Non-planar 3D Printing of Double Shells

Ioanna Mitropoulou1[0000-0001-7528-6216], Amir Vaxman2[0000-0001-6998-6689], Olga Dia-

manti3[0000-0003-2883-2194], Benjamin Dillenburger1[0000-0002-5153-2985]
1 Digital Building Technologies, ETH Zurich, Switzerland
2 Institute of Perception, Action and Behaviour, The University of Edinburgh, UK
3 Institute for Geometry of the Department of Mathematics, TU Graz, Austria.

Abstract. We present a method to fabricate double shell structures printed in

transversal directions using multi-axis fused-deposition-modeling (FDM) robotic
3D printing. Shell structures, characterized by lightweight, thin walls, fast
buildup, and minimal material usage, find diverse applications in prototyping and
architecture for uses such as façade panels, molds for concrete casting, or full-
scale pavilions. We leverage an underlying representation of transversal strip net-
works generated using existing methods and propose a methodology for convert-
ing them into printable partitions. Each partition is printed separately and assem-
bled into a double-shell structure. We outline the specifications and workflow
that make the printing of each piece and the subsequent assembly process feasi-
ble. The versatility and robustness of our method are demonstrated with both
digital and fabricated results on surfaces of different scales and geometric com-
plexity.

Keywords: Robotic 3D Printing, Non-planar, Strips, Digital Fabrication, Dou-

ble Shell, Print Path Design, FDM, Digital Fabrication.
2

Fig. 1. FDM robotic non-planar 3D printing of double shells along transversal directions. Fabri-
cation of the Costa minimal surface using paths aligned with principal curvature directions. (a)
Input Strip-Decomposable Quad (SDQ) mesh. (b) Fabrication-aware partitioning for the 𝑈𝑈 (top)
and the 𝑉𝑉 (bottom) strip networks, each partition is shown with a different color. (c) Printing of
individual pieces and assembly. (d) Final prototype.

1 Introduction

Latest developments in Additive Manufacturing (AM) have opened new possibilities

for 3D printing objects with unprecedented geometric complexity. We focus on fused
deposition modeling (FDM), whereby a linear molten thermoplastic material is ex-
truded from a hot nozzle into layers that accumulate to produce an object. FDM is rel-
atively inexpensive and flexible (Gao et al. 2015) and has gained much popularity
among industrial and hobbyist practitioners.
The introduction of robotic arms in FDM printing enables the scaling up of the print-
ing process. In larger-scale printing, material deposition often takes the form of a hol-
low shell, where each layer is realized by a single path without multiple offsets or infill.
Shells have the advantage of less material usage and faster buildup, and thus are used
for various architectural applications such as façade panels (Sarakinioti et al. 2018;
Taseva et al. 2020), bespoke floors (Aectual 2019), and molds for concrete casting (Jipa
et al. 2017; Leschok and Dillenburger 2019; Burger et al. 2020).
However, the realization of single-layer FDM shells has some inherent disad-
vantages. First, the thickness of the shell is constrained by the size of the extrusion
nozzle. For FDM printing using filament extrusion, this is typically between 0.5 and
2.5 mm, which can cause instabilities when scaling up. Second, due to the one-dimen-
sional nature of the paths, there is an asymmetry in the quality of the approximation of
the surface: the direction parallel to the paths is smoothly approximated, while the di-
rection orthogonal to the paths is only approximated with a staircase of subsequent sec-
tions (Fig. 2a). This adversely affects both aesthetics, with a staircase effect in the di-
rection orthogonal to the paths, and structural strength, as the resulting object is
 3

anisotropic, with considerably higher strength along the print direction (Ulu et al.
2015). While both the staircase effect and anisotropy are hard to avoid since they are
inherent to the printing process, they are properties that can be controlled to improve
the strength and quality of the printed object.
To that end, we propose the design of double (Fig. 1, Fig. 2c) instead of single (Fig.
2b) shells, with two outer walls printed using non-planar paths in transversal directions
and connected with sparse links (ribs) along a hollow interior. This adds structural depth
to the resulting object, thus improving stability. It also allows the distribution of imper-
fections more evenly along different directions on each side of the shell.

2 Related work

2.1 Non-planar slicing and multi-axis printing

Traditional planar 2.5D printing (Gibson, Rosen, and Stucker 2009) is based on the
accumulation of planar layers along a single build direction. Multiple approaches for
non-planar printing still follow the layered approach but allowing for curved layers. In
perhaps the first work in that direction (Chakraborty, Aneesh Reddy, and Roy
Choudhury 2008), the printer can dynamically change 𝑧𝑧-values within individual lay-
ers. Since then, multiple works have pursued the direction of decomposing a shell or
volume into curved layers (Ahlers et al. 2019; Allen and Trask 2015; Chen et al. 2019;
Etienne et al. 2019; B. Huang and Singamneni 2015; Lim et al. 2016; Llewellyn-Jones,
Allen, and Trask 2016; Pelzer and Hopmann 2021). Curved slicing produces non-planar
print paths that can greatly reduce support, improve surface quality (in particular, re-
duce staircase artifacts), and improve mechanical properties.
To facilitate the fabrication of non-planar paths, one can use a robotic arm with more
degrees of freedom (DOF) compared to the standard triaxial system of traditional desk-
top printers (Keating and Oxman 2013; Pan et al. 2014; C. Wu et al. 2017). Multi-DOF
systems have been used to print wireframe models (Y. Huang et al. 2016; R. Wu et al.
2016). They have also been used for layered printing; for example, Dai et al. (Dai et al.
2018) use a multi-DOF system that keeps the printer head fixed and freely moves the
object during printing to print curved layers that eliminate sacrificial support. However,
moving the object during the fabrication process can be challenging when printing
larger objects. Moving the print tool instead of the object is an alternative used in vari-
ous research projects (Mitropoulou, Bernhard, and Dillenburger 2020; Fang et al.
2020a; Zhang et al. 2022).

2.2 Fabrication in transversal directions

An important challenge of FDM 3D printing is the anisotropic behavior of printed ob-
jects due to the weak adhesion between layers along the vertical axis (Ulu et al. 2015;
Fang et al. 2020b). In response to this challenge, researchers have explored alternative
printing paths using transversal directions, such as aligning print directions with stress
lines (Tam and Mueller 2017), also integrating carbon fiber (Kwon et al. 2019). Utiliz-
ing transversal directions to enhance fabrication isn’t confined to just 3D printing. This
4

approach is prevalent in fabrication processes that involve linear elongated segments,

where the discontinuous direction has weaker bonding than the continuous one. Exam-
ples include crafting surfaces with paper strips (Takezawa et al. 2016) or metal rods
(Ma et al. 2020). The key advantage is that when directions intersect, the strengths in
one direction compensate for the vulnerabilities in the other, resulting in a more robust
final product.

2.3 Strips
Utilizing strips to represent the underlying geometry is common in such fabrication
processes because the geometric characteristics of strips match the elongated fabrica-
tion units. Strips have been used to design single-curved façade panels (Pottmann et al.
2008), textile ribbons (Schüller, Poranne, and Sorkine-Hornung 2018) and developable
surfaces (Verhoeven et al. 2022).
In the same spirit, (Mitropoulou et al. 2024) design transversal strip layouts over-
layed into a Strip-Decomposable Quad (SDQ) mesh (Fig. 2d), considering fabrication-
related properties and user-controlled alignment. They propose using SDQ meshes as
an intermediate representation to describe non-planar 3D printing paths. Our work fol-
lows that paradigm; however, we take advantage of both transversal strip networks to
fabricate double instead of single shells printed along transversal directions. Further,
we extend this scheme with a partitioning method that guarantees both topological and
geometric feasibility for printing.

3 Overview

We begin with a set of two transversal strip networks, 𝑈𝑈 (depicted in blue) and 𝑉𝑉 (de-
picted in green), combined in an SDQ mesh (Fig. 2d). The novelty of this work consists
of considering each strip network to represent the non-planar print paths of each side
of a double shell, printed in transversal directions. The challenges we tackle to enable
this fabrication mode are the following.
─ Partitioning strip networks into printable patches that can be fabricated with our print
setup (Sections 4.1, 4.2).
─ Addition of fabrication-enabling details and assembly into one connected double
shell object (Section 4.3).

3.1 Paths representation

Paths are generated by subdividing 𝑁𝑁 times each strip 𝑆𝑆𝑖𝑖 . Unlike planar printing, here,
the print direction 𝑇𝑇 (i.e., the orientation of the robot’s print head while printing) and
the layer height ℎ are variable. Fig. 2e shows how paths are generated on a strip 𝑆𝑆𝑖𝑖 and
the resulting 𝑇𝑇 and ℎ for two points 𝑎𝑎 and 𝑏𝑏. 𝑇𝑇 is given by the direction of the transversal
(green) edge, and ℎ is given by the distance from the neighboring subdivision.
 5

Fig. 2. (a) In 3D printed surfaces, one direction is smoothly approximated, while the other is
approximated with a staircase of subsequent sections. (b, c) Diagram of a single and a double
shell with each side printed in transversal directions. (d) Two transversal strip networks, 𝑈𝑈 (left)
and 𝑉𝑉 (middle), overlayed into one SDQ mesh (right), illustraton from (Mitropoulou et al. 2024).
(e) Paths generation by subdividing a strip 𝑆𝑆𝑖𝑖 showing the layer height ℎ and print direction 𝑇𝑇 for
two points 𝑎𝑎, 𝑏𝑏.

3.2 Printing standing shells

Next, we describe our non-planar 3D printing setup of standing shells (Fig. 3, top), i.e.,
shells where only the first path lies on sacrificial support, and all subsequent paths lie
on previously deposited paths. The robotic arm orients the extruder (i.e., the print tool)
towards the print direction 𝑇𝑇 at each position and prints sequentially on a sacrificial
base, from first to last path, in an uninterrupted motion. This sequential logic requires
a print sequence, where each path being printed fully lies on previously deposited paths.
Since the layer height ℎ, i.e., the distance between consecutive layers, varies, the ex-
truder updates the material flow rate at each print position to match the quantity of
material necessary for filling the current layer height. On the other hand, the layer width
𝑤𝑤 is considered constant throughout the process.
The robot’s degrees of freedom provide considerable flexibility for reorienting the
tool around the object and depositing material in various orientations. However, the
range of possible orientations for material deposition isn't unlimited. During the buildup
of paths, print angles that are unfavorable for material deposition can appear (Fig. 3,
bottom right). For that, we set a maximum print angle variation γ that should not be
exceeded (for our setup γ = 𝜋𝜋/2). In addition, the robot cannot print outside of its
reachable space; we define a bounding box of reachable space (for our setup, it has
dimensions 50 x 50 x 50 cm), and all print geometry must lie within it.
6

Fig. 3. Top: Overview of a standing shell's robotic 3D printing process. Bottom: Print angles
during the buildup of paths can become unfeasible (right, in red) if the print direction varies a lot.

Partitioning constraints

Considering the characteristics of the process outlined above, we propose a partitioning

scheme of strips that considers the following constraints.
1. Topologically feasible. All patches must have a topologically simple sequence of
paths so that there is an unambiguous print sequence. This should consist of a single
non-branching sequence of curves that layer upon each other so that each path being
printed lies on a previously deposited path.
2. Geometrically feasible. For each patch, the dimensions must fit within the bounding
box, and the angle variation should not exceed the maximum print angle variation.
3. Non-overlapping seams. The partitioning cuts on the two strip networks 𝑈𝑈 and 𝑉𝑉
must not overlap so that upon assembly, the two shells form one connected object.

3.3 Workflow
Our workflow proceeds as follows.
─ Begin with an SDQ mesh.
 7

─ Carry out topological partitioning that separates the strip networks into simply-
connected patches (Section 4.1).
─ Carry out geometric partitioning into patches that respect the angle and size bounds
(Section 4.2).
─ Incorporate assembly-related details and compute fabrication data (Section 4.3).
─ Print and assemble the individual pieces to obtain the final output (Section 5).

4 Method

4.1 Topological partitioning

Topological partitioning (Fig. 4) aims to separate the strip networks 𝑈𝑈 and 𝑉𝑉 into
simply-connected patches so that each patch has an unambiguous ordering (up to ori-
entation) and can be printed as a standing shell.
In (Mitropoulou et al. 2024), a partitioning scheme is detailed that cuts the strip net-
work along the separatrices of the singularities and separates handles, thus splitting a
strip network into simply-connected parts, as shown in Fig. 4a and Fig. 4d. These cuts
continue until they form closed loops or terminate on open boundaries. The difference
here is that we aim to partition both 𝑈𝑈 and 𝑉𝑉 (i.e., both Fig. 4a and Fig. 4b, similarly
both Fig. 4d and Fig. 4e), and we have the additional constraint that the cuts of the two
networks should not overlap. As long as the partitioning is parallel to the direction of
strips (i.e., no strip is cut perpendicularly), then the cuts of 𝑈𝑈 and 𝑉𝑉 are non-overlap-
ping, as they are applied in transversal directions, for example, see the partitioning of
the D6 singularity (Fig. 4 a, b, c).
However, around the D2 singularity (Fig. 4 d, e, f), cutting along the separatrix par-
allel to the strips alone would not separate the self-incident strip at the cut, which creates
paths unfeasible for printing due to their self-obstructing geometry. As a result, around
the D2 singularity, topological cuts are added that are both parallel to the strips (Fig. 4
d, e, in black) and intersect them transversally (Fig. 4 d, e, in red). This causes an over-
lap of cuts as, essentially, the cuts of the 𝑈𝑈 and the 𝑉𝑉 network are identical. If those cuts
were allowed to terminate naturally, they would create an extensive sequence of edges
where both sides share a cut. Then, upon assembly, the final object would be discon-
nected there. To avoid this, we introduce two additional transversal cuts 𝐶𝐶𝑈𝑈 and 𝐶𝐶𝑉𝑉 in
distance 𝑑𝑑𝑞𝑞 edges from the singularity, that interrupt the overlapping cut (the resulting
reduced overlap is marked in cyan). As a result, the pieces are still connected upon
assembly.

4.2 Geometric partitioning

We next partition the patches further to adhere to the sizing and angle constraints (Fig.
5). To avoid overlapping cuts in this step, we keep track of the edges that participate in
cuts on each network and disallow adding overlapping cuts. Since every edge sequence
of the mesh can potentially become a cut, there are always more than enough edges to
8

use as cuts despite this restriction. The geometric partitioning cuts are interrupted at the
boundary of each patch; thus, patches can be partitioned independently.

Fig. 4. Topological partitioning of the two singularities most commonly found in SDQ meshes
for both 𝑈𝑈 and 𝑉𝑉 strips. The strips’ outline is displayed as smooth for easier comprehension. (a,
b, c) D6 singularity partitioning along separatrices. (d, e, f) D2 singularity partitioning along
separatrices. The cyan line illustrates the overlapping cut on the two sides, bounded by the trans-
versal cuts 𝐶𝐶𝑈𝑈 , 𝐶𝐶𝑉𝑉 in distance 𝑑𝑑𝑞𝑞 edges from the singularity.

Partitioning to fit size bounds.

In each patch, we find the two transversal strips 𝑆𝑆𝑈𝑈 and 𝑆𝑆𝑉𝑉 with the largest area. In
addition, we compute the maximum extent of each patch using Principal Component
Analysis (PCA). Each resulting principal direction is best aligned with either 𝑆𝑆𝑈𝑈 or 𝑆𝑆𝑉𝑉 .
If a direction exceeds the maximum size, we partition the according strip by adding the
appropriate number of equidistant cuts (in number of quads), ignoring edges that would
cause overlap. The resulting partitions are checked again, and the process is repeated
until no piece exceeds the set bounds. Fig. 5 illustrates the partitioning of the 𝑈𝑈 and 𝑉𝑉
networks of a doubly-curved surface that is, by design, larger than the bounding box.

Partitioning to fit angle bounds.

While the robot is printing with the extruder oriented in the print direction 𝑇𝑇, its orien-
tation change should not exceed a certain threshold 𝛾𝛾. Note that the angle with which
the robot is printing depends on the piece’s orientation on the build platform. This is
 9

why we are not examining angles with the gravity direction (these are defined later
when the piece’s print orientation has been decided) but rather angle variations, which
are independent of the piece’s orientation on the print platform.
Assume we are working on a patch of 𝑈𝑈 strips; its 𝑈𝑈-strips (blue) lie along the paths’
direction, and its 𝑉𝑉-strips (green) are transversal to it. As a result, the total angle varia-
tion 𝐴𝐴 = ∑ |𝛼𝛼𝑉𝑉𝑖𝑖 | (the sum of absolute angle differences of neighboring 𝑇𝑇) within 𝑉𝑉-
strips measures the variation during the buildup of the piece. We consider the V-strip
with the greatest angle variation 𝐴𝐴𝑚𝑚𝑚𝑚𝑚𝑚 . If 𝐴𝐴𝑚𝑚𝑚𝑚𝑚𝑚 > 𝛾𝛾, then we cut the strip with 𝑘𝑘 =
𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐(𝐴𝐴𝑚𝑚𝑚𝑚𝑚𝑚 /𝛾𝛾) equidistant (in number of quads) cuts, again ignoring edges that would
cause overlap. All resulting patches are rechecked, and the process is repeated until no
patches have 𝑉𝑉 strips that exceed the 𝐴𝐴𝑚𝑚𝑚𝑚𝑚𝑚 . We show the process in Fig. 5, using dif-
ferent angle thresholds 𝛾𝛾.

Fig. 5. Partitioning steps: first, topological partitioning (left), and second, geometric partitioning
(right). Top: Partitioning a doubly curved surface from size bounds using two different bounding
box sizes. Bottom: Partitioning various surfaces from angle bounds using two different γ thresh-
olds. In both cases, only the 𝑈𝑈-strips partitioning is displayed.

4.3 From digital to physical

In the following, we provide the steps for preparing the digital double shell model to be
ready for printing.
10

Surface thickening – offsetting

The double-shell surface consists of two shells printed along transversal directions and
a hollow interior with sparse ribs connecting the shells. The total thickness 𝑡𝑡 of the shell
can be determined by the user, where in most prototypes, we use the default value t =
4n (Fig. 6a), where n is the nozzle diameter (2.5 mm in our setup). To account for the
thickness of the shells, we offset the original surface on each side with default distance
t/2 − 0.5n = 1.5n in the direction of the normals of the mesh, which we compute
for each quad by triangulating it and finding the average normal of its triangles. This
offsetting scheme is adequate for our application since our shapes have low surface
detail, and the quad mesh has sufficiently high resolution.

Managing inaccuracies during assembly

The assembly of the printed pieces might be hampered by small inaccuracies present in
the prints. They originate either from inevitable process imperfections or minor defor-
mations of the pieces during and after the print due to uneven shrinkage while cooling.
To account for them, we strategically introduce tolerances by removing a single strip
on each boundary between patches. We do this only on one side (by default, the 𝑉𝑉 side,
Fig. 6b) to avoid creating holes that would otherwise appear at the intersection of gaps.
For small inaccuracies, this strategy is sufficient to compensate for errors on both sides
of the shell.

Fig. 6. (a) Offsetting the original surface (in red dotted line) on both sides to create a double shell
of thickness 4𝑛𝑛. (b) Accounting for tolerances by creating one side with gaps (right, 𝑉𝑉-strips).

Orienting for fabrication

To decide on the global print orientation of a piece, we calculate the average ℎ�⃗ of all
directions 𝑇𝑇 (Fig. 7a, left, in black). We also calculate the average 𝑚𝑚
��⃗ of the first strip’s
 11

𝑇𝑇 directions (Fig. 7a, left, in red). Then, the piece’s global orientation is calculated so
that the vector 0.5(ℎ�⃗ + 𝑚𝑚
��⃗) is aligned with the vertical 𝑧𝑧⃗ direction (Fig. 7a, right).

Sacrificial support
Once the pieces have been oriented for fabrication, a supportive base is created upon
which the piece can be printed. It consists of a scaffolding, a lighter lower part produced
by a liner hatch to save time and material, and a platform, an upper part, which connects
to the printed object and consists of layered offsets of the first path of the patch (Fig.
7b).

Fig. 7. (a) Calculation of orientation for fabrication so that 0.5(𝑚𝑚 �⃗) is aligned with the vertical
��⃗ + ℎ
�⃗
direction 𝑧𝑧⃗, where ℎ is the average of all directions 𝑇𝑇 (vectors in black), and 𝑚𝑚 ��⃗ is the average of
the first strip’s 𝑇𝑇 directions (vectors in red). (b) Sacrificial base diagram (left) and printing (right).
(c) Rib interruptions for interlocking the two sides of the shell. The continuous ribs of one wall
are denoted with red lines. The ribs on the other side (colored per patch) are interrupted to create
an interlocking. (d) Printed prototype.

Ribs and connections

Along each wall, we print rigidifying ribs, which help with the overall stability and link
the two sides of the shell upon assembly. The ribs follow the direction transversal to
the one we are printing and have a default distance of 4 strips (chosen for aesthetics).
Similarly to the print paths, the ribs intersect transversely on the two sides of the shell.
To create an interlocking connection between the two sides, we interrupt the ribs of one
side on the intersections with the ribs from the other side (Fig. 7 c, d). This provides a
guide for positioning the pieces during assembly while allowing minor adjustments by
sliding, as we leave an additional tolerance gap between the ribs at their intersections.
The pieces are kept in place using screws on rib intersections (visible in Fig. 6b right),
which are drilled manually. The assembly process (displayed in Fig. 9 for prototypes d
and e) consists of assembling one piece from each side in an alternating pattern at every
step so that each new piece is half attached to the existing structure.
12

Printing process
Our printing setup consists of a universal robot UR10 mounted overhead to gain better
access to all sides of the printed object (Fig. 8). The plastic extruder tool has a nozzle
with a diameter of 2.5mm and is controlled by an Arduino Mega board. The printing
process is carried out at a constant linear robot speed at the extrusion point (15 mm/sec
while printing the object and 23 mm/sec while printing support). We vary the extrusion
rate to match the volume of material required to achieve the desired path thickness at
every printing position. The change of filament color is carried out manually without
stopping the printing process, and its purpose is solely aesthetic, to emphasize the path
orientations and differentiate the two shells visually.

Fig. 8. Top: Robotic setup for fabrication. Bottom: timelapse of printing, total time 35 min.

5 Results

We present a variety of printed prototypes created with the described methods. Fig. 1
presents the Costa minimal surface with the full pipeline from the initial input SDQ
mesh to the final fabricated result. The shape is partitioned into 7 pieces on one side
and 11 on the other. The prototypes shown in Fig. 9 a, b, c are created with an SDQ
mesh aligned to user-drawn directional constraints. The curved surfaces (Fig. 9 a, b)
are fabricated with 3 pieces on each side, while the elongated curved surface (Fig. 9c)
is partitioned into 5 and 6 pieces. The Chen Gackstatter minimal surface (Fig. 9d) is
generated using boundary-aligned constraints; the paths of one side are constrained to
 13

be parallel, and the paths of the other side to be orthogonal to the boundary. It is seg-
mented into 5 and 16 pieces. Finally, the Batwing minimal surface (Fig. 9e) is generated
with curvature-aligned directional constraints and is segmented into 9 and 12 pieces.
Table 1 presents fabrication details for all the prototypes.

Fig. 9. (a, b) Curved surfaces with paths aligned to two different user-drawn directions. (c) Elon-
gated curved surface with paths aligned to user-drawn directions. (d) Chen Gackstatter minimal
surface with paths aligned to boundary directions. (e) Batwing minimal surface with paths
aligned to curvature directions. More details on the prototypes are displayed in Table 1.

model Dimensions #sings #geomet- # pieces 𝑼𝑼 # pieces 𝑽𝑽 Total %

(cm) ric cuts print time sacrificial
(hrs) support
Fig. 1 60 x 60 x 51 4 12 7 11 51.5 49%
Fig. 9a 30 x 6 x 30 1 0 3 3 9.6 52%
Fig. 9b 30 x 6 x 30 1 0 3 3 8.9 48%
Fig. 9c 55 x 17 x 10 1 0 5 6 8.1 54%
Fig. 9d 50 x 50 x 60 4 8 5 16 46 42%
Fig. 9e 30 x 30 x 30 7 4 9 12 16.8 28%

Table 1. Details of prototypes presented in Fig. 1 and Fig. 9.

14

6 Conclusion

Our pipeline successfully produces quality print paths for various shapes. We note some
limitations and challenges.
Challenges of double shells. Despite the merits of printing double shells in transver-
sal directions, this fabrication mode also has notable challenges. Printing a double shell
implies that the surface is essentially printed twice, which leads to more pieces, more
sacrificial support for their base, and thus longer fabrication times. Further advancing
the partitioning method to consider the material efficiency of support structures can
help reduce the waste material. Also, the printing in transversal directions takes place
in separate pieces that are held together with sparse point links of small surface area
(screws). As a result, the connection of transversal directions is significantly weaker
compared to other approaches that print both directions directly on each other (Tam and
Mueller 2017; Kwon et al. 2019).
Assembly. A well-known challenge in FDM, which our work is also susceptible to,
occurs in the assembly of the printed paths. Minor inaccuracies caused by imperfections
of printing, such as over-extruding or deformations of the pieces during cool-down,
may cause them not to fit well together. When printing double shells, these can be es-
pecially problematic as the interfaces between pieces have a very large area; essentially,
the entire pieces are interfaces, meaning that any imperfection can impair the assembly
process. Future work into designing smarter part connections, such as puzzle joints or
joints that take advantage of the double-shell thickness of the components, could prove
very beneficial to the final quality of the assembled 3D-printed object.
Offsetting. A topic worth further study is how to offset shells optimally; our normal-
based offsetting only offers an approximate solution and may need to be adapted for
more intricate shapes.

6.1 Outlook
This work opens the door to a variety of promising future applications.
From multi-color to multi-material printing. We use colors to differentiate the
curved paths visually. Such changes in the extrusion’s content can have significant fu-
ture applications. Printing different materials, instead of just different colors, can create
functional material gradients that serve specific performance goals. For instance, fiber-
reinforced filament can be used strategically to improve the resulting piece's structural
properties. The development of specialized hardware that feeds the correct material at
the correct times during the print would contribute to such controlled functional mate-
rial changes.
Applications in architecture. 3D-printed shells in architecture find various applica-
tions but are often limited by their low thickness and anisotropy, which can be effec-
tively mitigated through the use of double shells printed using non-planar layers. Print-
ing each shell side separately and assembling them afterward is beneficial for embed-
ding functional equipment like tubes or wires and easily accessing them for repairs.
Also, it facilitates the incorporation of reinforcement between the two shells, enhancing
the overall functionality and durability of the shell structure. Further, add-on printing
 15

is a key application of non-planar shell printing, enabling direct material deposition

onto existing elements to improve efficiency, promote reuse, and facilitate repair or
reinforcement. Structural optimization is another important application. By exploiting
the anisotropic properties of 3D printed materials, objects can be optimized for specific
load cases. Printing each side of the shell with a different orientation allows further
control over the anisotropic properties, leading to additional mechanical improvements.

6.2 Reflection
Our research on the non-planar 3D printing of double shells using multi-axis FDM
resonates with the conference's theme, "Beyond Optimization," challenging traditional
approaches in robotic fabrication. We address this theme through two key aspects:
Embracing real-world complexity in assembly tasks. An important part of our as-
sembly method is the acknowledgment and proactive management of inaccuracies. Un-
like traditional approaches that strive for hyper-accurate results – often a challenging
and resource-intensive endeavor – our method anticipates and accommodates inaccu-
racies by incorporating gaps in the seams for handling tolerances. In the real-world
conditions of fabrication, absolute precision is often unattainable. This acknowledg-
ment of fabrication imperfections demonstrates a shift from pursuing unfeasible preci-
sion to developing robust and adaptable systems that do not fail when unavoidable im-
perfections occur.
Broadening the scope beyond the technical. Our research presents a novel way for
printing double shells using non-planar paths. Beyond functional efficiency, this con-
trol opens avenues for artistic expression and aesthetic enhancement. By manipulating
print paths, we unlock new design possibilities, creating textures and forms that en-
hance both the visual and tactile qualities of the final product, which can be particularly
valuable in artistic applications.

7 References

Aectual. 2019. "Pattern Terrazzo Floors." Accessed February 8, 2024.

https://www.aectual.com/architectural-products/flooring/variants/pattern-ter-
razzo/intro.
Ahlers, Daniel, Florens Wasserfall, Norman Hendrich, and Jianwei Zhang. 2019. “3D
Printing of Non-planar Layers for Smooth Surface Generation.” In 2019 IEEE
15th International Conference on Automation Science and Engineering (CASE),
1737–43. IEEE Press. https://doi.org/10.1109/COASE.2019.8843116.
Allen, Robert J A, and Richard S Trask. 2015. “An Experimental Demonstration of
Effective Curved Layer Fused Filament Fabrication Utilising a Parallel Deposi-
tion Robot.” Additive Manufacturing 8: 78–87.
https://doi.org/10.1016/j.addma.2015.09.001.
Burger, Joris, Ena Lloret-Fritschi, Fabio Scotto, Thibault Demoulin, Lukas Gebhard,
Jaime Mata-Falcón, Fabio Gramazio, Matthias Kohler, and Robert J Flatt. 2020.
“Eggshell: Ultra-Thin Three-Dimensional Printed Formwork for Concrete
16

Structures.” 3D Printing and Additive Manufacturing 7 (2): 48–59.

https://doi.org/10.1089/3dp.2019.0197.
Chakraborty, Debapriya, B Aneesh Reddy, and A Roy Choudhury. 2008. “Extruder
Path Generation for Curved Layer Fused Deposition Modeling.” Comput. Aided
Des. 40 (2). USA: Butterworth-Heinemann: 235–43.
https://doi.org/10.1016/j.cad.2007.10.014.
Chen, Lufeng, Man-Fai Chung, Yaobin Tian, Ajay Joneja, and Kai Tang. 2019. “Var-
iable-Depth Curved Layer Fused Deposition Modeling of Thin-Shells.” Robot-
ics and Computer-Integrated Manufacturing 57: 422–34.
https://doi.org/10.1016/j.rcim.2018.12.016.
Dai, Chengkai, Charlie C L Wang, Chenming Wu, Sylvain Lefebvre, Guoxin Fang,
and Yong-Jin Liu. 2018. “Support-Free Volume Printing by Multi-Axis Mo-
tion.” ACM Trans. Graph. 37 (4). New York, NY, USA: Association for Com-
puting Machinery. https://doi.org/10.1145/3197517.3201342.
Etienne, Jimmy, Nicolas Ray, Daniele Panozzo, Samuel Hornus, Charlie C L Wang,
Jonàs Martínez, Sara McMains, Marc Alexa, Brian Wyvill, and Sylvain
Lefebvre. 2019. “CurviSlicer: Slightly Curved Slicing for 3-Axis Printers.”
ACM Trans. Graph. 38 (4). New York, NY, USA: Association for Computing
Machinery. https://doi.org/10.1145/3306346.3323022.
Fang, Guoxin, Tianyu Zhang, Sikai Zhong, Xiangjia Chen, Zichun Zhong, and Char-
lie C L Wang. 2020a. “Reinforced FDM: Multi-Axis Filament Alignment with
Controlled Anisotropic Strength.” ACM Trans. Graph. 39 (6). New York, NY,
USA: Association for Computing Machinery.
https://doi.org/10.1145/3414685.3417834.
Gao, Wei, Yunbo Zhang, Devarajan Ramanujan, Karthik Ramani, Yong Chen, Chris-
topher Williams, Charlie Wang, Yung Shin, Song Zhang, and Pablo Zavattieri.
2015. “The Status, Challenges, and Future of Additive Manufacturing in Engi-
neering.” Computer-Aided Design 69 (August).
https://doi.org/10.1016/j.cad.2015.04.001.
Gibson, Ian, David W Rosen, and Brent Stucker. 2009. Additive Manufacturing Tech-
nologies: Rapid Prototyping to Direct Digital Manufacturing. 1st ed. Springer
Publishing Company, Incorporated.
Huang, Bin, and Sarat Singamneni. 2015. “A Mixed-Layer Approach Combining
Both Flat and Curved Layer Slicing for Fused Deposition Modelling.” Proceed-
ings of the Institution of Mechanical Engineers, Part B: Journal of Engineering
Manufacture 229 (12): 2238–49. https://doi.org/10.1177/0954405414551076.
Huang, Yijiang, Juyong Zhang, Xin Hu, Guoxian Song, Zhongyuan Liu, Lei Yu, and
Ligang Liu. 2016. “FrameFab: Robotic Fabrication of Frame Shapes.” ACM
Trans. Graph. 35 (6). New York, NY, USA: Association for Computing Ma-
chinery. https://doi.org/10.1145/2980179.2982401.
Jipa, Andrei, Mathias Bernhard, Benjamin Dillenburger, Nicolas Ruffray, Timothy
Wangler, and Robert Flatt. 2017. “SkelETHon Formwork 3D Printed Plastic
Formwork for Load-Bearing Concrete Structures.” Blucher Design Proceedings
3 (12): 345–52. http://dx.doi.org/10.5151/sigradi2017-054.
 17

Keating, Steven, and Neri Oxman. 2013. “Compound Fabrication: A Multi-Functional

Robotic Platform for Digital Design and Fabrication.” Robotics and Computer-
Integrated Manufacturing 29 (6): 439–48.
https://doi.org/10.1016/j.rcim.2013.05.001.
Kwon, Hyunchul, Martin Eichenhofer, Thodoris Kyttas, and Benjamin Dillenburger.
2019. “Digital Composites: Robotic 3D Printing of Continuous Carbon Fiber-
Reinforced Plastics for Functionally-Graded Building Components.” In Robotic
Fabrication in Architecture, Art and Design 2018, edited by Jan Willmann,
Philippe Block, Marco Hutter, Kendra Byrne, and Tim Schork, 363–76. Cham:
Springer International Publishing.
Leschok, Matthias, and Benjamin Dillenburger. 2019. “Dissolvable 3DP Formwork.
Water-Dissolvable 3D Printed Thin-Shell Formwork for Complex Concrete
Components.” In ACADIA 19: Ubiquity and Autonomy. Proceedings of the 39th
Annual Conference of the Association for Computer Aided Design in Architec-
ture, edited by Kory Bieg, Danelle Briscoe, and Clay Odom, 188–97. Associa-
tion for Computer Aided Design in Architecture (ACADIA).
Lim, Sungwoo, Richard A Buswell, Philip J Valentine, Daniel Piker, Simon A Austin,
and Xavier De Kestelier. 2016. “Modelling Curved-Layered Printing Paths for
Fabricating Large-Scale Construction Components.” Additive Manufacturing
12: 216–30. https://doi.org/10.1016/j.addma.2016.06.004.
Llewellyn-Jones, Thomas, Robert Allen, and Richard Trask. 2016. “Curved Layer
Fused Filament Fabrication Using Automated Toolpath Generation.” 3D Print-
ing and Additive Manufacturing 3 (4): 236–43.
https://doi.org/10.1089/3dp.2016.0033.
Ma, Z, A Walzer, C Schumacher, R Rust, F Gramazio, M Kohler, and M Bächer.
2020. “Designing Robotically-Constructed Metal Frame Structures.” Computer
Graphics Forum 39 (2): 411–22. https://doi.org/10.1111/cgf.13940.
Mitropoulou, Ioanna, Mathias Bernhard, and Benjamin Dillenburger. 2020. “Print
Paths Key-Framing: Design for Non-Planar Layered Robotic FDM Printing.” In
Symposium on Computational Fabrication. SCF ’20. New York, NY, USA: As-
sociation for Computing Machinery. https://doi.org/10.1145/3424630.3425408.
Mitropoulou, Ioanna, Amir Vaxman, Olga Diamanti, and Benjamin Dillenburger.
2024. “Fabrication-Aware Strip-Decomposable Quadrilateral Meshes.” Com-
puter-Aided Design 168 (March). Elsevier BV: 103666.
https://doi.org/10.1016/j.cad.2023.103666.
Pan, Yayue, Chi Zhou, Yong Chen, and Jouni Partanen. 2014. “Multitool and Multi-
Axis Computer Numerically Controlled Accumulation for Fabricating Confor-
mal Features on Curved Surfaces.” Journal of Manufacturing Science and Engi-
neering 136 (3). https://doi.org/10.1115/1.4026898.
Pelzer, Lukas, and Christian Hopmann. 2021. “Additive Manufacturing of Non-Planar
Layers with Variable Layer Height.” Additive Manufacturing 37: 101697.
https://doi.org/10.1016/j.addma.2020.101697.
Pottmann, Helmut, Alexander Schiftner, Pengbo Bo, Heinz Schmiedhofer, Wenping
Wang, Niccolo Baldassini, and Johannes Wallner. 2008. “Freeform Surfaces
from Single Curved Panels.” ACM Trans. Graph. 27 (3). New York, NY, USA:
18

Association for Computing Machinery: 1–10.

https://doi.org/10.1145/1360612.1360675.
Sarakinioti, Maria-Valentini, Thaleia Konstantinou, Michela Turrin, Martin Ten-
pierik, Roel Loonen, Marie L de Klijn-Chevalerias, and Ulrich Knaack. 2018.
“Development and Prototyping of an Integrated 3D-Printed Façade for Thermal
Regulation in Complex Geometries.” Journal of Facade Design and Engineer-
ing 6 (2): 29–40. https://doi.org/10.7480/jfde.2018.2.2081.
Schüller, Christian, Roi Poranne, and Olga Sorkine-Hornung. 2018. “Shape Represen-
tation by Zippables.” ACM Trans. Graph. 37 (4). New York, NY, USA: Associ-
ation for Computing Machinery. https://doi.org/10.1145/3197517.3201347.
Takezawa, Masahito, Takuma Imai, Kentaro Shida, and Takashi Maekawa. 2016.
“Fabrication of Freeform Objects by Principal Strips.” ACM Trans. Graph. 35
(6). New York, NY, USA: Association for Computing Machinery.
https://doi.org/10.1145/2980179.2982406.
Tam, Kam-Ming Mark, and Caitlin T Mueller. 2017. “Additive Manufacturing Along
Principal Stress Lines.” 3D Printing and Additive Manufacturing 4 (2): 63–81.
https://doi.org/10.1089/3dp.2017.0001.
Taseva, Yoana, Nik Eftekhar, Hyunchul Kwon, Matthias Leschok, and Benjamin Dil-
lenburger. 2020. “Large-Scale 3D Printing for Functionally-Graded Facade.” In
Anthropocene, Proceedings of the 25th International Conference of the Associ-
ation for Computer-Aided Architectural Design Research in Asia (CAADRIA)
2020, 1:183–92. Hong Kong.
Ulu, Erva, Emrullah Korkmaz, Kubilay Yay, O Burak Ozdoganlar, and Levent Burak
Kara. 2015. “Enhancing the Structural Performance of Additively Manufactured
Objects Through Build Orientation Optimization.” Journal of Mechanical De-
sign 137 (11). https://doi.org/10.1115/1.4030998.
Verhoeven, Floor, Amir Vaxman, Tim Hoffmann, and Olga Sorkine-Hornung. 2022.
“Dev2PQ: Planar Quadrilateral Strip Remeshing of Developable Surfaces.”
ACM Trans. Graph. 41 (3). New York, NY, USA: Association for Computing
Machinery. https://doi.org/10.1145/3510002.
Wu, Chenming, Chengkai Dai, Guoxin Fang, Yong-Jin Liu, and Charlie C L Wang.
2017. “RoboFDM: A Robotic System for Support-Free Fabrication Using
FDM.” In 2017 IEEE International Conference on Robotics and Automation
(ICRA), 1175–80. https://doi.org/10.1109/ICRA.2017.7989140.
Wu, Rundong, Huaishu Peng, François Guimbretière, and Steve Marschner. 2016.
“Printing Arbitrary Meshes with a 5DOF Wireframe Printer.” ACM Trans.
Graph. 35 (4). New York, NY, USA: Association for Computing Machinery.
https://doi.org/10.1145/2897824.2925966.
Zhang, Tianyu, Guoxin Fang, Yuming Huang, Neelotpal Dutta, Sylvain Lefebvre, Ze-
kai Murat Kilic, and Charlie C L Wang. 2022. “S3-Slicer: A General Slicing
Framework for Multi-Axis 3D Printing.” ACM Trans. Graph. 41 (6). New
York, NY, USA: Association for Computing Machinery.
https://doi.org/10.1145/3550454.3555516.