Article

Solid-body trajectoids shaped to roll along

desired pathways

https://doi.org/10.1038/s41586-023-06306-y Yaroslav I. Sobolev1,5 ✉, Ruoyu Dong1,5, Tsvi Tlusty1,2 ✉, Jean-Pierre Eckmann3 ✉,

Steve Granick1,4 ✉ & Bartosz A. Grzybowski1,4 ✉
Received: 2 January 2022

Accepted: 9 June 2023

In everyday life, rolling motion is typically associated with cylindrical (for example,
Published online: 9 August 2023
car wheels) or spherical (for example, billiard balls) bodies tracing linear paths.
Check for updates
However, mathematicians have, for decades, been interested in more exotically
shaped solids such as the famous oloids1, sphericons2, polycons3, platonicons4 and
two-circle rollers5 that roll downhill in curvilinear paths (in contrast to cylinders
or spheres) yet indefinitely (in contrast to cones, Supplementary Video 1). The
trajectories traced by such bodies have been studied in detail6–9, and can be useful in
the context of efficient mixing10,11 and robotics, for example, in magnetically actuated,
millimetre-sized sphericon-shaped robots12,13, or larger sphericon- and oloid-shaped
robots translocating by shifting their centre of mass14,15. However, the rolling paths
of these shapes are all sinusoid-like and their diversity ends there. Accordingly, we
were intrigued whether a more general problem is solvable: given an infinite periodic
trajectory, find the shape that would trace this trajectory when rolling down a slope.
Here, we develop an algorithm to design such bodies—which we call ‘trajectoids’—and
then validate these designs experimentally by three-dimensionally printing the
computed shapes and tracking their rolling paths, including those that close onto
themselves such that the body’s centre of mass moves intermittently uphill
(Supplementary Video 2). Our study is motivated largely by fundamental curiosity,
but the existence of trajectoids for most paths has unexpected implications for
quantum and classical optics, as the dynamics of qubits, spins and light polarization
can be exactly mapped to trajectoids and their paths16.

Let us begin by drawing some curve T on a plane, copying it multiple of the path T. When the starting point of the next segment is reached,
times to form an infinite periodic trajectory T∞ , and inclining the plane we trim the shell to define another local cylinder to roll along this new
slightly (Fig. 1a). The question we then ask is whether it is possible to direction. The procedure is then repeated until reaching the end of the
design a solid body—which we will call a ‘trajectoid’—to roll exactly polygonal path (or, in general, of a curvilinear path approximated by a
along T∞ without slipping or pivoting (in which ‘pivoting’ means spin- union of infinitesimally short linear segments). The body thus shaped
ning about a normal to the plane at a point of contact). can be construed as a piecewise combination of gradually reorienting
To develop intuition about engineering such objects, we begin with cylinders whose axes remain parallel to the plane below, and pass
the simplest case: a cylinder of radius r rolling without slipping or through the body’s CM. This CM always remains at a height r above the
pivoting over a horizontal plane is a trajectoid of a straight line (Fig. 1b). plane, and the planar path defines an equipotential ‘trench’ in the tra-
For the preferred, linear path, the cylinder’s centre of mass (CM) is jectoid’s landscape of gravitational energy (Fig. 1b). Whereas the sphere
always at a height r above the plane (Fig. 1b, green dot): in this sense, touches the plane at a single point at a time, the final shape touches it
the path is an ‘equipotential’ one, and any deviation from it would raise at many points at once (as with cylinder-to-plane contact), giving stabil-
the cylinder’s CM (to the red dot), thereby increasing the gravitational ity to the roll. Some examples illustrating the morphing procedure for
potential energy (green surface). different trajectories and for different values of R /r are shown in Fig. 1d,e.
Extending to a flat path T composed of several straight lines (Fig. 1c),
we consider a ‘virtual’ assembly of a heavy sphere (radius r) encased
inside a weightless spherical ‘ghost’ shell of radius R > r. We shave off Condition for trajectoid existence
part of the shell to leave behind a locally cylindrical region of radius r: For an object to qualify as a trajectoid, it must periodically regain its
in this way, the object can roll along the first segment, with the points initial three-dimensional orientation. Noting that the points of the
of attachment of the inner, heavy sphere tracing precisely this portion object’s contact with the planar trajectory T also trace a certain curve
1
Center for Soft and Living Matter, Institute for Basic Science (IBS), Ulsan, South Korea. 2Department of Physics, Ulsan National Institute of Science and Technology (UNIST), Ulsan, South Korea.
3
Department of Theoretical Physics and Mathematics Section, University of Geneva, Geneva, Switzerland. 4Department of Chemistry, Ulsan National Institute of Science and Technology
(UNIST), Ulsan, South Korea. 5These authors contributed equally: Yaroslav I. Sobolev, Ruoyu Dong. ✉e-mail: [email protected]; [email protected]; [email protected];
[email protected]; [email protected]

310 | Nature | Vol 620 | 10 August 2023

a c a Area Area
Incli A 2πr2 2πr2
plan ned Shell R
e
T
Translated A
copy of T n(t)
A n(t)

r n(t) n(t)
b Potential CM Ω
surface Spherical
Core curve AΩ
Ω

CM Remnant Path Tn as
Polygonal flat metal strip Ω
chain path
Cylinder b c
deflects
from path

CM in
ground Shaved off part
state
of the shell

Preferred,
linear path

d R/r = 1.01 R/r = 1.05 R/r = 1.30 R/r = 2.80

d e

e R/r = 1.01 R/r = 1.05 R/r = 1.30 R/r > 2.80

Fig. 2 | Path-to-sphere mapping and trajectoid existence. a, Condition

Fig. 1 | Rolling paths and trajectoid shapes. a, An arbitrary curve T (dark necessary for trajectoid existence. A flat path Tn = AΩ (starting at point A and
blue) drawn on an inclined plane (grid) serves as one period of an infinite ending at point Ω) can be construed as a sticky metal strip that easily bends out
translationally symmetric trajectory T∞ to be traced by the trajectoid. of the plane yet maintains its length and geodesic (initially, in-plane) curvature.
b, A cylinder (gold) rolling along the plane (blue grid) must tilt sideways, as This metal strip attaches (maps) to the sphere being rolled along it and
shown, to deviate from its preferred linear path of rolling (thick grey line on the becomes the sequence of the sphere-plane contact points. For a trajectoid of
grid). Such a tilt inevitably raises the cylinder’s CM (red dot) along a cycloid-like AΩ to exist, on reaching the point Ω, the sphere must arrive at the same 3D
⌢
curve (solid black) away from CM’s original level (green dot): the preferred orientation as it had at the starting point A. This condition is equivalent to AΩ
linear path corresponds to a trench in the gravitational potential landscape being a closed spherical curve splitting the sphere’s surface into two equal
(green surface, also see Extended Data Fig. 6). For an illustration in frontal areas 2πr2 (shaded green and yellow on the rightmost sphere in a). b–d, When a
projection, see Extended Data Fig. 1a. c, The design of a trajectoid shape starts path T (b, one period) does not map onto a closed curve on a sphere it can be
with a rigid core (blue) of radius r surrounded by a massless concentric shell either appended with a bridge (c, coloured in red and orange here) or deformed
(gold colour) of radius R > r . This composite structure, shown here in central (d, green lines connect the points of original black T to respective points of its
cross-section, is forcibly moved in infinitely small steps such that the spherical deformed blue version, see main text and Methods for details). Some T can be
core rolls along the target path (grey polygonal chain) without slipping or transformed in this way to yield viable one-period trajectoids (experiments in
pivoting (Supplementary Video 1). To allow such rolling, the shell is ‘shaved’ to Fig. 4d,e,g,n,o). e, By contrast, a two-period trajectoid exists for most paths
leave a patch of cylindrical surface (orange). The point CM is the centre of mass and no amendments to the path are needed (the green circle shows where the
of the core, and therefore of the whole assembly. d, Trajectoids for a linear path two periods meet).
using different ratios R /r consist of a cylindrical part (orange) of radius r and
two surviving spherical segments of the shell (gold). e, Trajectoids for the
curvilinear path (same as in Fig. 2a and experimentally realized in Fig. 4d) using The path Tn = AΩ can be represented by the dependence of its normal
different ratios R /r . At R /r = 1.01, the body is still roughly spherical but for n(t) = (cosψ(t), sinψ(t), 0) (blue in Fig. 2a) on the arc length t along
higher values shown, the shape becomes complex (and does not change for
the trajectory, where ψ is the angle n(t) forms with the planar projec-
R /r ≳ 2.80) and has improved stability due to deeper potential energy trench
tion of gravity. Alternatively, Tn can be specified by the curvature
(b and Extended Data Fig. 1a).
κ(t) = dψ /dt and the initial angle ψ(0) (ref. 17). The translational perio-
dicity of the path Tn (as in Fig. 1a) indicates that the curvature obeys
an index theorem18,
on the surface of this object’s inner, heavy sphere (Figs. 1c and 2a), the
condition for periodic motion can be recast in terms of the rotation ψ(t) = ∫ κ(t)dt = 2πIT, (1)
T
group acting on this rolling sphere. Specifically, an n-period trajectoid T

exists for a translationally invariant path T∞ if there exists a ball of finite

radius r, which, after rolling (without slipping or pivoting) along n where IT is the rotation index of the period T, for example, IT = 0 if T
periods T composing the subpath Tn = AΩ, regains the same orientation does not self-intersect. A complementary scenario was recently con-
at endpoint Ω as it initially had at starting point A. Repeating this cycle, sidered19 where bodies are designed such that their rolling paths are
the trajectoid then rolls indefinitely along T∞. confined to a finite region on the plane. These ‘non-trajectoids’

Nature | Vol 620 | 10 August 2023 | 311

Article
therefore violate the property in equation (1) and are bound to halt a b M c
when rolling autonomously down a slope. M T
Let us now model the rolling motion in a frame of reference in which U
the sphere’s centre is stationary and the sphere rotates around it. We M
approximate the planar path AΩ = n (t) by a polygon made of N short U S(r) = πr2
U S(r)
straight segments of length δt . In the said coordinate system, each
T T
segment δt with starting point ti induces a rotation of the sphere around A A
S(r) A
an axis parallel to the normal to the path n (ti) (blue arrows in Fig. 2a)
by an angle δt /r (Methods). The rotation is captured by a rotation matrix d T Straight e
Semicircle
Ri = exp[(δt /r )(n(ti) ⋅ L)] , where the generators of the rotation group Small
M
f
A M Ω T
SO(3) are L = (Lx , Ly , Lz ) . As the sphere is rolling, the rotations Ri act on arcs α π−α
e
the sphere sequentially, progressing from the first to the last segment

Normal spherical
f 0 10 20 30 40 50 60 70 O
πr2 d C
(Fig. 1c), and the overall rotation RAΩ is then their ‘time-ordered pro­

area S(r)/r2
π c πrr2
b
duct’, RAΩ = RN ⋯R2R1. In the limit N → ∞, the overall rotation RAΩ becomes a U
0 π−α α
an infinite product of minute rotations, which takes the form of a –π/2 T180
0 1 2 3 4 5
A=Ω
time-ordered exponential, familiar from quantum field theory20, Scale V = L/(2πr)

 1 Ω  Fig. 3 | Property π and TPT theorem. Illustration of property π ensuring the

RAΩ = T exp ∫ A dt n(t) ⋅ L  (2) existence of TPTs (see text for details) for a path whose period (green or blue
 r  in d) consists of a semicircle, two straight lines and two small arcs to smooth
⌢
the corners. a–c, Spherical trace T (green) of the contact point on rolling a unit
where the time-ordering operator T ensures that the rotations and the sphere along one period of the path. We use control parameter σ = L /(2πr )—the
generators are multiplied in the correct order according to their posi- ratio of the path’s length L and the sphere’s circumference (for example, σ = 1
tion t along the path. In terms of rotation matrices, the trajectoid must for a great circle). a, σ = 0.304. b, σ = 0.526. c, σ = 0.81. As σ increases from a to c,
⌢
obey RAΩ = 1, where 1 is the identity matrix. the area S (σ ) (shaded green) enclosed by the green spherical trace T and
⌢
Per the illustration in Fig. 2a, this condition can be visualized by the red arc U also increases (plotted inf, orange curve and bottom axis) and
imagining that the flat curve Tn = AΩ is cut out of a flexible metal sheet eventually reaches π in d (also shown by a left red dot in f). For an animation, see
as a narrow ribbon that can bend out of plane and stick onto the rolling Supplementary Video 3. Points marked on the orange curve S(σ ) in f correspond
⌢ to a–c. The purple curve in f is a respective plot for a random walk path (Extended
ball, producing a curve AΩ on the sphere. As the sticking ribbon
deforms only out-of-plane, it conserves the arc length t (by forbidding Data Fig. 4g), its trajectoid solution marked by the rightmost red circle. Note
stretching due to slipping) and the geodesic curvature κ(t) (by forbid- that this function is non-monotonic (text). e, Illustration of the proof of the TPT
theorem. See text and Supplementary Video 3 for details.
ding in-plane bending due to pivoting). Thus, κ(t) would now become
⌢
the geodesic curvature of the ‘attached’ spherical curve AΩ.
⌢
The RAΩ = 1 condition implies that AΩ is a closed curve on the
⌢
sphere. By the Gauss–Bonnet theorem18,19,21, AΩ encircles an area,
2
S = 2πr (1 − I AΩ), where the rotation index is I AΩ = (2π)−1  dt κ(t). The
⌢ ⌢ Two-period trajectoid theorem
invariance of κ (t) , as illustrated by the sticky ribbon, indicates that Two-period trajectoids (TPTs) can be constructed for paths T∞ whose
its integral, the rotation index, is also conserved, I ⌢
AΩ = IAΩ = nI T. Hence, period T has the following property π: As a sphere of finite radius r > 0
⌢
by equation (1), the area enclosed by AΩ , on either side (green and rolls along the period T = AM, the contact point traces on the sphere’s
⌢
yellow in the rightmost panel of Fig. 2a) is S = 2πr 2(1 − IAΩ) = 2πr 2, for surface a spherical curve T , which is not necessarily closed (green in
⌢ ⌢ ⌢
IT = 0. If AΩ self-intersects, then the standard definitions of signed Fig. 3a–c). Together, T and a great arc U (red) connecting its ends M
oriented area are in effect18 and, with this convention, the area must and A enclose some oriented spherical area S(r ). The condition requires
be a multiple of 2πr2. that there exists an r for which S(r )/r 2 = ±π (red dots in Fig. 3f).
The proof relies on the following simple observation: by the
Gauss–Bonnet theorem, the integral of geodesic curvature along a
One- and multiple-period trajectoids spherical curve with property π is, modulo 2π,  dt κ(t ) = 2π − S(r )/r 2 = π
It is obviously not guaranteed that an arbitrary one-period path T will (details in Methods). However, the curvature integral vanishes along
⌢ ⌢
map onto the rolling sphere as a closed loop, let alone enclose an area T (by periodicity, equation (1)), and along the great arc U (because it
of exactly 2πr2. In fact, these conditions are typically not met and is geodesic), so the only remaining contribution comes from the corners
one-period trajectoids (n = 1) are expected to be rare. This said, a given A and M whose angles therefore add up to π (π − α and α in Fig. 3e).
⌢
path T can sometimes be heuristically adjusted to produce a similar This property allows us to draw a 180°-rotated image T180 (Fig. 3e, blue)
⌢
trajectory that does close up on a sphere and encloses 2πr2 areas. One of the original curve T (Fig. 3e, green), such that the two curves con-
⌢
way to do so is by appending a specially constructed bridge to the given nect smoothly at A = Ω and M to form a closed two-period curve, T2 ,
2
path (Fig. 2b,c, see details in Extended Data Fig. 3 and Methods). enclosing an area S(r ) = 2πr , thus obeying condition RAΩ = 1. This
Another is to deform the given path as illustrated in Fig. 2d and Extended proves the TPT theorem. We make three remarks about this result.
Data Fig. 2. In both approaches, the quantity to minimize is the angle (1) Having property π is not obvious, because S(r )/r 2 may often be a
θ of mismatch between the initial and final orientations of the sphere non-monotonic function of r (Fig. 3f and Extended Data Fig. 4e,h
on completing one path period T (for details, see Methods). As we will and figure captions). Also, property π is sufficient but not necessary
see later, these procedures generally produce trajectoids that, when for the existence of a TPT: for instance, a degenerate loop in which
physically fabricated, trace the adjusted paths faithfully and periodi- a flat curve is traversed forward and then in reverse, automatically
cally (Fig. 4). yields a period-one trajectoid, and therefore a TPT as well. Alterna-
In striking contrast to n = 1, we prove below that two- and higher-period tive forms of property π are discussed in the Methods.
trajectoids, n ≥ 2, exist for a broad class of paths defined by a general (2) The TPT theorem proof implies, in particular, that splitting the
⌢
property π described below. In our numerical experiments, a random trajectoid through the plane of the great arc U produces two iden-
graph drawn on the plane always belonged to this class. This includes tical halves: we illustrate this twofold rotational symmetry in Sup-
wildly self-intersecting and convoluted graphs (Extended Data Fig. 4). plementary Video 2 (first experiment) with a trajectoid built from

312 | Nature | Vol 620 | 10 August 2023

 a d e f g h i j k l m n o

CM
dge
Bri

b
Light

Downhill direction
Heavy

CM

Stop
c Stop

dge
Bri

Theory
Experiment

Fig. 4 | Experimental validation. a,b, Methods of CM engineering: introducing single-period, n = 1 trajectoids of strictly downhill paths. f,i, Examples of paths
strategically placed voids (a) or using a heavy sphere concentric with the desired that have no single-period trajectoid; the body shaped to match this path
location of the CM (b). c, Photograph of two three-dimensionally printed halves stops after one period at the point marked red. g, However, the single-period
of one shape, each containing a hemispherical void for housing a steel ball trajectoid exists for a path similar to f but with a bridge (orange) appended
bearing (1-inch diameter). d–o, Examples of specific paths and corresponding to it. h, Alternatively, the input path from f can be traced by a two-period, n = 2
bodies. Solid black lines are intended trajectories of the CM. Blue curves are trajectoid. Correspondence to f and g is emphasized by brown dashed lines.
experimental trajectories obtained by image-processing of top-view Along path h, the trajectoid gains enough kinetic energy to cut some corners
experimental videos (Methods and Supplementary Video 2). Filled black (for example, purple arrows) and occasionally escape the main potential
(or black and red in f) markers denote locations corresponding to the body’s trench illustrated in Extended Data Fig. 7c. j,m, Further n = 2 trajectoids. The
single revolution. The smallest period is between any two adjacent black circles trajectoid in j is manufactured from two identical halves (magenta and green)
(empty or filled). Green circles highlight sharp corners at which tumbling or corresponding to two periods (symmetry axis OC in Fig. 3e). n,o, The n = 1
recoiling was observed. Yellow arrows point to slight zig-zag wiggling of trajectoids that roll, respectively, intermittently upwards and in loops. Scale
the experimental trajectory around the intended path. d,e, Examples of bars are 1 cm. Each experiment was replicated at least 5 times.

such halves (Fig. 4j). Generalizing this construction to n-fold sym- For example, voids can be strategically introduced into the design of
metry, we prove that multi-period trajectoids (MPT) with any n > 1 a three-dimensionally printed object22,23 (Fig. 4a), including voids that
exist for paths having property 2π/n (MPT theorem in the reach the surface: these would not affect the rolling as long as they do
Methods). not influence the shape’s convex hull. Instead, we simply placed a heavy
(3) Property π is obeyed by a rich variety of unusual paths, including 1-inch (2.54 cm) ball bearing (steel, ρ = 7.8 g cm−3) into a spherical
self-intersecting, spiralling and two-dimensional random walk cavity centred on CM (Fig. 4b) inside a trajectoid made of a light,
paths shown in Extended Data Fig. 4. On the other hand, there ρ = 1.25 g cm−3, polylactic acid (PLA). As shown in the image in Fig. 4c,
exist special paths that violate property π. A simple example is a sym- we three-dimensionally printed the shape in two halves, each one hav-
metric, V-shaped path with an acute angle between the equal arms ing a hemispherical cavity for the steel ball, and glued them together
(Extended Data Fig. 5a–c and Supplementary Video 4) or some more after the ball was inserted. The shape of the PLA shell was determined
complex V-like paths with perfect symmetry of the arms (Extended by the shaving procedure described earlier (Fig. 1) and implemented
Data Fig. 5d and further details in Methods). Because of this extra numerically as described in the Methods. Readers can do it for a path
symmetry, they are much rarer than paths having property π. Also of their own by using the online Google Colab notebook released with
note that any slightest tapering of the acute angle of V-shaped path this paper. We used the R /r value of 1.3 to ensure adequate depth of
already gives it property π (Extended Data Fig. 5g). Although we the potential energy groove (Fig. 1 and Extended Data Fig. 1a) and
cannot prove that these are the only paths that violate property better stability along the path. Overall, this fabrication scheme ensured
π, we conjecture that paths without property π are infinitely rare, that our trajectoids were heavy despite the low density of the PLA part,
because they never appeared among random paths in our numeri- which is experimentally convenient as it makes gravity dominant rela-
cal experiments. tive to stray forces of other nature. Furthermore, it makes rotational
inertia more isotropic, with moments of inertia dominated by those
of the heavy steel ball.
Experimental validation The fabricated trajectoids were released to roll down a flat plane
Turning to the experimental validation of the algorithm, we recall that inclined at angles 0.6–1.7°, with the instantaneous locations of the
the location of the trajectoid’s CM was assumed to be insensitive to the centre of gravity estimated and tracked (Methods). Figs. 4d,e,n,o and
final shape itself, and so the CM of the manufactured object must be Extended Data Fig. 6a,b show the trajectories traced by one-period
at this predefined point CM. This can be achieved in several ways. (n = 1) trajectoids. As seen, there is generally close correspondence

Nature | Vol 620 | 10 August 2023 | 313

Article
between the intended and experimental results, save for some wiggling crystal electro-optic effects for LIDAR (laser imaging, detection and
at sharp corners that is due to non-zero inertia of the physical bodies ranging) and holographic displays35–37.
(if the trajectoid loses too little energy to friction and inelastic colli- Further examination of the basic geometric problem of trajectoids
sions, its total energy might eventually become sufficient for escap- may lead to unexpected analogies and applications in other physical
ing the potential trench that follows the target path; see the last two settings, for example, when the trajectoids in time-varying driving
periods in Fig. 4h and Supplementary Videos 1 and 2). Next, Fig. 4f,i field (for example, magnetic12,13) could be used to achieve complex
shows bodies for which the single-period path T mapped onto the motions with fewer moving parts and inside sealed spaces. Moreover,
sphere does not close up: as predicted, the rolling for such a body is trajectoids morphing on-the-fly by means of piezoelectric actuators
not periodic and it stops after one period (red dot). However, a trajec- for solids38 or gas inflation of soft hollow shapes39 could be useful
toid can be constructed when T is either corrected by adding a bridge in robotics.
(orange colour in Fig. 4g) or when, in accordance to TPT theorem, two
periods T are used (n = 2 trajectoids in Fig. 4h,j). More examples of
n = 2 trajectoids are shown in Figs. 4j–m. Apart from being a source Online content
of local wiggling, inertia can be harnessed to follow trajectories that Any methods, additional references, Nature Portfolio reporting summa-
intermittently roll upwards (Fig. 4n) or even in loops (Fig. 4o). Although ries, source data, extended data, supplementary information, acknowl-
the energy of a downward-rolling object decreases owing to friction, edgements, peer review information; details of author contributions
its kinetic component (inertia) may be sufficient to climb along the and competing interests; and statements of data and code availability
potential groove yet insufficient to escape it (Extended Data Fig. 7). are available at https://doi.org/10.1038/s41586-023-06306-y.
In practice, such trajectoids performed reliably only for sufficiently
smooth paths, because recoils and oscillations near sharp turns 1. Dirnböck, H. & Stachel, H. The development of the oloid. J. Geom. Graph. 1, 105–118
render the energy loss stochastic and thereby the progression along (1997).
the periods unpredictable. 2. Hirsch, D. A device for generating a meander motion. Patent no. 59720 (1980).
3. Hirsch, D. & Seaton, K. A. The polycons: the sphericon (or tetracon) has found its family.
J. Math. Arts 14, 345–359 (2020).
4. Seaton, K. & Hirsch, D. Platonicons: the platonic solids start rolling. In Proc. Bridges 2020:
Extensions and outlook Mathematics, Art, Music, Architecture, Education, Culture (eds Yackel, C. et al.) 371–374
(Tessellations Publishing, 2020).
The TPT theorem is a universal property of rotations that can be 5. Stewart, A. T. Two-circle roller. Am. J. Phys. 34, 166–167 (1966).
translated from trajectoids to other physical systems governed by 6. Phillips, A. Meander mazes on polysphericons. In The Visual Mind II (ed. Emmer, M.) 667–684
the rotation group SO(3). In quantum-mechanical systems, the (MIT Press, 2005).
7. Muramatsu, T. The construction of a rideable geometric object based on a conical form:
overall rotation RAΩ (2) can be seen as a time-evolution operator, regarding ‘space walk on the earth’ with ellipsoidal rolling plane. J. Geom. Graph. 15,
Ω
UAΩ = T {exp[−(i /ħ)∫ A dt H (t)]} , with a time-dependent Hamiltonian 203–212 (2011).
8. Kuleshov, A. S., Hubbard, M., Peterson, D. L. & Gede, G. Motion of the oloid on the horizontal
H (t)  n(t) ⋅ L and a time-varying field n(t). Notable examples are a plane. Russ. J. Nonlinear Dyn. 7, 825–835 (2011).
magnetic dipole in a planar magnetic field24 and the Bloch sphere 9. Bäsel, U. & Dirnböck, H. The extended oloid and its contacting quadrics. J. Geom. Graph.
representation of a qubit driven by an electric field25–27 (Extended Data 19, 161–177 (2015).
10. Schatz, P. The Study of Rhythms and Technology: The Evertible Cube. Polysomatic Form-
Fig. 9a and Methods). By ‘almost any’ rotation sequences we hence-
Finding (Niggli, 2013).
forth mean those having property π. The TPT theorem suggests that 11. OLOID Solution GmbH. OLOID—Technology. OLOID https://www.oloid.de/index.php/en/
almost any planar field pulse n(t), once scaled by an appropriate fac- procedures (2021).
12. Jeon, S. Self-positioning ability of a sphericon-shaped magnetic millirobot rolling on an
tor r > 0 and applied twice in a row, will return the system exactly to
inclined surface. AIP Adv. 8, 056729 (2018).
its original state, UAΩ = 1. Such 360°-rotation pulses play a role in com- 13. Jeon, S. A sphericon-shaped magnetic millirobot rolling on a surface actuated by an
mon pulse sequences, as found in rotary echo28,29 and Wimperis external wobbling magnetic field. AIP Adv. 7, 056708 (2017).
14. Rhodes, T., Gotberg, C. & Vikas, V. Compact shape morphing tensegrity robots capable of
sequences30, and the TPT theorem may inspire a new sequence design locomotion. Front. Robot. AI https://doi.org/10.3389/frobt.2019.00111 (2019).
for nuclear magnetic resonance (NMR)16, atom interferometry31 or 15. Haller, J. I. & Samija, L. The moving OLOID–a different pet in different times. Autodesk
verification of quantum simulators32. The scaling factor r can be inter- Instructables https://www.instructables.com/The-Moving-OLOID-a-Different-Pet-in-
Different-Time/ (2020).
preted as modulating the magnitude of the pulse, n(t) → n(t)/r , or as 16. Merrill, J. T. & Brown, K. R. Progress in compensating pulse sequences for quantum
stretching it in time, n(t) → n(tr ) (Extended Data Fig. 9b,c and Sup- computation. In Quantum Information and Computation for Chemistry (ed. Kais, S.)
plementary Video 5). In this analogy, very small trajectoids, r → 0, 241–294 (John Wiley & Sons, Ltd, 2014).
17. Chirikjian, G. S. Stochastic Models, Information Theory, and Lie Groups Vol. 1, 155–159
represent the adiabatic limit, where the field n(tr ) changes very slowly (Birkhäuser, 2009).
relative to the revolutions of the state vector: to such a tiny rolling 18. Kobayashi, S. Differential Geometry of Curves and Surfaces (Springer, 2021).
trajectoid, smooth paths seem straight and therefore trace great cir- 19. Segerman, H. Rolling acrobatic apparatus. Not. Am. Math. Soc. 68, 1 (2021).
20. Dyson, F. J. The S matrix in quantum electrodynamics. Phys. Rev. 75, 1736–1755 (1949).
cles on the core ball (Fig. 2), such that the trajectoid becomes a cylin- 21. Levi, M. Geometric phases in the motion of rigid bodies. Arch. Ration. Mech. Anal. 122,
der (shown in Supplementary Video 8, Extended Data Fig. 10 213–229 (1993).
and Methods). 22. Prévost, R., Whiting, E., Lefebvre, S. & Sorkine-Hornung, O. Make it stand: balancing shapes
for 3D fabrication. ACM Trans. Graph. 32, 1–10 (2013).
In classical optics, the evolving state vector can be taken as polariza- 23. Bächer, M., Whiting, E., Bickel, B. & Sorkine-Hornung, O. Spin-it: optimizing moment of
tion vector of light rotating on the Poincaré sphere (Extended Data inertia for spinnable objects. ACM Trans. Graph. 33, 96 (2014).
Fig. 9d). When light passes through a thin uniaxial waveplate, its polar- 24. Rojo, A. G. & Bloch, A. M. The rolling sphere, the quantum spin, and a simple view of the
Landau–Zener problem. Am. J. Phys. 78, 1014–1022 (2010).
ization vector rotates by an angle proportional to the plate’s thickness 25. Feynman, R. P., Vernon, F. L. & Hellwarth, R. W. Geometrical representation of the Schrödinger
δt , around a planar axis n = (cosψ , sinψ , 0) (refs. 33,34), where ψ /2 is equation for solving maser problems. J. Appl. Phys. 28, 49–52 (1957).
the angle of the waveplate’s optical axis (Extended Data Fig. 9e and 26. Shore, B. Manipulating Quantum Structures using Laser Pulses 146–150 (Cambridge Univ.
Press, 2011).
Methods). Here, the TPT theorem implies that for almost any given 27. Cohen-Tannoudji, C., Dupont-Roc, J. & Grynberg, G. Atom—Photon Interactions: Basic
sequence of waveplates—with arbitrary thicknesses δti and optical axis Process and Appilcations 353–363 (Wiley-VCH, 1998).
directions ψi /2—there exists a scaling factor r such that any light enter- 28. Butts, D. L., Kinast, J. M., Timmons, B. P. & Stoner, R. E. Light pulse atom interferometry at
short interrogation times. J. Opt. Soc. Am. B 28, 416 (2011).
ing the doubled sequence of waveplates with thicknesses δti /r regains 29. Rakreungdet, W. et al. Accurate microwave control and real-time diagnostics of neutral-
its original polarization when it exits (Extended Data Fig. 9e–g and atom qubits. Phys. Rev. A: At. Mol. Opt. Phys. 79, 022316 (2009).
Supplementary Video 6). This basic property may become another 30. Wimperis, S. Broadband, narrowband, and passband composite pulses for use in advanced
NMR experiments. J. Magn. Reson. A 109, 221–231 (1994).
tool in an optical designer’s toolbox, perhaps an aid in creating 31. Dunning, A. et al. Composite pulses for interferometry in a thermal cold atom cloud.
polarization-insensitive (‘phase-only’) optical devices based on liquid Phys. Rev. A 90, 033608 (2014).

314 | Nature | Vol 620 | 10 August 2023

32. Shaffer, R., Megidish, E., Broz, J., Chen, W.-T. & Häffner, H. Practical verification protocols 39. Walker, J. et al. Soft robotics: a review of recent developments of pneumatic soft
for analog quantum simulators. npj Quantum Inf. 7, 46 (2021). actuators. Actuators https://doi.org/10.3390/act9010003 (2020).
33. Chekhova, M. & Banzer, P. Polarization of Light: in Classical, Quantum, and Nonlinear Optics
(Walter de Gruyter GmbH & Co KG, 2021). Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
34. Svirko, Y. P. & Zheludev, N. I. Polarization of Light in Nonlinear Optics (Wiley, 1998). published maps and institutional affiliations.
35. Liu, J. & Wang, J. Demonstration of polarization-insensitive spatial light modulation using
a single polarization-sensitive spatial light modulator. Sci. Rep. 5, 9959 (2015).
36. Li, S.-Q. et al. Phase-only transmissive spatial light modulator based on tunable dielectric Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this
metasurface. Science 364, 1087–1090 (2019). article under a publishing agreement with the author(s) or other rightsholder(s); author
37. Xiong, J., Hsiang, E.-L., He, Z., Zhan, T. & Wu, S.-T. Augmented reality and virtual reality self-archiving of the accepted manuscript version of this article is solely governed by the
displays: emerging technologies and future perspectives. Light: Sci. Appl. 10, 216 (2021). terms of such publishing agreement and applicable law.
38. Wu, Y. et al. Insect-scale fast moving and ultrarobust soft robot. Sci. Robot. 4, eaax1594
(2019). © The Author(s), under exclusive licence to Springer Nature Limited 2023

Nature | Vol 620 | 10 August 2023 | 315

Article
Methods projection never exceeds 8.7% of the trajectoid’s minimal radius r = 1
for all possible orientations of the trajectoid (Extended Data Fig. 8d,f).
Numerical calculation of the trajectoid shape Note 1: the location of the CM can also be estimated more precisely
Once we approximate the planar path AΩ by a polygon made of short by applying the standard algorithms of six-degree-of-freedom (6D)
line segments of length δt , as described in the main text, the matrix pose (position and orientation) tracking to the experimental video:
Rk←0 for rotation from the start of the path (t0 = 0) to the position tk is these algorithms yield global location and orientation of the solid’s own
a product: reference frame, in which location of the true CM is known a priori. To
test this method, we applied markers to the surface of a manufactured
k
trajectoid (black dots in Extended Data Fig. 8e), filmed its rolling at
Rk ←0 = ∏ Ri = Rk ⋯R2R1 (3)
i =1 120 frames per second and applied 6D pose estimation algorithms of
the Blender software41 (Supplementary Video 6). Data presented in
The removal of a shell fragment (shaving) is performed by applying Extended Data Fig. 8h,i evidenced that the 6D pose approach produces
Boolean mesh operations to subtract a series of cubes Ck = (Rk←0)−1C only a negligible improvement of precision as compared to the shape
from the initially geodesic mesh of the shell that envelopes (and is centroid method described above.
concentric with) the core sphere. The inverse rotations slide the cube Note 2: the symmetry of the two halves of the TPT is illustrated at
along the spherical path, instead of rotating the core with respect to the beginning of Supplementary Video 2 with a two-colour trajectoid.
C. Meshes of the cubes Ck were produced with Python by rotations of However, only one-colour trajectoids were used for analysis of experi-
a cube C whose position and orientation are shown in Extended Data mental trajectories.
Fig. 1b, then the Ck were imported into Autodesk 3ds Max 2018 (ref. 40)
for executing standard Boolean mesh operations, although any other Path integral form of the RAΩ = 1 condition
3D CAD software would also be suitable for the purpose. Readers can For an infinitesimal path segment dt, the rotation matrix can be written
supply their intended paths to the online Google Colab notebook in terms of the rotation group generators: R(n(t), dt /r ) = e(dt /r )(n ⋅L) =
released with this paper (Code availability statement at the end of the 1 + (dt /r )(n ⋅ L). Hence, the matrix of net accumulated rotation after
paper) to calculate trajectoid shapes. rolling along the entire period AΩ is a product integral,
Ω Ω Ω
Fabrication  dt  dt  dt 
RAΩ = ∏ R n(t ),  = ∏ e r (n(t )⋅ L) = ∏ 1 + (n(t ) ⋅ L),
r  t =A r
After the trajectoid’s shape was calculated numerically as described t =A  t =A  
above and using parameters r = 15.88 and R = 20.64 mm, three-
dimensional (3D) printing of calculated shape was performed with where each further rotation matrix is multiplied from the left-hand
PLA on Ultimaker 3 Extended printer (Ultimaker B.V.) with triangular side. Equivalently, one can write a path-ordered exponential,
infill at 50% density (generated by Ultimaker Cura slicer). To ensure that
the CM of the manufactured object was always at the predefined point  1 Ω 
RAΩ = T exp ∫ A dt n(t) ⋅ L  (4)
(CM in Fig. 1c), a heavy steel ball (diameter 2.54 cm or 1 inch = 1.6r) was  r 
placed into a spherical cavity centred on CM (Extended Data Fig. 1d and
Fig. 4c). As shown in Fig. 4c, the trajectoid was three-dimensionally where the time-ordering operator T makes sure that the rotations and
printed in two halves, each having a hemispherical cavity for the steel the generators are multiplied in the correct order according to their
ball. After the ball was inserted, the halves were glued together with positions along the path. This compact form links the trajectoids to
cyanoacrylate (Loctite) glue. This approach yielded a heavy object the basic concept of propagators and path integrals of quantum field
despite the low density of the 3D-printing material (PLA) and ensured theory20. In this language, the trajectoid existence condition means
that gravity dominated stray forces of different natures. Furthermore, that the time-evolution operator in equation (2) is equal to the unity
it made rotational inertia more isotropic, with moments of inertia rotation RAΩ = 1.
dominated by those of the heavy steel ball.
Further considerations for the proof of the TPT theorem
Analysis of experimental trajectories Here, we give the proof for a simple path T = AM, which, in the plane,
Silicon carbide abrasive paper (electro-coated, 2000Cw) was glued has the same tangent direction at the beginning and the end, such that
onto a sloped 12-mm-thick glass slab to prevent slippage and pivo­ the integral IT of the curvature (1) vanishes modulo 2π. When the
ting during rolling. Slope angle and direction were measured by high- curve is mapped onto the sphere, the beginning and end points are
precision, two-axis digital inclinometer (Dong-Do IM-2DT). Videos were map­ped to two points A and M and we draw the geodesic arc from M
acquired at either 60 or 120 frames per second using a digital camera to A (the shorter of the two). By property π discussed in the main text,
(D850, Nikon) placed 1.62 m above the slope. The centroid of the object’s there exists a radius r of the sphere, for which the area enclosed by the
apparent projection on the top-view videos was found by thresholding curve and the arc is exactly πr 2. By the Gauss–Bonnet theorem, the
the red, green, blue values of video frames to detect the projection out- (interior) angles at A and M add up to 2π − (enclosed surface)/r 2 = π .
line. Before finding the centroid, a convex hull operation was applied to Because the tangents at A and M in the plane are the same, the exit angle
the detected projection to eliminate any possible non-convexity: as the at M is equal to the entry angle at A, as illustrated in Fig. 3e. It follows
object’s shape is convex by construction, its projection must be convex. that on rolling over the second period, the graph on the sphere MΩ is
To evaluate whether the centroid of the visible shape can be used as the same as the first, but now turned by 180°. Therefore, the area
an adequate estimate of the object’s CM location, we simulated shape enclosed by the second part is again πr 2 and the sum is 2πr 2, which is
projections of the trajectoid onto the plane as it rolled along the target half the area of the sphere. This means that the RAΩ = 1 condition is
path. We then compared the trajectory of the simulated projection satisfied, and therefore the trajectoid will be in exactly the same
centroid (orange solid lines in Extended Data Fig. 8a–c) to the target orientation as when it started at A, once it has rolled over two copies of
path, that is, to the theoretical trajectory of projection of the CM onto the original path. The path can be continued indefinitely. For the more
the plane (black solid lines in Extended Data Fig. 8a–c). The difference general cases, the argument is the same, taking into account that the
between these two trajectories was found to be negligible. Further- tangents at A and M might be unequal, and that the areas might overlap,
more, for trajectoids of paths as in Extended Data Fig. 9a,c, the planar in which case the orientation of the curvature will come into play. For
distance between the projection of the CM and the centroid of the shape an animated illustration of this proof, see Supplementary Video 3.
 Extended Data Fig. 1d), each rotated by an angle 2πi/n around OΛ with
⌢
Paths without property π respect to the original T , such that all n paths connect smoothly to
⌢
Let us consider a path consisting of just two straight segments whose form a closed n-period path, Tn, enclosing an area S(r) = 2πr 2, thus obey-
normals are n1 and n2, such that the sphere undergoes just two rota- ing condition RAΩ = 1, which proves the MPT theorem. The correspond-
tions: around axes n1 and n2 by angles ασ and βσ where α and β are ing trajectoid will show n-fold rotational symmetry and can therefore
arbitrary positive fixed values and σ > 0 is the scaling factor. The rota- be assembled from n identical parts. The case n = 2 corresponds to the
tion angle of the composition of the two rotations is equal to π if and TPT theorem and property π, where the wedge becomes a single great
only if 42 arc. For n = 1, the arc vanishes and T must be closed with property 2π,
which is simply the Gauss–Bonnet form of condition RAΩ = 1.
cot(ασ /2)cot(βσ /2) = n1 ⋅ n2 (5) In the context of ref. 19, the property 2π/n says that the holonomy
(change in the orientation and position of the trajectoid with respect
If α ≠ β, a solution σ > 0 of equation (5) always exists, because in this to the plane) is a pure translation after having covered exactly n copies
case the right-hand side of equation (5) is limited to an interval [−1, 1], of the original path.
but the left-hand side spans (−∞, ∞). If α = β, the left-hand side of equ­ Because of the limits limr →∞ S(r )/r 2 = 0 and limn→∞ 2π/n = 0, one may
ation (5) is positive—spans (0, ∞)—for all σ > 0 and equation (5) can be tempted to prove that for any given T, there exists a large enough
only be satisfied for n1 ⋅ n2 ≥ 0 and has no solutions if n1 ⋅ n2 < 0. n such that an n-period trajectoid exists for T. The obstacle to
Supplementary Video 4 and Extended Data Fig. 5 explore paths that such attempts is the discontinuity of S(r ) for n > 2 . Note that the
⌢
are similar to acute-angle isosceles V paths. We found that an above mentioned MΛA can be constructed only when the arc length
⌢
acute-angle symmetric V-like path still does not have property π even MA is not larger than 2πr/n. This can be seen from the⌢ spherical ⌢
after we introduce a kink into the arms of V-path at points K1 and K2 as rule⌢of cosines applied to the spherical triangle MΛA, cos MA AΛ
= cos 2 r +
AΛ 2π 2π ⌢ r
shown in Extended Data Fig. 5d: note that S(σ) plot in Extended Data sin 2 r cos n ≥ cos n and therefore we find MA ≤ 2πr/n , where equality
⌢ ⌢
Fig. 5e never reaches π or −π. Sweep of the kink angle α is shown in applies only when MΛ = AΛ = πr /2 . As a consequence, S (r ) is not a
⌢
Supplementary Video 4. At the same time, introducing a slightest taper continuous function of r: it is undefined when MA > 2πr/n. In practice,
at the corner of the V-path, as shown in Extended Data Fig. 5g, produces this means that increasing n typically reduces the domain of S (r ) . It is
⌢
a path that has property π: solution of S(σ) = π is marked by the red dot also noteworthy that S(r ) is multivalued: whenever MΛA can be con-
in Extended Data Fig. 5h. Furthermore, deviation from the perfect sym- structed, it can be constructed in two different ways that are mirror
metry of the V-like path’s arms (Extended Data Fig. 5k) also produces reflections of each other in the plane AMO. This ambiguity allows S (r )
a path that has property π: the function S(σ) reaches −π (red dot in to have two values for each r in its domain.
Extended Data Fig. 5l). Asymmetry in the last example is controlled by Following the example of the TPT, we find that counterexamples to
a parameter Ξ responsible for the difference between kink angles α property 2π/n are also single isosceles V paths, with corners γ sharper
and β in two arms: β = α(1 + Ξ) as shown in Extended Data Fig. 5k. The than π/n (generalizing the TPT case, γ < π/2 where n = 2). To show that
sweep of asymmetry Ξ is shown in Supplementary Video 4. these are counterexamples, note that for V to have property 2π/n, roll-
Last, we discuss a version of how one could define rare paths. We have ing along one V must give a matrix of rotation by an angle 2π/n around
shown that certain piecewise linear functions, with some extra sym- some axis. Then, using the formula for the composition of two rotations
metry, define paths for which neither one nor TPTs exist. Clearly, the with α = β to obtain the rotation angle 2π/n requires:
space of such functions has finite dimension, whereas the functions of
differentiability class C 1 considered above form an infinite dimensional cos(π/n) = cos 2(ασ /2) − sin2 (ασ /2)(n1 ⋅ n2)
space. However, there might be other curves for which no TPT exists, = cos 2(ασ /2) + sin2 (ασ /2)cosγ
and we leave this as an open question. Apart from examples such as those ≥ cos 2(ασ /2)cosγ + sin2 (ασ /2)cosγ = cosγ
in Extended Data Fig. 5a,d, numerical experimentation yielded no other
curves without TPTs, and so we conjecture that there may be none. However, for any γ < π/n , we find a contradicting inequality,
cos(π/n) < cosγ . It follows that such a V-shaped path with γ < π/n
Generalization of the TPT theorem to MPT violates property 2π/n.
For any n ≥ 2, we prove the following theorem:
Bridging and scaling paths to support one-period trajectoids
MPT theorem Paths T typically do not satisfy condition RAΩ = 1 for one-period (n = 1)
The n-period trajectoids can be constructed for paths T∞ whose period trajectoids and, consequently, the corresponding one-period trajec-
T has the following property 2π/n: as a sphere of finite radius r > 0 toids are rare. Still, one can find a path similar to the given one that
rolls along the period T = AM, the contact point traces on the sphere’s allows for a one-period trajectoid. We developed two methods to
⌢
sur­face a curve T (which is generally not closed; green in Extended achieve this: either by appending a specially constructed ‘bridge’ to
⌢
Data Fig. 1c,d). T can then be closed by an isosceles V-shaped wedge the given path, or by deforming the given path itself. In both approaches,
⌢
made of two great arcs MΛA (red in Extended Data Fig. 1c,d), with an the quantity to minimize is the angle θ of mismatch between the initial
⌢
angle 2π/n at Λ (orange angle in Extended Data Fig. 1c). Together, T and final orientations of the sphere on completing one path period T.
⌢
and the wedge MΛA connecting its ends enclose some oriented spher- This angle is computed from Euler’s axis-angle representation of the
ical area S(r ) (shaded green in Extended Data Fig. 1c,d). The condition overall rotation matrix
requires that there exists an r for which S(r )/r 2 = ± 2π/n.
The proof again uses the Gauss–Bonnet theorem: the area having  tr RAΩ − 1 
θ = arccos  , (6)
property 2π/n is S(r )/r 2 = 2π/n = 2π −  dt κ (t ). However, the curvature’s  2 
⌢
integral vanishes along T (by periodicity, equation (1)), and along the
⌢
two big arcs of MΛA (as these are geodesic), so the only remaining con- where tr is the trace operator.
tribution comes from the corners with their angles, α at A, λ = 2π/n at Deforming the input path was performed as follows. Keeping a con-
Λ and µ at M: 2π/n = 2π − [(π − α ) + (π − λ) + (π − µ)] = 2π/n + (α + µ − π). It stant r = 1, we applied non-uniform scaling to T with two scaling coef-
follows that, for any n ≥ 2, the two side corners of the wedge add up ficients, kx along the downward direction and ky perpendicular to it.
⌢
to 180°, α + µ = π. This allows us to draw n copies of T , i = 1, …, n (blue For an example path from Fig. 4d (same as path in Fig. 1a and Fig. 2a),
and yellow in Extended Data Fig. 1c, also magenta and sky-blue in the map of angle θ(kx , ky) is shown in Extended Data Fig. 2a. Our goal
Article
is then to find scaling factors (kx , ky) that give θ closest to zero, as long or by rescaling time ( B(t) → B(r ) ). By the TPT theorem, almost
as ky is not too small, because there is always a degenerate solution any given B(t) can be scaled so that applying the scaled magnetic
with ky = 0, which turns the curve into a straight line whose trajectoid field twice, sequentially, brings any magnetic dipole back to the
is a cylinder, but this solution is of no interest. For the example in orientation it had before the experiment, whatever that
Extended Data Fig. 2, we required ky > 0.5 . In Extended Data Fig. 2a, orientation was.
there are two locations with small θ. For the one in Extended Data Fig. 2b, (2) In the NMR context, a strong static magnetic field B0 along the z
⌢
the corresponding curve T traced on the surface of a rolling sphere axis causes precession of the dipole moment around that axis
does not self-intersect; by contrast, for the minimum in Extended Data with Larmor frequency ω0 = γB0 , and a pulse of magnetic field
⌢
Fig. 2c, T has many self-intersections. B1(t) = B1(t)(xˆcos(ω0t − ψ(t)) + yˆsin(ω0t − ψ(t))) rotating in the xy
If one needs a one-period trajectoid, deforming T by just two scaling plane is applied. For example, a linear increase of the phase
factors in orthogonal directions is not always enough to fulfil RAΩ = 1. shift ψ(t) = t ∆ will occur when the pulse is detuned slightly off-
In practice, for some paths, no combination of kx and ky can yield resonance25,26: that is, it oscillates with frequency (ω0 − ∆) instead
θ(kx , k y) < 2° (a mismatch angle acceptable in light of the imperfections of ω0. If the phase shift ψ(t) and magnitude B1(t) change slowly with
of our fabrication process). In these cases (Fig. 4d,e,n,o), we slightly time t, it is common to use the reference frame that rotates togeth-
adjusted the nodes of the Bezier spline through which the initial T was er with Larmor precession and formulate optical Bloch equations
defined (shown in Extended Data Fig. 2e) and recalculated the θ(kx , ky) in a rotating wave approximation26,27, where state vector ρ(t) is the
map again (Extended Data Fig. 2d,f). Generally, global optimization normalized magnetic dipole moment in the rotating frame, and
algorithms must be used for such deformation of the input path, its instantaneous angular velocity n(t ) = −γB1(t )[ˆcosx ψ(t ) + yˆsinψ(t )]
but, for instance, a path in Extended Data Fig. 2e needed just three to belongs to the xy plane. For the application of the TPT theorem, see
four iterations of adjustment to end up with a minimum value the discussion after (3).
θ(kx , ky) < 2°. (3) The same Bloch sphere representation of state evolution applies
Instead of deforming the input path, a special bridge may be also to the electric dipole transitions in a two-state quantum system
appended to it such that a one-period trajectoid exists for the overall under driving field E (t )cos(ω0t − ψ(t )) near the resonant frequency
curve (input path plus the bridge). To improve the practical perfor- ω0 = W /ħ, where W is the energy difference between the two eigen-
mance, we avoid using sharp corners in the bridge: curvature of bridge states. In this case, the state vector ρ(t) with unit length and com-
elements must not exceed certain κmax. The bridge thus consists of two ponents (u(t ), v(t ), w(t )) (Extended Data Fig. 9a) is not in real space
straight arms (Extended Data Fig. 3, orange) connected to each other coordinates anymore: coordinates u and v are called ‘coherences’26
and to the ends of input path by arcs of curvature κmax (Extended Data and w is the population inversion25,26, which is equal to +1 and −1 for
Fig. 3, red). Each bridge is made to close the spherical trace of the over- the system’s eigenstates (poles of the Bloch sphere in Extended
all curve (input path plus the bridge, Extended Data Fig. 3b). To ensure Data Fig. 9a). Under applied driving field, the state vector ρ(t), whose
that trajectoid does not need to rely on inertia, straight sections of the end is shown by a red circle in Extended Data Fig. 9a, has instan­
bridge must always be directed ‘downhill’ with respect to the projection taneous angular velocity vector n(t) = − κE (t)(x^ cosψ(t) + y^ sinψ(t))
of gravity onto the plane (in Extended Data Fig. 3a taken to be from the (blue arrow in Extended Data Fig. 9a), where κ = µab /ħ and µab is
left to the right). For a more robust practical performance, the elements the transition dipole moment25. Rotation of ρ(t) is shown by an
of bridge must not self-intersect. Under all these constraints, for a given orange arrow in Extended Data Fig. 9a.
scale σ of the input path (for example, blue in Extended Data Fig. 3a,b), As in the case (1), in cases (2) and (3) a scale factor r can be applied
we construct a family of bridges shown in red and orange Extended either to the pulse magnitude ( B1(t) → B1(t)/r , E (t) → E (t)/r , green
Data Fig. 2b and parameterize by declination of arms (orange) away curves in Extended Data Fig. 9b,c and right side of Supplementary
from the arc connecting the ends of the spherical trace of the scaled Video 5), or by stretch­ing pulse functions in time ( B1(t) → B1(rt) ,
input path. In this family, we then find a bridge such that the overall E (t) → E (r ), ψ(t) → ψ(rt), blue curves in Extended Data Fig. 9b,c and
spherical trace (of input path and the bridge together) splits the sphere left-hand side in Supplementary Video 5). By the TPT theorem, al-
into two spherical regions of equal area (2πr2) and by the Gauss– most any given pulse can be scaled such that applying the scaled
Bonnet theorem satisfies the RAΩ = 1 condition. An example of such pulse twice brings the system back to the original state. Note that
a solution is shown in Fig. 2c and validated experimentally in Fig. 4g. only the pulse envelope (B1(t) or E (t), dotted curve in Extended Data
The entire algorithm begins with a value of σ giving the smallest pos- Fig. 9b) and phase shift curve ψ(t) (Extended Data Fig. 9c) must be
sible distance between the ends of the input path’s spherical trace copied and shifted in time by a delay L, but the carrier wave (at fre-
(Extended Data Fig. 3b, blue). If we are unable to find a proper bridge quency ω0) must be shared between the scaled copy and the scaled
among the family of bridges for a given σ , we decrease σ and try again. original: for example, if scaling is applied to the electric field mag-
nitude, then the scaled shifted copy is (1/r )E (t − L)cos(ω0t − ψ(t − L))
Application of the TPT theorem in quantum and classical systems and the scaled origi­nal is (1/r )E (t)cos(ω0t − ψ(t)).
Here, we illustrate how the TPT theorem can be translated from trajec- (4) In classical optics, the polarization state of light can be represented
toids to four other physical systems governed by the rotation group by a Stokes vector on the Poincaré sphere (Extended Data Fig. 9d),
SO(3): (1) magnetic dipoles in planar magnetic field, (2) NMR, (3) elec- with purely circular polarizations at the sphere’s poles, on the S3
tric dipole transitions in a two-state quantum system under driving axis, and various linear polarizations in equatorial S1, S2 plane.
electric field, (4) the polarization state of light in classical optics. The Consider a single thin uniaxial waveplate whose optic axis lies in
Bloch sphere representation describes cases (2) and (3), and the closely the plane of the waveplate and forms angle ψ /2 with the vertical
related Poincaré sphere representation describes case (4). By almost any axis of the global reference frame as shown for one of the waveplates
sequences, we henceforth mean those with property π. in Extended Data Fig. 9e. When light of wavelength λ passes nor-
(1) In a time-varying planar magnetic field B(t) = B(t)(xˆcosψ(t) + mally through such a waveplate of thickness δt , the Stokes vector
yˆsinψ(t)), a magnetic dipole moment rotates with instantaneous on the Poincaré sphere rotates by an angle 2π(ne − no)δt /λ around
angular velocity n(t) = − γB(t), where γ is the gyromagnetic ratio. the axis n (blue arrow in Extended Data Fig. 9d) that lies in the equa-
The dynamics of a dipole in a field of constant magnitude ( B(t) = B) torial S1, S2 plane (shaded blue in Extended Data Fig. 9d) at an angle
has been described24 by rolling a sphere of radius r = 1/(γB) along ψ to the S1 axis33,34. Here, ne − no is the difference of refractive indices
a flat curve that has n(t) as its normal vector. A scale factor r can be between extraordinary and ordinary waves. Extended Data Fig. 9e
applied in two ways: either to the field magnitude ( B(t) → B(t)/r ), shows a sequence of waveplates of various thicknesses and
 orientations of optic axes (indicated by arrows). By the TPT theorem
in this case, after scaling all thicknesses by an adequate factor 1/r Code availability
as illus­trated in Extended Data Fig. 9f, and then doubling the An online open-source demonstration on Google Colab allows users
sequence of waveplates as in Extended Data Fig. 9g, the resulting to design trajectoids for a path of their choice: https://colab.research.
system will have no net effect on the polarization state of light pass- google.com/drive/1XZ7Lf6pZu6nzEuqt_dUCHormeSbCCMlP. All
ing through it. This analogy is illustrated in Supplementary Video 6. Python code used to obtain results that are reported in the paper is avail-
able from a public repository: https://doi.org/10.5281/zenodo.8116412.
Small-radius limit of trajectoids
Consider a minute sphere rolling along a differentiable planar curve T 40. Autodesk 3ds Max (Autodesk, 2018).
with a geodesic curvature κ(t). At the scale of the tiny sphere, r → 0 or 41. Hess, R. Blender Foundations: The Essential Guide To Learning Blender 2.6 (Focal Press,
2010).
σ → ∞, the geodesic curvature vanishes, rκ(t) → 0, such that T is effec- 42. Nielsen, M. A. & Chuang, I. L. Quantum Computation and Quantum Information 177
⌢
tively straight and the mapped curve on the sphere T is a geodesic, (Cambridge Univ. Press, 2010).
that is, a great circle (Extended Data Fig. 10a–c and Supplementary
Video 8, 0:00–0:13). Therefore, the total 3D curvature of the spherical Acknowledgements This work was supported by the Institute for Basic Science, Republic of
⌢ Korea, project code IBS-R020-D1 and by Swissmap (to J.-P.E.).
curve is κT = 1/r 2 + κ(t) 2 ≃ 1/r . The path T will remain in the vicinity
of the great circle and changes in the direction κ = dψ /dt will only tilt
Author contributions Y.I.S. designed the algorithm that solves for trajectoid’s shapes,
it very slightly (Extended Data Fig. 10c). The resulting trajectoid is implemented it in the software, fabricated the trajectoids, performed experiments,
therefore a simple cylinder. Only the introduction of infinitely sharp trajectory tracking and analysis. R.D. helped with experiments. T.T. and J.-P.E. developed
the mathematical treatment. T.T. and Y.I.S. developed quantum-mechanical and classical
corners, where κT = κ → ∞, can make the curve markedly depart from
analogies. S.G. and B.A.G. conceived and supervised research. Y.I.S., S.G., J.-P.E., T.T. and
the great circle. In a polygonal curve (Extended Data Fig. 10d), each B.A.G. wrote the paper.
smooth segment will be mapped to a great circle, and corners will
induce transition among the circles (Extended Data Fig. 10e–g and Competing interests The authors declare no competing interests.
Supplementary Video 8, 0:13–0:26).
Additional information
Supplementary information The online version contains supplementary material available at
https://doi.org/10.1038/s41586-023-06306-y.
Data availability Correspondence and requests for materials should be addressed to Yaroslav I. Sobolev,
All data in support of the findings of this study, including 3D shape files Tsvi Tlusty, Jean-Pierre Eckmann, Steve Granick or Bartosz A. Grzybowski.
Peer review information Nature thanks Peter Varkonyi and the other, anonymous, reviewer(s)
of demonstrated trajectoids, are available from a public repository: for their contribution to the peer review of this work.
https://doi.org/10.5281/zenodo.8116412. Reprints and permissions information is available at http://www.nature.com/reprints.
Article

Extended Data Fig. 1 | Additional illustrations. a, Potential energy curve

corresponding to one half of a period of a sideways rolling of a capped-cylinder
trajectoid (orange, yellow) from Fig. 1d. The potential energy curve (green)
starts at a minimum energy mgr , follows a cycloid until reaching maximum
energy mgR and remains constant afterwards until the edge E contacts the
plane. In other words, spherical side caps of radius R define the maximum of
the potential surface. Increasing R increases the mg (R − r ) difference and
deepens the “trench” in the gravitational potential surface, making the
adherence to the target path T∞ more reliable. b, Relative positions of objects
defined in the numerical implementation of the algorithm. See Methods for
details. Sphere G (the initial “shell” mesh, radius R) is concentric with r -radius
core W and intersects with the cube C , whose top face is tangent to W and lies in
the plane Q of rolling. c,d, Illustration of the MPT Theorem (see Methods for
details) with the same path as in Fig. 3d for n = 3 (c) and n = 5 (d).
Extended Data Fig. 2 | Deforming the input path non-uniformly to make it corresponding to the local minima marked as b and c in the θ (kz , ky ) map in
have a one-period trajectoid. a, Angle θ between starting and final orientations panel (a). d–f, Deformations additional to non-uniform scaling affect the
of a sphere rolled along a target path (same as in Fig. 2a and Fig. 4d) that was minimal mismatch on the θ (kz , ky ) map: a path (blue, same as in Fig. 2a and
deformed non-uniformly in two directions by scaling factors kx and ky (see Fig. 4d) is partially modified by adjusting its Bezier curve nodes; the modified
Methods for details). Examples of deformed path on the left correspond to section is shown in orange in e. Maps (d,f) of mismatch angle θ for both the
ky = 1, 2, 3 and fixed kx = 1; examples below correspond to kx = 0.5, 2, and ky = 1. original path (d) and its modified version (f) are plotted for the vicinity of the
Fulfilling the RAΩ = 1 condition requires that θ be zero (blue regions). b,c, Curves useful minimum of the original path (see full map in panel a). Note the increase
traced over the sphere by sphere-plane contact point when using paths of mismatch in f compared to d.
Article

Extended Data Fig. 3 | “Bridging” the input path to make it have a period-
one trajectoid. Example of input path (blue) with “bridges” (red and orange)
appended to it. Bridge consists of three arcs (red) and two straight “arms”
(orange). b, Family of bridges constructed for the path in a at a fixed scale.
Extended Data Fig. 4 | Existence of trajectoids completing two periods in formula (6) in Methods: θ = arccos((trRAΩ − 1)/2). Bottom plot in each panel
one revolution: paths having Property π. a,d,g, Two periods of the input shows oriented spherical area S(σ ) enclosed by the spherical trace of scaled
planar, complex paths. Color represents progression within a single period first period, also plotted against the scale σ . Scale corresponding to a two-
(from blue to yellow, see color scale in a). Orange circle shows diameter of the period trajectoid is marked by red dots. c,f,i Spherical trace of contact point of
sphere from c,f,i relative to the path. b,e,h, Correspond to paths in a,d,g a unit-radius sphere rolling along scaled paths in a,d,g corresponds to value
respectively. Top plot is the mismatch angle (degrees) between initial and final of σ indicated by red dots in respective plots b,e,h. a, Archimedes spiral with
orientations of the sphere after completing two periods of the scaled path – random noise added. d,g, Path obtained by a 2D random walk (making equal
plotted against the path scale σ . This angle is obtained by Euler’s axis-angle steps in random directions) – in piecewise linear version (d) and smoothed
representation of the matrix of net rotation accumulated by the sphere version (g).
Article

Extended Data Fig. 5 | Search for paths not having Property π. Representation g, V-like path similar to the one in panel a, but here the acute corners are tapered
analogous to Extended Data Fig. 4, but for different input paths. a, Acute-angle (“dulled”) by straight segments A1 T1, T2T3, T4M1. Trajectoid for this path exists
isosceles V-path does not have Property π : for this path, |S(σ )| < π for all σ . The (red points in h). k, Asymmetric V-like path with kinks – as in panel d, but here
trace c corresponds to a one-period trajectoid as shown by the green dot in b. kinks in two arms are unequal: angle of the left kink is α , but angle of right kink is
d, V-like path whose arms have “kinks” parameterized by angle α . It does not β = α (1 + Ξ). Trajectoid for this path exists (red points in l). In panels k,l,m, we
have Property π , and a two-period trajectoid does not exist for this path. The used asymmetry Ξ = 0.37. See Supplementary Video 4 for dependence of these
trace f corresponds to a green dot in e showing a near miss (i.e., not a trajectoid). plots on angle α in d, taper ratio in g, and asymmetry Ξ in k.
Extended Data Fig. 6 | Trajectoid tumbling at sharp turns. In Fig. 4d, g, h, the
trajectories feature some sharp turns (marked therein by green circles) at which
trajectoids tumbled/recoiled. These features of dynamics can be understood
by looking at the trench made by the trajectoid in the gravitational potential
surface. The representation here resembles Fig. 1b except that the trench
makes a sharp turn. In this dynamic analogy, resulting motion can be illustrated
by a small particle (colored here in blue) rolling down such a trench: prior to
encountering a sharp turn, the particle may have a zig-zag trajectory bouncing
between the trench walls (cf. Fig. 4d, h yellow arrows), and then might recoil
back at the sharp turn instead of making the turn immediately. It may take a few
bounces before the particle finally turns the corner. For the actual trajectoids
we fabricated, these bounces and zig-zagging correspond to precessions that
may result in net rotation of the object around the axis normal to the sloped
plane. In the case of the second corner in Fig. 4d (bottom orange circle), the
unintended net rotation due to recoils causes a turn of the entire subsequent
trajectory with respect to the intended path T. We believe that these dynamic
defects may be minimized by engineering the trench profile (presently we keep
it cycloid-like, which might be overly sharp) or making smoother turns of path T.
Article

Extended Data Fig. 7 | Energetic considerations for intermittent uphill insufficient to escape walls of the potential trench that follow the target
rolling. Energy landscapes illustrating design considerations for traversal of path. b, When losses are too high, the net energy is insufficient for an uphill
uphill excursions (as in the path in the main-text Fig. 4n, o). The gravitational excursion. When losses are too low, as in c, net energy decreases more slowly
acceleration vector g (vertical arrow) points down. The trajectoid’s total than the potential energy and becomes sufficient to escape the potential
energy (blue surface) decreases with descent due to losses such as friction. trench as indicated by the dashed arrow. Note that strictly speaking, the net
The potential energy surface (green) is defined by trajectoid’s shape and the energy is not a function of only the 2D location on the plane and, furthermore,
angle of the slope. Parts of the potential surface that are below the total energy the potential energy depends on the trajectoid’s 3D orientation, which is not an
surface are accessible to the rolling trajectoid. Proper balance between losses, unambiguous function of the 2D planar location even for slipless rolling. Still,
shape, and slope are shown in a: the total energy suffices to overcome the the concept of potential landscape is useful as a first approximation.
potential barrier along the target path (uphill excursion, short arrows), yet is
Extended Data Fig. 8 | Centroid of a visible shape as an estimator of the trajectoid (same as in Fig. 4k) whose surface has black dots (markers)
center of mass. a–c, Comparisons between the target path of a trajectoid painted on it; g shows the same raw frame overlaid with the trajectoid’s green
(solid black curves) and the theoretical trace of the centroid of trajectoid’s 3D mesh (the one used for 3D-printing the trajectoid) matching the location
shape projection, simulated assuming perfect performance (orange curves). and orientation evaluated by 6D pose tracking algorithm applied to apparent
The difference between curves in a is barely visible. Note that the discrepancy trajectories of black markers on the video (reconstructed 3D locations of
between orange and black curves in a,b,c is smaller than or comparable to the markers are shown in orange). h, Comparison of theoretical (intended) path of
discrepancies between the experimental trajectories and the respective the 2D projection of the center of mass (black curve) and the experimental path
target paths (Fig. 4o, n, l). d,f, For every possible orientation of the trajectoid evaluated by two methods: by the centroid of the visible shape (blue curve) or
(parameterized by latitude and longitude of contact point on the trajectoid), by the 6D pose tracking algorithm (orange curve). Both methods were applied
color on these maps shows planar distance between the projection of the to the same video. i, Difference in the Y coordinate between the experimental
center of mass and the centroid of the shape projection. Maps in d and f are and the intended path for the two methods of evaluating the center-of-mass
computed for the trajectoids constructed for paths b and c, respectively. Note location: here, the black curve in h has been subtracted from the blue and
that the distance in these maps never exceeds of the trajectoid’s minimal radius orange curves in h, respectively. The standard deviations for the two methods
r = 1. e,g, Illustration of 6D pose (full orientation and position) tracking from are: 0.845 mm when using centroid of visible shape, 0.765 mm when using the
video frames (see Supplementary Video 7): e is the raw video frame of the 6D pose tracking.
Article

Extended Data Fig. 9 | Optical and quantum-mechanical analogies to polarization state of light. Squares show polarization states corresponding to
the existence of two-period trajectoids. a, Illustration of the Bloch sphere respective black points: two circular polarizations at the poles and four linear
representation of a single qubit. The state represented by a red circle on the polarizations at the equatorial plane. e–g, Given almost any (i.e. those having
sphere rotates (orange arrow) around an instantaneous axis n(t ), which is Property π ) sequence of waveplates (dark blue in e), their thicknesses can be
defined by the driving field. b,c, Field pulse equivalent to a single period T scaled (f) by such a factor 1/r that the doubled sequence (g) has no net effect on
(b, solid curves), its envelope (b dashed curves) and phase shift (c) as functions polarization state of light (yellow helices) passing through it. In this example,
of time. Shown are two possible analogies to varying the radius r of the rolling curved green arrow shows left-handed circular polarization, curved orange
sphere in case of a given field pulse (black): either scaling the applied pulse’s arrows show right-handed elliptic (f) and right-handed circular (g) polarization.
magnitude (green) or stretching the pulse’s functions (envelope and phase See also Supplementary Videos 5, 6.
shift) in time (blue). d, Illustration of the Poincaré sphere representation of
Extended Data Fig. 10 | Small-radius (“adiabatic”) limit of trajectoids. details. d,e–g, Effect of ball radius r (or path’s scale σ = L /(2πr )) on the trace of
a–c, Effect of ball radius r (or path’s scale σ = L /(2πr )) on the trace of the contact the contact point upon rolling along a polygonal path (path shown in d). Value
point upon rolling along a finite smooth path (two periods of path from Fig. 4k). of r decreases from e to f to g. See Supplementary Video 8 for more details.
Value of r decreases from a to b to c. See Supplementary Video 8 for more