In this project, significant progress was achieved in the development of a dedicated soft body solver in Blender. As part of the project a volumetric remesher was implemented in the Blender code-base. The tetrahedralization also gives the option for addition of internal points to allow for more uniform tetrahedralization and to represent the object in 3 dimensions more accurately which can be controlled using Internal Resolution parameter. Additionally a basic flow for soft body simulations using Extended Position Based Dynamics (XPBD) was implemented including distance and volume constraint. The stiffness of the object can be controlled using compliance which is implemented as inverse of stiffness.

While there are still a few advancements required to deploy the solver to the main branch, this project sets the first stone to a dedicated soft body solver using XPBD.

Volumetric Remeshing

Before simulations can be performed, the surface meshes need to be converted to volumetric meshes, in particular tetrahedral mesh. To perform this conversion, a modified version of Incremental Delaunay Tetrahedralization was implemented. Since Blender does not have data structures to store and work with volumetric meshes, the tetrahedron information was stored as quadrilateral planes. This assumption works well since all vertices of a tetrahedron are connected, thus eliminating the requirement to store edge information for each tetrahedron primitive.

Delaunay method works by adding vertices incrementally and briefly follows the steps:
1) Find tetrahedron containing the point
2) Find all tetrahedrons that violate Delaunay Condition
3) Delete violating tetrahedrons and make new tetrahedrons centered at the new point
4) Update the neighbors

We start by defining the Delaunay condition for a tetrahedron and an external point. The Delaunay condition is satisfied for this pair if the point lies outside the circum-sphere of the tetrahedron.

For the first step to find the tetrahedron that contains the new point to be added, a ray is cast from center of the current tetrahedron to the point. The next tetrahedron to be tested is the neighbor of current tetrahedron that shares the face intersected by the cast ray.

Once the containing tetrahedron is found, we recursively check neighbors of all violating tetrahedrons whether they satisfy the Delaunay Condition and add the ones that don’t to the stack to subsequently check their neighbors. In addition, all tetrahedrons that are tested for satisfying the condition are marked.

Once all the violating tetrahedrons are found, we need to take all boundary faces - outermost faces or more technically, faces that either don’t have a neighboring tetrahedron or that are common to violating and non-violating tetrahedrons. Finally new tetrahedra are constructed by joining boundary faces with the new point.

Finally the neighbors for all the new tetrahedrons are updated. This can be divided into 2 parts, updating information around boundary faces and updating information for internal faces. For boundary, the non-violating tetrahedron’s neighbor is set to the new tetrahedron formed using the boundary face. Same is done for the violating tetrahedron. For the internal faces, we make use of the fact that each face is shared by only 2 tetrahedrons and can be uniquely represented using it’s 3 vertices.

Since the algorithm makes tetrahedrons using existing vertices, to allow for more consistent and uniform tetrahedralization, the option of adding points inside the structure has also been added in the form of interior resolution. Additionally to visualize the tetrahedralization, the option to display all faces individually was also given where each tetrahedron is scaled down so that it can be comfortably analyzed for correctness.

Suzanne tetrahedralized and represented with individual tetrahedrons scaled down to a factor of 0.8

Extended Position Based Dynamics (XPBD)

To perform the actual simulations, Extended Position Based Dynamics(XPBD) was implemented. As opposed to force based methods that calculate the amount of force on vertices and determine the motion of individual vertices, position based methods move the vertices in the direction of their current velocity and then enforce hard constraints as necessary.

Since the current soft-body solver was just a cloth simulation solver with high surface tension, the data structures had to be changed completely to store tetrahedron information. New data-structures BodyTet and BodyEdge were defined while the existing data-structures SoftBody and BodyPoint were completely rewritten. This required subsequent changes to be made DNA and RNA code for soft body.

The various intrinsic properties of an object are implemented in the form of constraint functions in XPBD. Keeping that in mind, basic constraints like distance conservation constraint (distance between connected vertices is preserved) and volume conservation constraint (volume of individual tetrahedrons must be preserved) were implemented. As a general rule these constraints must ensure that the linear momentum and angular momentum of the object is conserved.

To ensure that the center of mass of the object is in the middle of the object and the density is uniform in the body, the mass of vertices has been set as sum of inverse of volume of tetrahedrons that they are a part of. Because of this, the mass of vertices is dependent on the resolution of tetrahedralization and eventual volume of individual tetrahedrons.

The stiffness of the object can be controlled using the edge and volume compliance values which is inversely proportional to the stiffness. Thus a compliance value of 0 gives the object infinite stiffness causing the object to behave like a rigid body. An important thing to note here is how the masses for vertices have been set. The compliance values for specific softness can vary in magnitude of order from object to object or for that matter even same object with different internal resolutions. The reason being that the compliance values are closely tied to the mass of vertices.

Tetrahedralized cube bouncing on imaginzary x-y plane

Future Work

1) Make a dedicated modifier for tetrahedralization: The current code for tetrahedralization was added to a random modifier file ( since it was supposed to be a part of simulator and this allowed for easier debugging but it turned out to be a request by many and generally since the code is in place, a separate modifier for the same only helps

2) Shift tetrahedralization code : Shift tetrahedralization to part where mesh is converted to soft-body data-structures. While this is contradicting to the first point, this will add the layer of abstraction where the users don’t see the individual tetrahedrons stored as quads which distort the view and don’t allow for practical usage of the simulator since there’s not proper way to go back to the original object once the modifier has been applied. The first point will simply allow users to make use of the tetrahedralization whenever required outside of simulations.

3) Implement self collisions and collision : This is fairly self explanatory, a part of the project scope and undoubtedly one of the most important parts of any simulation engines is collision and self collision that could not be completed due to various challenges.

4) Vertex groups : Add vertex group options that can be used for multiple purposes including pinning vertices and changing mass of certain regions of the soft-body

5) Cloth simulation : Make option for performing cloth simulation using the current solver. This would require additional bending constraint to be implemented. The bending constraint would require the initial angle between neighboring faces to be saved which is not a part of the current data structures defined. An easy fix to this would be to store the initial normal vector of all planes and use them to calculate the initial angle in subsequent steps.

6) Refactor soft body code : Define new flow for soft-body data-structure. Currently the soft-body data-structure is present inside the object data-structure. This makes the code less readable and a little more convoluted to navigate around. Would prefer structuring the data like it’s structured for cloth right now.


Thank you so much Sebastian Parborg for being my rubber duck. You were always extremely helpful and patient and I really look forward to working a lot more with you. I would also like to thank Matthias Müller whose work have been implemented throughout the project.