give different mountain-valley assignments. If they came out the same there would be no choice to make and the whole essay would be about nothing, so the generator throws rather than drawing a picture of a distinction that does not exist.

What the figure cannot show is a real actuator. The spring model is the crudest possible: linear, identical at every crease, with no rate, no hysteresis and no coupling to the panel stiffness. A real self-folding sheet has all of those and the energy landscape is correspondingly less tidy — but the ridge is topological, and a more careful model moves the curves without removing it.

What symmetry breaking looks like in practice #

Since the energy does not choose, something else has to, and in every working design something else does.

Pre-creasing. Fold the sheet slightly by hand along the intended lines before actuating. The sheet then starts a little way down one side of the ridge and stays there. This is what almost every laboratory demonstration does and it is why the videos look so convincing.

Asymmetric actuators. Put the contracting layer on one face rather than in the middle. The moment then has a definite sign, and the crease folds the way the layer is on. This is the same trick as putting a hinge on one face of a thick panel, and it has the same cost: the assignment is built into the hardware and cannot be changed afterwards.

Sequencing. Actuate some creases before others, so that by the time the ambiguous ones move the sheet is already committed. Thermally this is done by giving different creases different activation temperatures; it converts a selection problem into a scheduling problem, which is easier.

Stops. Add mechanical features that block one branch physically. Reliable and expensive in complexity.

Every one of these is a way of breaking the symmetry before the sheet reaches the branch point. None of them is more torque.

What actually gets built with it #

The applications are worth naming, because they explain why a problem this awkward is worth solving at all.

Things that have to be assembled where nobody can go. A structure that unfolds inside a body, inside a spacecraft or inside a sealed chamber has no assembler available, and a sheet that folds itself is the only mechanism that fits. Stents are the mature case: a tube small enough to thread through an artery, which then opens.

Things too small to handle. Below about a millimetre, tweezers stop being an option and surface forces dominate. Self-folding at that scale is not a convenience but the only manufacturing route, and a great deal of the microfabrication literature is about folding two-dimensional lithography into three-dimensional devices.

Things that have to be made in quantity. A sheet that folds itself is made flat, which is the cheapest way to make anything, and then becomes three-dimensional without a jig. That is the argument for printed robotics, and it is an economic argument rather than a technical one.

In every case the branch problem is present and in every case it is solved by breaking symmetry in the design rather than by controlling the actuation. That is the practical content of this essay.

Scale changes which forces matter #

The energy argument above is purely geometric, and at small enough scales geometry stops being the whole story.

A crease has a bending stiffness that scales with the cube of the thickness, so a sheet ten times thinner is a thousand times easier to fold. Surface tension, adhesion and electrostatic forces scale much more slowly, and below a certain size they overwhelm the elasticity entirely.

That has a useful consequence and an awkward one. The useful one is that a micron-scale sheet can be folded by capillary action alone — a droplet placed on a flat template pulls it closed as it evaporates, which is a standard technique. The awkward one is that the same forces make the sheet stick to itself in configurations nobody asked for, and the branch selection problem acquires a competitor.

At larger scales the balance goes the other way. A metre-scale deployable has to fight gravity and its own weight, the actuator has to be substantial, and the branch problem is solved by hinges that physically cannot go the wrong way — which is the surface-hinge trick doing double duty.

What the sheet is being told #

There is a way of framing self-folding that makes the difficulty sound less like an obstacle and more like an information problem, and it is the framing worth carrying away.

The crease pattern is a program. Actuating it is running the program. And the program as usually written is under-specified: it says where to fold and how far, and it does not say which of the consistent outcomes is wanted.

An under-specified program does not fail; it produces one of its permitted outputs. So a self-folding sheet that ends up in the wrong configuration has not malfunctioned — it has executed a program that permitted that configuration, and the fix is to write a program that does not.

Pre-creasing, asymmetric actuators, sequencing and stops are all ways of adding the missing information. They differ in where the information lives: in the initial state, in the hardware, in the schedule, or in the geometry. What none of them does is add force, because the missing thing was never force.

That reading also explains why the problem gets worse with complexity rather than better. A pattern with n ambiguous vertices has up to two to the n permitted outputs, and specifying which one is wanted takes n bits — which have to be supplied somewhere, by something, in the physical design.

Where the model stops #

One vertex. A real pattern has many, and their branch choices are not independent — a crease shared between two vertices has one assignment, so choosing at one constrains the other. Whether the constraints leave one global choice or many is a pattern-by-pattern question.

No self-intersection. The kinematics finds configurations satisfying the angles. It says nothing about whether the panels pass through one another on the way, and a branch that requires them to is not available in practice however good it looks in the model.

No layer ordering. The energy argument is about one vertex. Whether the branch chosen at each vertex assembles into a globally consistent stacking is a separate constraint and can rule out combinations that are individually fine.

Rigid panels. Self-folding sheets are usually not rigid. They bend, and the bending is often what makes the fold happen at all.

Quasi-static. The energy argument assumes the sheet moves slowly enough to be at equilibrium throughout. A fast actuator can carry a sheet past a branch point by momentum, which is either a nuisance or a mechanism depending on the design.

A crude spring. Real creases have a rest angle, a stiffness that changes with fold angle, and a memory. None of that is here.

No gravity, no friction, no air. All three matter at the scales these things are built at, and all three are asymmetries that break the tie — sometimes helpfully and sometimes not.

The pattern is assumed valid. Everything here is about which of the permitted outcomes occurs. Whether the pattern has any valid folded state at all is a separate and harder question.

Zero thickness. The vertex here is four rigid sectors of no depth. A real self-folding sheet is a laminate several layers deep, and getting that thickness through the fold changes both the kinematics and the energy.

The surprise: this is a bifurcation, and the field has a name for it #

The shape of the difficulty is not special to folding, and recognising it saves reinventing a great deal.

A system sitting at a symmetric configuration, with two equally good asymmetric states available and nothing to choose between them, is a pitchfork bifurcation. Buckling columns do it, snapping shells do it, and the whole apparatus of stability theory applies.

What that apparatus says, translated, is exactly the practical advice above. A bifurcation is resolved by an imperfection: introduce a small asymmetry and the system follows the perturbed branch reliably, and the size of the perturbation needed falls as the system moves away from the branch point. So pre-creasing works for the same reason that a column with a small initial bow buckles predictably while a perfectly straight one does not.

It also says something less comfortable. Near the branch point the sensitivity to imperfection is unbounded, so the outcome is determined by whatever asymmetry happens to be largest — a thickness variation, a temperature gradient, the way the sheet was resting. A design that does not deliberately introduce an asymmetry is a design whose outcome is set by manufacturing noise.

That is the whole content of “self-folding is unreliable”, stated in the vocabulary of a field that has been thinking about it since Euler.

Who found it, and when #

Self-folding has an unusually short and well-documented history.

The first programmable self-folding sheets are from the mid-2000s, from work at MIT and Harvard on shape-memory composites and on printed actuators. Robert Wood, Daniela Rus, Erik Demaine and colleagues produced the self-folding robot demonstrations that made the field visible around 2010 to 2014.

The branch-selection problem was recognised early by the people building the things and articulated more formally later. Work on the energy landscape of folding vertices — including the observation that the flat state is a degenerate point where branches meet — is from the 2010s and belongs as much to the mechanics community as to the folding one.

The vocabulary of bifurcation arrived last of all, and importing it is the step that turns a list of workarounds into a theory.

Why it is worth the trouble #

The difficulty described here is real and the field has not gone away, which is worth explaining.

The alternative to a self-folding sheet is an assembly process, and an assembly process needs somebody or something to do the assembling. At laboratory scale that is a person; at production scale it is a machine with a jig; at micron scale it is nothing at all, because there is no manipulator small enough and cheap enough.

So the branch problem is a difficulty inside an approach whose competitor is impossibility. A sheet that folds itself into one of two configurations, with a design trick to make it reliably the right one, is a working manufacturing route. A sheet that has to be folded by a machine that does not exist is not.

That is the honest case for the field, and it is stronger than the demonstrations suggest. The videos show sheets folding themselves into shapes a person could have folded in a minute; the applications are all in the regime where the person cannot get there. The same argument justifies every deployable: the fold is not the cheapest way to make the shape, it is the only way to make the shape where it is needed.

The ladder from here #

Later rungs against this anchor: actuator materials, and what each one’s moment-angle curve looks like. Sequenced folding as a scheduling problem. Branch coupling across a pattern, and when the choices are independent. Self-folding at small scales, where surface forces dominate and the whole energy balance changes. Self-assembly proper, where the sheet is not creased in advance. And the reverse problem — designing a pattern whose branch structure has only one outcome — which would remove the difficulty rather than manage it and does not appear to have been attempted.

The torque is the easy half. Deciding what the sheet becomes is the half that has no force behind it.