Communications of the ACM

Research highlights

Designing Plush Toys With a Computer

By Yuki Igarashi, Takeo Igarashi
Communications of the ACM, December 2009, Vol. 52 No. 12, Pages 81-88
10.1145/1610252.1610275
Comments

We introduce Plushie, an interactive system that allows nonprofessional users to design their own original plush toys. To design a plush toy, one needs to construct an appropriate two-dimensional (2D) pattern. However, it is difficult for nonprofessional users to appropriately design a 2D pattern. Some recent systems automatically generate a 2D pattern for a given three-dimensional (3D) model, but constructing a 3D model is itself a challenge. Furthermore, an arbitrary 3D model cannot necessarily be realized as a real plush toy, and the final sewn result can be very different from the original 3D model. We avoid this mismatch by constructing appropriate 2D patterns and applying simple physical simulation to it on the fly during 3D modeling. In this way, the model on the screen is always a good approximation of the final sewn result, which makes the design process much more efficient. We use a sketching interface for 3D modeling and also provide various editing operations tailored for plush-toy design. Internally, the system constructs a 2D cloth pattern in such a way that the simulation result matches the user's input stroke. We successfully demonstrated that nonprofessional users could design plush toys or balloon easily using Plushie.

Back to Top

1. Introduction

A computer can be a powerful tool for designing real-world objects. One can build a virtual three-dimensional (3D) model on a computer using computer-aided design (CAD) and use the model to run various simulations with computer-aided engineering (CAE) without the need to build or damage costly real objects. The benefits are evident in many areas from architecture to automobile design. However, these tools are mainly designed for professional users and are not particularly accessible to the ordinary person. The construction of a 3D model using a standard CAD system is tedious, and running a physical simulation using a standard CAE system requires a certain level of expertise.

Our goal is to bring the benefits of CAD and CAE to the hands of nonprofessional users including children. This article introduces our plush-toy design system,¹⁸ Plushie, as an example of our efforts to achieve this goal. Plush toys are familiar objects in our daily lives, but their design is difficult. One must design an appropriate two-dimensional (2D) pattern to obtain a particular 3D shape, but the relationship between the two is nontrivial, and intensive experience and knowledge are required to achieve satisfactory results. As a result, most people simply buy ready-made plush toys and do not enjoy the design and construction of their own. We have provided a way for people to design their own toys using a simple but powerful modeling tool that tightly integrates a sketching interface with physical simulation in the modeling process.

Plushie allows the user to design a plush toy from scratch by simple sketching operations.¹⁸ The user first draws the desired silhouette, and the system automatically generates a 3D plush-toy model and corresponding 2D cloth pattern. The user can also edit the model, e.g., cut it or add a part, using a simple sketching interface, and the 3D model and 2D cloth pattern are automatically updated. The 3D model is the result of a physical simulation that mimics the inflation of the sewn 2D cloth patch. Therefore, the model on the screen is always a good estimate of the final sewn result (Figure 1). When we ran workshops in a museum to have novice users try our system, we observed that even children could design their own plush toys.

We first give an overview of sketching interfaces for 3D modeling and previous efforts to enable end users to design physical objects. We then describe the user interface and implementation of the Plushie system, followed by results and user experiences. Finally, we conclude the article with some discussion of future work.

Back to Top

2. Sketching Interfaces for 3D Modeling

The sketching interface part of Plushie is an evolution of the Teddy system we presented in 1999.¹¹ That system allowed the user to create an interesting 3D model simply by sketching a silhouette of the target model (Figure 2, left). It was designed for the modeling of free-form rotund models, a task that is particularly difficult using standard modeling interfaces. Figure 3 shows an example of a modeling sequence using Teddy. The user's strokes are shown in red, and everything else is inferred and drawn by the system. The user first draws the silhouette of the base primitive, and the system generates the corresponding 3D geometry. The user then draws a stroke across the model and the system cuts the model at the line. The user can also add parts to the base model by drawing two strokes. Figure 2 right shows several 3D models created using the system.

Several sketching systems for free-form shapes were developed after Teddy. The original Teddy system used polygonal meshes, but some later systems experimented with other representations such as voxels²⁰ and implicit surfaces.² Some systems extended the interface to support subsequent editing by direct manipulation. ShapeShop²¹ represents a model as a collection of blob primitives and allows the user to move or scale each primitive. Fibermesh¹⁹ keeps the original stroke as a control curve on the model surface and allows the user to adjust the shape by deforming the curves. However, all these systems are designed for purely virtual 3D models without consideration of the physical properties of materials. Plushie is innovative in that it shows the feasibility of using a sketching interface for free-form shapes in the design of physical objects.

Back to Top

3. Designing Physical Models With A Computer

Another key aspect of our work is the tight integration of physical simulation into the 3D modeling process. In traditional applications, modeling and simulation are completely separate. A virtual model is created in 3D modeling software without considering any physical constraints, and it is then passed to a simulation environment. If the simulation result reveals a problem, the user returns to the model to fix the problem. We made this process more efficient by running the simulation concurrently with the modeling to create only models that are physically realizable. In this way, the user can more efficiently explore the design dimensions within realistic constraints. From the user's point of view, the model generated by the system may not correspond exactly to the shape that was input, but it will be a physically realizable shape reflecting the input shape.

Some recent systems have tried to incorporate fast physical simulation into an interactive design process. Igarashi and Hughes developed a mark-based interface for putting clothing on a virtual character,¹⁰ and Decaudin et al. proposed a system for designing an original garment via sketching.⁶ Both used simple geometric simulations to represent the physical properties of cloth material. Masry and Lipson described a system in which the user can quickly build a CAD model by sketching and immediately apply a finite element analysis to the model.¹⁵ However, the model construction is computed before the simulation in these systems, and no dynamic feedback loop exists between the simulation result and the original user input.

Several efforts have also been made to support the design of physical objects by end users in the computer graphics research community. Mitani and Suzuki¹⁶ and Shatz et al.²² presented automatic segmentation of a 3D model into surface patches that can be perfectly flattened onto a plane without distortion for constructing paper craft models. Similarly, Julius et al. proposed similar method for plush toys¹⁴ allowing small distortion. Pillow system¹⁷ facilitates the manual segmentation of a model by providing automatic flattening and by showing the result of physical simulation. These systems make plush-toy design more accessible, but the fundamental challenge of creating an original plush toy is still unresolved.

Back to Top

4. The Plushie System

The system consists of two windows: one shows the 3D plush-toy model being constructed and the other shows the corresponding 2D pattern (Figure 4). The user works on the 3D view, interactively building the 3D model by using a sketching interface. The 2D view is mainly for reference but the user can also edit the 2D pattern directly when desired. The 3D model is produced from a physical simulation of the assembled 2D pattern. After each input from the user, the system updates the 2D pattern so that the simulation result matches the user input. This guarantees that the model is always realizable as a real plush toy and that the 2D pattern is readily usable as a template for cutting and sewing real fabric.

The modeling operations are based on the Teddy system.¹¹ The user interactively draws free-form strokes on the canvas as gestures and the system automatically generates a 3D model and corresponding 2D cloth pattern. We also provide some special editing operations tailored for plush-toy design.

Creating a New Model: Starting with a blank canvas, the user creates a new plush-toy model by drawing its silhouette as a closed free-form stroke. The system automatically generates two cloth patches corresponding to the stroke and visualizes the shape of the resulting plush toy by applying a simple physical simulation (Figure 5).

Cut: The user can cut the model by drawing a stroke that starts outside of the model, crosses it, and ends outside of the model (Figure 6). The model is cut at the intersection and flat patch is generated at the cross-section. This operation is useful for creating relatively flat surfaces, such as those in a foot or belly.

Creation of a Part: The user can add protruding parts such as the ears and arms to the base model by drawing a single stroke that defines the silhouette of the part. The stroke should start and end on the base model (Figure 7a). The system generates two candidate shapes and presents them to the user as suggestions⁹ (Figure 7b). One is for fat, rounded parts like the body, arm, and leg (Figure 7c). Their base is connected to the base model with an open hole. The other candidate shape is for thin parts like ears and the tail whose base is closed (Figure 7d). The user clicks the desired thumbnail and the system updates the main model accordingly. We found that the ability to create thin parts with a single stroke is particularly useful. They are frequently seen in real toys and are difficult to design using standard modeling software. Figure 20 shows a couple of example models with thin parts.

Pull: The user can grab a seam line and pull it to modify the shape. For example, the user can pull an ear to make it larger when it is smaller than the other (Figure 8). The pulling operation begins when the user starts dragging on the background region near a seam line. The system changes the mouse cursor when it approaches a seam line to indicate that the user can start pulling. We use the peeling interface introduced by Igarashi et al.¹² to adjust the size of the region to be deformed, that is, the larger area is deformed as the user pulls more. The system continuously updates the 2D cloth pattern during pulling and shows the simulation result in the 3D view.

Insertion and Deletion of Seam Lines: The modeling operations performed thus far automatically generate 2D patches according to predefined algorithms and seam lines (patch boundaries) appear on the 3D model surface without the user's explicit control. However, it is sometimes desirable for knowledgeable users to design seam lines manually, for more detailed control. This is especially important when using nonstretchy cloth as in balloon models because one needs to divide a rounded surface into many almost-developable small patches (Figure 19 bottom).

The user can add a new seam in the seam line drawing mode by drawing a free-form stroke on the model surface (Figure 9). The corresponding cloth patch is then automatically cut along the new seam line. If the stroke crosses the entire patch, the patch is divided into two separate patches. If the stroke starts or ends in the middle of a patch, it becomes a dart. The 3D geometry does not change immediately after the insertion of these seam lines, but the user can pull the seam line afterwards to modify the shape. This operation is very useful for creating a salient feature in the middle of a flat patch. Deletion is achieved by tracing the target seam line in the erasing mode. This merges the separated patches together and thus flattens the area (Figure 10).

Operations on the 2D Pattern View: The 2D pattern view is mainly used to preview the pattern to be printed for sewing, but it also works as an interface for advanced users to edit the pattern directly. The preview helps the user to understand the relationship between the 3D model and 2D patches and gives a sense of the labor required for assembling the patches. The system can display how patches are connected by showing connectors or paired numbers (Figure 11). Connectors are useful for understanding the relationship on the screen and numbers are useful as a printed reference on each patch. The system provides an automatic layout and manual arrangement interface for preparing the final pattern to be printed.

The system also allows the user to edit the patches directly by using the pulling interface. The user can grab the boundary of a patch and pull it to deform the shape.¹² We again use a peeling interface to adjust the size of area to be deformed. The effect of 2D deformation immediately appears in the 3D view because of the physical simulation. The ability to deform an individual patch is useful for designing asymmetric shapes such as a penguin belly (Figure 12). The pull operation is also useful for opening a dart line to make a flat patch swell more (Figure 13).

Back to Top

5. Implementation

This section briefly describes the implementation of Plushie. A more detailed description is found in our original paper.¹⁸ We use a standard triangle mesh for the representation of the 3D model and 2D patches. We use a relatively coarse mesh (1000–2000 vertices) to achieve interactive performance. Each vertex, edge, and face of the 3D mesh is associated with corresponding entities in the 2D mesh. A 3D mesh is always given as a result of applying a physical simulation to the assembled 2D pattern. To be more precise, the physical simulation applied to the 3D mesh is governed by the rest length of each edge and the rest length is defined in the 2D mesh. For each modeling operation, the system constructs the initial 2D patches and the 3D geometry corresponding to the input stroke, and then runs a physical simulation to update the 3D geometry. The system then adjusts the patch shape so that the simulation results match the input strokes.

 5.1. Physical simulation

We use a simple static method for the physical simulation. We examined other, more elaborate methods, such as finite element methods,⁸ dynamic simulation,⁵ and energy minimization,⁴ but we found that the simple approach is best suited for our purpose. It is easy to implement, fast enough for interactive modeling, and sufficiently robust for dealing with adverse user operations. More importantly, it produces a reasonable estimation of the resulting plush-toy shape. As shown in Figure 14, it successfully reproduces characteristic behaviors seen in the stuffed cloth. This algorithm is also used in a garment capture system.³

In each simulation cycle, the system first moves each face slightly in its normal direction to mimic the effect of internal pressure (Figure 14a). The system then adjusts the length of each edge to preserve the integrity of the cloth material⁵ (Figure 14b). We decided to prevent stretching only and tolerate compression because plush toys' rotund shape is generated from compression (small wrinkles) along the seam lines. The second part (adjustment of edge length) runs 10 times in each cycle to prevent excessive stretch. It takes ~30 simulation cycles (2 s) to converge in our typical examples. Although it is possible to show the result only after the convergence, we decided to show the intermediate shape because test users preferred to see the inflation process.

 5.2. 3D modeling operations

Creating a New Model: The input stroke is projected onto an invisible plane at the center of the world facing the screen, and the system generates an initial two-sided mesh inside of the closed region. Each side of the mesh is used directly as a 2D patch for the model. The system then applies the physical simulation to the mesh. It inflates the mesh to the direction perpendicular to the viewing direction, but its silhouette actually becomes smaller as it inflates (Figure 15). The system waits until the simulation converges and then starts to adjust the 2D pattern so that the simulation result matches the input stroke. Specifically, the system calculates the distance d_i from a vertex v_i of the 3D mesh along the seam line to the corresponding point p_i in the projected input stroke along its normal direction, and moves the corresponding 2D vertex u_i on the patch boundary to its normal direction by that amount d_i. After modifying the patch boundary, the system updates the 2D mesh so that vertices are uniformly distributed inside of the patch. The length of the edges in the updated 2D mesh is then used as the new rest length in the simulation. The system repeats this adjustment process and the physical simulation until convergence. It takes approximately 20 iterations (2 seconds) to converge in our typical examples.

This simple algorithm works well in practice for our application. Figure 16 shows some examples in which our algorithm successfully found appropriate 2D patches that yielded the desired 3D shapes. In some situations, the input shape is not realizable as a plush-toy model consisting of two patches. For example, a sharp concavity is not realizable without causing self-intersection in the 2D patch. In these cases, the system terminates the optimization process, leaving a gap between the input stroke and the 3D model. This indicates that the desired shape is not possible with two patches. The user must add additional seam lines to obtain more control.

Cut: The system constructs a curved surface by sweeping the cutting stroke on the screen to the viewing direction and dividing the mesh along the surface. The right-hand side of the surface is removed and a new mesh is created on the cross-section. The cross-section is always developable, so the system simply flattens it and uses it as a 2D patch. The system then applies the inflation process to the model. Note that the silhouette of the inflated 3D model does not exactly matches with the input cut stroke because we do not apply any adjustment as in the initial creation.

Creation of a Part: The system first projects the two endpoints of the input stroke onto the base model surface. A plane that passes through these 3D points and is facing toward the screen is constructed and the input stroke is projected onto it. The system then draws an ellipse on the model surface for constructing a fat part and draws a line for creating a thin part (Figure 17). The ellipse or the line (what we call base curves) is also projected to the plane. The system generates a 2D mesh on the projection plane in the area enclosed by the projected input stroke and the projected base curve. The 2D mesh is duplicated and serves as 2D pattern and as the initial 3D geometry for the added part. As in the initial model creation case, the flat two-sided 3D mesh is inflated by physical simulation. The silhouette of the added part gradually shrinks and the system enlarges the 2D pattern so that the silhouette matches the input stroke as in initial creation.

In case of a part with an elliptic base curve, the system cuts open the base surface and stitches it with the newly created mesh. The result is a single connected mesh, and physical simulation is applied uniformly to the entire mesh. On the other hand, the system does not open the base mesh in case of the linear base curve. The new part is created as an independent closed mesh and the simulation is applied separately to the base mesh and the new part. The base mesh inflates independently of the part mesh, and the base curve is treated as a positional constraint in the simulation of the part mesh (we simply do not move these vertices in the simulation cycle).

Pull: The pull operation is a bit involved because the system cannot directly modify the 3D mesh and must do so indirectly by deforming the corresponding 2D pattern. As the user starts pulling a vertex on a seam line, the system first constructs a projection plane that passes through the seam line (Figure 18). The mouse cursor position on the screen is projected onto the projection plane, and it is used as a target position for the pulled vertex during subsequent dragging. The system computes the displacement δ_i in the local coordinate frame on the projected plane from the original position v_i to the target position h_i, and moves the corresponding vertices u_i⁰ and u_i¹ in the 2D mesh in their local coordinate frames by that amount δ_i. These 3D and 2D coordinate frames are defined by the pulled vertex's normal vector and the direction of the seam line. The system iterates this displacement process with physical simulation until it converges. To achieve a smooth deformation, the system also moves the surrounding vertices in the 2D mesh using the curve manipulation method introduced in Igarashi et al.¹² It enlarges the region to be deformed proportional to the displacement of the pulled vertex.

Insertion and Deletion of Seam Lines: Insertion of a new seam line is straightforward. The system simply cuts the patch along the added seam line, and basically does not change the result of simulation. Deletion is more complicated because the merged patch is not necessarily developable. The system applies an approximate flattening operation²³ to the merged 3D surface to obtain the geometry of the new 2D patch.

Back to Top

6. Results

Plushie is implemented as a Java™ program. Construction of 2D patterns and a physical simulation run in real-time on a 1.1 GHz Pentium M PC. We designed a couple of plush toys using our system and created a real toy based on the printed pattern. A modeling session typically takes 10–20 min and sewing takes 3–5 h. Figure 19 shows a plush toy and balloon model designed in our system. It shows that the physical simulation successfully captures the overall shape of the real objects. We interviewed with professional balloon designers and they supported our system, saying that it can significantly reduce the time necessary for designing original balloon.

The user can assign different textures to individual patches (Figure 20). Therefore the user can explore various design possibilities before actuary sewing the real fabric (such as Figure 20 right). These models also demonstrate the effectiveness of thin parts.

We ran four small workshops to test the usability of the system and found that novice users, mainly children, can successfully create original plush toys using our system. Here is an observation from one of these workshops. Nine children, approximately 10–14 years old, joined the workshop accompanied by their parents. We gave a brief tutorial at the beginning and had them design their own plush toys using the system. It took about an hour for the design. They then printed the designed pattern and sewed a real toy in ~3 h. Figure 21 shows a couple of plush toys created in the workshop. Participants quickly learned how to use the system and successfully designed the 3D models they wanted, with some help from volunteers. Furthermore, they enjoyed the process. These toys were their own creations and one-of-a-kind designs. Participants also gave us valuable feedback for future improvement. They wanted to have some auxiliary functions such as the ability to design symmetric parts and remove existing parts, but no one complained about the quality of the visual simulation. A perfectly accurate simulation is not necessary because many small variations inevitably occur during the real sewing and stuffing process.

Back to Top

7. Conclusion

We introduced a plush-toy design system as an example of our efforts to make CAD and CAE accessible to end users. The system allows the user to design a plush toy quickly and simply by combining simple sketching operations. The user draws the desired silhouette on the canvas, and the system automatically generates a 3D plush-toy model and a 2D cloth pattern. The system runs a simple physical simulation in the background so that the resulting 3D model is always a good estimate of the final sewn result. The user can construct a real plush toy by printing the pattern and sewing the resulting pieces together.

To demonstrate the effectiveness of the approach even further, we also developed a system for the design of knitted toys.¹³ A knitted toy is a toy made of knitted yarn instead of cloth patches. One can construct a knitted toy by knitting according to a specific knitting pattern, but it is difficult to produce an appropriate knitting pattern for a desired 3D shape. The Knitty system allows the user to design an original knitted toy by simply drawing the desired silhouette (Figure 22). The system then generates a 3D knitted animal model and corresponding knitting diagram. We ran a workshop using this system and observed that children could use it to design their own knitted animals.

Interactive 3D modeling assisted by concurrent physical simulation can be a powerful tool in many other application domains. For example, if one could run an aerodynamic simulation during the interactive design of a paper airplane model, the entire geometry could be adjusted in an intelligent manner in response to the user's simple deformation operations to produce a model that could actually fly. This kind of interaction would make it easier for designers to pursue aesthetic goals while satisfying engineering constraints. Real-time simulation does require high-performance computing resources, but some meaningful support should be possible by carefully limiting the target task and designing appropriate interfaces as shown in this paper. We hope that our work inspires more work in this direction.

Back to Top

Back to Top

Back to Top

Back to Top

Back to top

The Digital Library is published by the Association for Computing Machinery. Copyright © 2009 ACM, Inc.

---

No entries found