Text from the interactive explanation
Bistable origami robot for adaptive multimodal motion actuated by nonuniform magnetic fields
Chong Huang, Zhuangzhi Miao, Sida Ma, Zheguo Tang, Heming Wang, Zhanpeng Cui, Yihao Fu, Yang Li
Original paperOne 0.81-gram origami robot, two stable shapes, eight moves — driven by a deliberately uneven magnetic field.
Could a capsule smaller than a fingernail crawl, roll, grab and let go — steered by magnets that never have to surround you?
Replay ×3 real time
reconstruction
Animated reconstruction: shape from the paper's crease pattern; motion timed and scaled to the measured strides, speeds and pulse timings in Figs. 2–4. Field shading is a simplified sketch, not the paper's field maps.
1 Starting fold
2 Demonstrated magnetic input
3 Switch shape
Touring all eight moves until you pick one.
1
Why not just use a better magnet?
A magnetic field is the invisible push-and-pull around a magnet or a wire carrying current. A tiny magnet sitting in a field can feel two different things. It can be twisted until it lines up with the field — that turning effect is a torque . And it can be pulled bodily towards where the field is stronger — a force .
The authors model each of the robot's four magnets as a pair of opposite "magnetic charges", a north end and a south end that the field pushes in opposite directions. Drag the magnet below and compare the two kinds of field.
Net pull: 0 — the even field only twists the magnet
Drag the magnet (or focus the picture and use the arrow keys). Arrows: blue on the north charge, red on the south charge, copper = net pull. illustration, arbitrary units
In an even field the pushes on the two ends are equal and opposite. They cancel, so the magnet only twists. Only where the field changes from place to place — a gradient — do the two ends feel different pushes, leaving a net pull. In the authors' analysis (Table S12), snapping shut or open, tumbling and scampering all need that pull.
Most magnetic micro-robot systems are built from Helmholtz coils : pairs of rings around a box that make a very even, precisely controlled field — but only in a small region enclosed by the coils. This team did the opposite. They mounted three coils and one small electromagnet under an open table, on two perpendicular sliders that follow the robot (Fig. 1G). Their field is uneven on purpose. The robot's rolling workspace was a circle ≈120 mm across, and the sliders are only accurate to ±2 mm. The precision burden moves into the robot's folds.
Prediction game: which actions need a pull?
Tap every action you think needs a net force (not just a twist), then reveal the authors' analysis. analytical prediction, Table S12 — not an experiment
Pick the actions you think need a pull.
Fixed Helmholtz coils: 4 of 12 actions predicted possible
2
Two shapes, one hill
Bistable means "having two stable shapes". A light switch is the classic example: it is happy up or down, resists in between, and then suddenly flips. The robot's fold is tracked by one angle, θ : 0 when it lies flat, about 1.5 rad (≈85°) when it has curled shut.
Why is there a hill at all? The panels are thick plastic and the hinge lines are deliberately offset. The pieces only fit together at the start and at the end of the fold. In between they collide, so the thin silicone hinges must stretch and bend — storing elastic energy, like a bent ruler.
Zero-thickness crease pattern (Table S1) folded to θ · mountain · valley kinematic model
Elastic energy vs θ model, ≈digitized Fig. 1C
Predict, then release: if you let go at this angle, which shape wins?
Set an angle, make your prediction, then let go.
Release, wiggle and kick are toy dynamics on the model curve; kick strength is in arbitrary units.
Measured snap-through forces measured
A force probe slowly pushed one panel until the robot snapped (Fig. 1F; three prototypes, five times each).
Values ≈digitized. Smaller a₂ offsets made a taller hill. After 205 close–open cycles the selected design was still bistable (Fig. S14).
Self-locking measured
Once shut, the robot stays shut with no power at all. Gripping a small sphere in its locked state it resisted a drag of 26.5 mN before letting go, and it carried up to ≈2.7 g — more than three times its own 0.81 g — when holding a 6 mm sphere (Fig. 3J).
That is why the switching pulse only has to be brief: the geometry, not the coils, holds the shape.
3
How a wobble becomes a walk
A small oscillating field cannot flip the robot over its hill. It just flexes it inside the open valley: the field folds it a little, then the hinges spring it back. On its own that is only a wobble. The authors made the robot's two ends different on purpose: the field pulls harder on the rear end, and the two feet grip the floor differently. "The combination of asymmetric force and friction converts the small-amplitude folding and unfolding motions into a forward movement."
Net step per wobble: 0 mm — it rocks in place
Travelled 0.0 mm
Rule of this toy: while the field folds the robot, the foot whose pull beats its grip slides; while the hinges spring back, the less grippy foot slides. Full step scaled to the measured ≈0.8 mm per cycle (5 A, 1 Hz). Arch and step drawn larger for visibility. toy model
Faster wobbles, faster walking measured
Crawling speed rose with both the current amplitude and the frequency, reaching ≈2.75 mm/s at 6 A and 4 Hz (Fig. 2F, mean of five runs, ≈digitized).
Turning is the same wobble with the pull angled across the body (≈3.3° per cycle at 4 A, 3 Hz). Tumbling is different: a large, short pulse flips the whole flat robot end over end.
4
Eight moves, measured
Each move belongs to one shape and one kind of signal (Fig. 1I). Small oscillating fields jiggle the robot inside its current valley; large short pulses throw it as a rigid body — or kick it over the hill into the other shape. The clips are the authors' own footage; success rates are from 50 tries each on one prototype (Table S10).
Open flat crawler
A −2 A, 0.1 s pulse snaps it shut in ≈0.06 s (Fig. 3B); tracked points reach several hundred mm/s (Fig. 3G). Lighter pulses grasp (−0.8 A) or capture (−0.3 A).
50/50 closings · silkworm capture 4/100 tries (a proof of concept)
Clip: Supplementary Movie S5 — authors' labels ×1 / ×5, further sped up 2×.
≈0.8 mm per cycle at 5 A, 1 Hz (Fig. 2E); up to ≈2.75 mm/s (Fig. 2F). Turning: ≈3.3° per cycle at 4 A, 3 Hz (Fig. 2K), up to ≈28.5°/s at 6 A, 4 Hz (Fig. 2L).
crawling 50/50 · turning 46/50
Clip: Supplementary Movie S4 — filmed ×25 (on-screen label), further sped up 1.5× (≈×37 overall).
+5 A pulses when lying face down, −3 A when face up, 0.1 s each; half a turn in ≈0.10 s (Fig. 2N). The coils' symmetric field lets it tumble in any direction.
42/50
Clip: Supplementary Movie S4 — first part ×0.1 slow motion as labelled; second part labelled ×10, further sped up 4× (≈×40).
Closed capsule
A +2 A, 0.1 s pulse snaps it open in ≈0.06 s (Fig. 3B). The clip shows both switches: −2 A to close, +2 A to open.
opening 50/50
Clip: Supplementary Movie S5 — authors' labels ×1 and ×0.1 slow motion; no further speed change.
A 0.23 A, 3 Hz half-wave signal on the single electromagnet: ≈1.5 mm per stride, ≈4.5 mm/s (Fig. 4E–F); up to ≈9.45 mm/s at 0.31 A, 4 Hz.
48/50
Clip: Supplementary Movie S6 — filmed ×2 (label), further sped up 3× (≈×6 overall).
Identical 5 A, 1 Hz sines on coils 1–3; each half field cycle rolls it half a circumference (Fig. 4N). Top speed 5.45 body lengths per second with a 2 Hz field.
50/50
Clip: Supplementary Movie S6 — ×1 real time as labelled.
A −3 A, 0.25 s pulse launches it (Movie S6). In a sweep of −0.6 to −1.2 A, hops reached ≈68 mm high and ≈96 mm far (Fig. 4K–L).
44/50
Clip: Supplementary Movie S6 — ×0.2 slow motion as labelled; no further speed change.
All clips: Huang et al., Nat. Commun. (2026), Supplementary Movies S4–S7, CC BY 4.0; trimmed, resized, some sped up as noted, re-encoded, audio removed. On-screen labels and speed markers are the authors' own.
Jump planner
Where should the robot sit relative to the electromagnet, and how hard should the pulse be? Pick a spot and a pulse to see what the team measured.
Measured: ≈35 mm high, ≈34 mm far
Closest to the axis the pull points mostly upward: tall hops, but short ones.
Height and distance: measured means of five jumps, ≈digitized from Fig. 4K–L. The arc between them is drawn as a simple parabola. measured endpoints
5
The obstacle course
To show why switching shape matters, the team built tracks where no single move would do (Fig. 5). Too wide for a channel? Close up and roll. Too tall to roll under a beam? Open up and crawl. A low wall? Tumble over it.
Schematic inspired by Fig. 5A, not to scale illustration
Step 1 of 9 · Crawl ② towards a narrow channel (open)
Track 2
On a second course the robot gripped a cargo stuck to the floor and freed it when the coil platform shifted slightly, creating a shearing force; jumped a high wall while still gripping; and switched to scampering to pass a low tunnel inside a water tank, before rolling to the target and opening to release (Fig. 5B).
How reliable? measured, Table S10
Single moves mostly succeeded 84–100 % of the time; some changes between moves only 28–32 %. An uninterrupted run of the whole track worked 4 of 56 times (≈7.1 %). The authors note that much of the switching time went into an operator visually monitoring and adjusting the coils' position, and that tracking delays and positioning errors can interrupt a run.
6
Back to the question
The robot is sub-centimetre, and when closed it resembles a capsule. With its structural stability and room for small sensors, the authors call it "a promising candidate for the development of functional endoscopes". They are equally clear about what has to come first:
clinically relevant, biocompatible materials;
tests in realistic conditions — so far it moved on dry tracks and in simple water, while the gut is full of viscous fluids;
a more precise gantry, better coil geometry and control, a larger workspace, portability and heat management;
closed-loop navigation guided by medical imaging such as X-ray or ultrasound;
validation in ex vivo gastrointestinal tissue, then animal studies.
The bigger idea travels further than this one robot: put the "intelligence" into the geometry, and the magnet system can afford to be simple, large and open.
Source & credits
Huang, C., Miao, Z., Ma, S., Tang, Z., Wang, H., Cui, Z., Fu, Y. & Li, Y. Bistable origami robot for adaptive multimodal motion actuated by nonuniform magnetic fields. Nature Communications (2026), Article in Press. https://doi.org/10.1038/s41467-026-79010-w
The article is published under a Creative Commons Attribution 4.0 International (CC BY 4.0) licence.
Video: clips adapted from Supplementary Movies S4–S7 of Huang et al. 2026, CC BY 4.0; trimmed, cropped, speed-changed, re-encoded, audio removed. Speed changes are listed under each clip.
Plots replotted from approximate digitization of Figs. 1C, 1F, 2F and 4K–L (values marked ≈). Success rates from Table S10; the Helmholtz comparison from the authors' analytical prediction in Table S12. The 3-D robot, the field sketch, the crawling toy model, the energy-curve dynamics and the track schematic are independent reconstructions for explanation, not the authors' simulations.
Tags: measured experiments reported in the paper · model the authors' theory or analysis · reconstruction our own illustration.
Topics in this iPaper
Engineering
a large-workspace magnetic actuation system is developed to generate low-uniform magnetic
Abstract, page 1
Physics
simultaneously magnetic gradient forces (F) to govern the rigid-body motions
Results, 'Design procedure of the robotic system', page 4
miniature robots
Untethered miniature robots are promising tools for accessing narrow
Introduction, page 1
magnetic actuation
magnetic driving offers several advantages
Introduction, page 1
bistability
bistability is introduced into the thick-panel origami by selecting special geometric parameters
Introduction, page 2
thick panel origami
Another way to resolve this is to employ thick-panel origami
Introduction, page 2
origami robots
origami-based designs excel in providing transformation from 2D sheets to various 3D shapes
Introduction, page 2