Creating a soft object that squashes or stretches is one challenge; allowing it to be cut, peeled, fractured, and freely manipulated in real time is considerably harder. Mesh-based simulation methods often require expensive remeshing when an object’s topology changes, while physically rigorous approaches may depend on global solvers that are difficult to run at interactive frame rates.

Developed by Korea Advanced Institute of Science and Technology (KAIST) researchers Jewoo Shin and Jinah Park, OmniBody replaces a conventional connected mesh with voxel-aligned, oriented ellipsoidal particles. Local Extended Position-Based Dynamics constraints control contact, bending, elasticity, volume preservation, and fracture, allowing global behavior to emerge without matrix assembly or linear system solving. The project was presented as part of SIGGRAPH 2026’s Animation & Simulation posters.

The resulting GPU framework can simulate a 64,000-particle object in approximately 3.1 milliseconds on an NVIDIA GeForce RTX 3080 Ti, while a demanding cutting test remained below 10 milliseconds. We spoke with Shin about designing materials through modular constraints, breaking objects without remeshing, the remaining challenges surrounding exposed interior surfaces, and why freely deformable objects could be especially valuable for XR and sandbox games.

Could you introduce yourselves, your research backgrounds, and the simulation challenges that motivated you to develop OmniBody?

Jewoo Shin: My name is Jewoo Shin, and I’m currently a master’s student at Korea Advanced Institute of Science and Technology (KAIST), working under the supervision of Professor Jinah Park. I don’t have an extensive research background yet, but I have a deep passion for computer graphics and physics simulation. I majored in computer science as an undergraduate, and OmniBody is the first project I’ve brought to a meaningful result as a graduate student.

The main challenges that motivated me to develop this simulation model were the lack of accessibility and the complexity of implementation. The conventional methods for softbody simulation require sophisticated domain knowledge to set up and control, and achieving a desired physical state is rarely straightforward. More specifically, I was motivated by the goal of building a real-time interactive simulation that goes beyond the simple squishy bodies commonly seen in many environments — one that supports more realistic physical behaviors and interactions such as cutting and peeling.

Furthermore, particle-level deformation plays a central role in interactive simulation techniques. In the research project I have been working on, there was a need for a deformable body model capable of handling various material behaviors. It was crucial to handle both global deformation for visual coherence and local response at contact points.

Therefore, OmniBody was built on Extended Position-Based Dynamics(XPBD) chosen for its simplicity and stability, and demonstrates that a wide variety of interesting scenarios can be produced through the composition of simple constraints.

OmniBody builds global deformation from entirely local interactions between voxel-aligned ellipsoidal particles. How does this representation work, and what advantages does it offer over tetrahedral meshes, conventional particle systems, or finite-element methods?

Jewoo Shin: The oriented ellipsoidal particle representation carries more information than conventional point-mass particles, which enables a complex and distributed XPBD constraint configuration. This is conceptually closer to finite element methods than to conventional particle systems, in that the body is subdivided into explicit elements and each particle acts as an individual element whose properties are computed locally. However, unlike FEM, it requires no expensive operations such as matrix assembly or linear system solving. Instead, local constraints are resolved and connected across neighboring particles, keeping the computation lightweight while maintaining physical coherence.

Unlike tetrahedral or hexahedral meshes, ellipsoidal particles have significantly less overhead when the topology changes, such as during destruction or cutting. Mesh-based methods typically require vertex duplication and remeshing in these cases, which can be prohibitively costly in interactive real-time scenarios. OmniBody addresses this by simply deactivating the contact constraint between affected particles, and this operation is easily parallelizable.

Another advantage of our representation lies in collision handling. Since the particles themselves serve as the collision boundary, detecting and resolving collisions with other OmniBody objects, particle-based materials, or the body itself is far more straightforward than collision handling between meshes.

The framework separates contact, bending, anisotropic elasticity, volume preservation, and fracture into individual constraints. How are these constraints combined without producing unstable or contradictory behavior, and what controls can artists use to define different materials?

Jewoo Shin: These constraints are categorized into two levels: particle-wise constraints and neighbor-level constraints. The particle-wise constraints consist of anisotropic elasticity and volume preservation, which determine how the shape of each ellipsoid deforms under stress and how it restores to its rest shape. The stiffness parameters of elasticity control the sensitivity of strain along each axis, while the compliance of volume preservation indicates the object's tolerance to volumetric stretch and compression.

The neighbor-level constraints, on the other hand, determine which particles are connected to one another and how closely they are held together. The contact constraint enforces geometric attachment between particles by restricting the distance between their surfaces to zero, while the bending constraint prevents adjacent particles from rotating too freely relative to each other, similar to ball joint.

Putting these together, each solver step iterates through the constraints in a sequential manner to produce global behavior. Following the predict-project-update scheme of position-based dynamics, the solver first resolves positional constraints and predicts the scale and orientation of each ellipsoid. The particle-wise constraints then determine the shape of each particle, and once all particles have settled into their configurations, the fracture constraint may deactivate connections between adjacent particles if they have been pulled too far apart. Since each step is handled through substepping, which divides each timestep into shorter intervals, the system converges slowly but steadily toward equilibrium, as long as there is no severe misalignment between particles.

Thanks to the constraint-based formulation, each constraint is easily decoupled and tuned through its compliance parameter. All artists need to do is select which constraints act on an object and set the stiffness of each, allowing material behavior to be adjusted at runtime. For example, increasing the compliance of the contact constraint produces a more fluid, jelly-like behavior, while a higher compliance for volume preservation makes the object more stretchable.

Destruction commonly requires remeshing or explicitly tracking cracks. How does OmniBody’s distance-based fracture model allow objects to tear and break without remeshing, and how well can it reproduce specific fracture patterns rather than general fragmentation?

Jewoo Shin: Our distance-based fracture model offers a simple yet effective peeling feature for soft bodies. The distance between particle surfaces is ideally kept at zero at all times, but if it exceeds a certain threshold, the constraint is permanently decoupled, allowing the body to separate. Importantly, this threshold can vary along each axis, which enables more varied scenarios. For example, a piece of meat that tears more easily along a particular direction can be naturally represented.

Another form of destruction is cutting, which works by directly disabling contact constraints. A virtual cutting object, such as a plane, disconnects any constraint it intersects as it passes through the geometry. The result is similar to cutting a spring in a voxelized jelly block.

That said, reproducing specific fracture patterns is currently less reliable than achieving general fragmentation. Since all particles are treated as having equal mass, they tend to respond uniformly to external forces. This means that a pinching interaction, for instance, only detaches the particles nearest to the contact point rather than propagating inward in a controlled way. Addressing this would require treating the inverse mass of each particle differently to steer fracture along intended paths, which likely calls for a differentiable approach. We leave this as a direction for future work.

What performance did you achieve across different object resolutions, particle counts, and destruction scenarios? Which parts of the simulation benefit most from GPU parallelization, and where do the current computational bottlenecks appear?

Jewoo Shin: For a standard cubic grid, the per-frame simulation time scales quite well with particle count. On an NVIDIA RTX 3080 Ti, a 20³ grid with 8,000 particles runs at 0.9ms per frame, and even a 40³ grid with 64,000 particles stays within 3.1ms. For a more complex scenario such as cutting a stretched 40³ cube, the peak frame time remained under 10ms, which we consider sufficient for real-time interactive use.

The key to this performance lies in how the voxelized neighbor structure naturally decomposes constraints into three independent sets along the x, y, and z axes. This allows the Gauss-Seidel solve to be completed in just three passes, reducing the most computationally expensive part of the simulation approach to near-linear complexity, which maps well to GPU parallelization.

Even so, there are still areas we are actively working to improve. Since OmniBody is still under active development, we have intentionally kept each constraint modular and easy to modify, which means the GPU currently processes them as a large number of small sequential operations. Because many of these operations access the same data, the hardware has to wait for each one to finish before starting the next, which adds up to a meaningful source of overhead. Similarly, the memory layout per particle is currently designed with flexibility in mind rather than being fully optimized. As the system matures and the constraint design stabilizes, consolidating these operations and tightening the memory layout will be our primary focus for optimization.

What limitations or visual artifacts remain, and what would be required to integrate OmniBody into a production game engine for effects such as deformable environments, destructible creatures, cloth-like objects, or soft-body gameplay?

Jewoo Shin: There are still limitations on both the simulation and visual representation sides. On the simulation side, OmniBody's constraints are not derived from strictly physically based equations or energy formulations, so its behavior can sometimes differ from what we would expect from a physically accurate simulation. For example, objects cannot swing repeatedly like a pendulum because the current formulation does not include angular terms.

On the visual side, rendering currently relies either on directly displaying individual particles or on skinning a mesh to follow the particle positions. This works well for many deformable objects, but it becomes more challenging for fully destructible models. When an object is cut or broken apart, the newly exposed interior surfaces are not explicitly defined in a way that can be rendered cleanly. Volumetric representations, such as texture-based representations or Gaussian splats, could be a better way to visualize these newly exposed regions.

There are also several engineering challenges that would need to be addressed before OmniBody could be integrated into a production game engine. Game environments typically contain many dynamic objects, so the simulation needs to remain stable and efficient across a wide range of interactions. Because OmniBody is still an approximation of real physical behavior, unexpected behavior can occur in certain situations. Robust interactions between multiple objects, as well as further optimization, would also be necessary.

Jewoo Shin: At the same time, I don't think OmniBody needs to replace existing solutions for every use case. If fully destructible soft bodies are not essential, well-established mesh-based softbody or cloth simulation methods would likely provide better performance. For many types of game content, predefined fragmentation systems or mesh-cutting techniques are already sufficient. Where I think OmniBody becomes particularly interesting is in highly interactive environments such as XR, where users can manipulate objects freely with their hands rather than being limited to predefined or scripted interactions.

As someone who genuinely loves games, I believe that simulations that behave more like the real world can create richer and more immersive experiences. Giving players the freedom to interact with objects in unexpected ways, rather than limiting them to predefined actions, can open up new possibilities for creative and engaging gameplay.

I've always been drawn to simulation and sandbox games, and I've long imagined virtual environments where objects could be explored and manipulated much as they are in the real world. Imagine being able to actually knead dough, slice bread and cook a meal in a virtual kitchen. That is the kind of experience I find most exciting, and it is ultimately what motivates this work.

Jewoo Shin, Computer Graphics MS at KAIST