XPBD: Position-Based Simulation of Compliant Constrained Dynamics

Miles Macklin Matthias Müller Nuttapong Chentanez

NVIDIA

Figure 1: In this example, we see the effect of changing the relative stiffness of volume conservation and stretch and shear constraints on a
deformable body. Unlike traditional PBD, our method allows users to control the stiffness of deformable bodies in a time step and iteration
count independent manner, greatly simplifying asset creation.

Abstract simulation [Bender et al. 2014b]. Specifically, constraints become

arbitrarily stiff as the iteration count increases, or as the time step
We address the long-standing problem of iteration count and decreases. This coupling of parameters is particularly problematic
time step dependent constraint stiffness in position-based dynam- when creating scenes with a variety of material types, e.g.: soft
ics (PBD). We introduce a simple extension to PBD that allows it to bodies interacting with nearly rigid bodies. In this scenario, raising
accurately and efficiently simulate arbitrary elastic and dissipative iteration counts to obtain stiffness on one object may inadvertently
energy potentials in an implicit manner. In addition, our method change the behavior of all other objects in the simulation. This of-
provides constraint force estimates, making it applicable to a wider ten requires stiffness coefficients to be re-tuned globally, making
range of applications, such those requiring haptic user-feedback. the creation of reusable simulation assets extremely difficult. Itera-
We compare our algorithm to more expensive non-linear solvers tion count dependence is also a problem even in the case of a single
and find it produces visually similar results while maintaining the asset, for example, setting the relative stiffness of stretch and bend-
simplicity and robustness of the PBD method. ing constraints in a cloth model. To make matters worse, the effects
of iteration count are non-linear, making it difficult to intuitively
Keywords: physics simulation, constrained dynamics, position adjust parameters, or to simply rescale stiffness values as a simple
based dynamics function of iteration count.

Concepts: •Computing methodologies → Real-time simulation; The recent resurgence in virtual-reality has given rise to the need for
Interactive simulation; higher fidelity and more physically representative real-time simula-
tions. At the same time, the wide-spread use of haptic feedback
devices require methods than can provide accurate force estimates.
1 Introduction PBD does not have a well defined concept of constraint force, and
as such it has mostly been limited to applications where accuracy
Position-Based Dynamics [Müller et al. 2007] is a popular method is less important than speed, and where simulations are secondary
for the real-time simulation of deformable bodies in games and in- effects.
teractive applications. The method is particularly attractive for its
simplicity and robustness, and has recently found popularity out- In this paper we present our extended position-based dynamics
side of games, in film and medical simulation applications. (XPBD) algorithm. Our method addresses the problems of iteration
and time step dependent stiffness by introducing a new constraint
As its popularity has increased, the limitations of PBD have be- formulation that corresponds to a well-defined concept of elastic
come more problematic. One well known limitation is that PBD’s potential energy. We derive our method from an implicit time dis-
behavior is dependent on the time step and iteration count of the cretization that introduces the concept of a total Lagrange multiplier
to PBD. This provides constraint force estimates that can be used
Permission to make digital or hard copies of all or part of this work for per- to drive force dependent effects and devices.
sonal or classroom use is granted without fee provided that copies are not
made or distributed for profit or commercial advantage and that copies bear To summarize, our main contributions are:
this notice and the full citation on the first page. Copyrights for components • Extending PBD constraints to have a direct correspondence to
of this work owned by others than the author(s) must be honored. Abstract-
well-defined elastic and dissipation energy potentials.
ing with credit is permitted. To copy otherwise, or republish, to post on
servers or to redistribute to lists, requires prior specific permission and/or • Introducing the concept of a total Lagrange multiplier to PBD
a fee. Request permissions from [email protected]. © 2016 Copyright allowing us to solve constraints in a time step and iteration
held by the owner/author(s). Publication rights licensed to ACM. count independent manner.
MiG ’16,, October 10 - 12, 2016, Burlingame, CA, USA
ISBN: 978-1-4503-4592-7/16/10 • Validation of our algorithm against a reference implicit time
DOI: http://dx.doi.org/10.1145/2994258.2994272 stepping scheme based on a non-linear Newton solver.

49
2 Related Work 3 Background
We briefly describe the core of the position-based dynamics algo-
There is a large body of work in computer graphics that attempts
rithm and refer to the survey paper by Bender et al. [2014b] for
to solve constrained dynamics simulations [Nealen et al. 2005].
further detail. PBD can be thought of as a semi-implicit integra-
Our method builds on the position-based dynamics (PBD) algo-
tion step using the Stömer-Verlet method, followed by a number of
rithm presented by Müller et al. [2007], which solves constraints at
constraint projection steps. The projection step is performed using
the position-level in an iterative Gauss-Seidel fashion. Stam [2009]
local linearizations of each constraint function with mass weighted
proposed a similar Gauss-Seidel constraint solver that operates on
corrections. The main step in PBD’s constraint solver is the calcu-
velocities. These approaches operate on constraints locally and are
lation of the per-constraint position delta, given by:
popular due to their efficiency and ease of implementation, however
both methods suffer from iteration count stiffness dependence.
∆x = kj sj M−1 ∇Cj (xi ). (1)
Many authors have applied PBD to deformable object simulation.
Recently, Bender et al. [2014a] proposed a continuum-based for- Here the subscript i denotes the iteration index, j is the constraint
mulation that treats strain-energy as a constraint function directly. index, and k ∈ [0, 1] is the constraint stiffness applied simply as
Concurrently, Müller et al. [2014] proposed constraining the entries a multiplier on each constraint correction. The scaling factor, s, is
of the Green-St Venant strain tensor to allow controlling strain in di- given by the following equation, derived from a single Newton step
rections independent from the model discretization. Shape match- of the constraint function:
ing [Müller et al. 2005] introduces a geometric constraint for the
simulation of deformable objects amenable to PBD, but again, all −Cj (xi )
these methods suffer from a strong dependency of stiffness on iter- sj = . (2)
ation count. In contrast, our compliant constraint formulation has ∇Cj M−1 ∇CjT
a direct correspondence to tradtional constitutive models, and con-
verges to a well-defined solution. An undesirable side-effect of simply scaling the position change by
k is that the effective constraint stiffness is now dependent on both
the time step and the number of constraint projections performed.
Servin et al. [2006] proposed a compliant constraint formulation of
Müller et al. [2007] attempted to address this by exponential scaling
continuum models and a semi-implicit, velocity level integrator to
of the stiffness coefficient. However this does not take into account
handle a large range of stiffness in a time step independent manner.
the time step, and does not converge to a well-defined solution in
Our work adapts their compliant constraint formulation to PBD,
the presence of multiple constraints.
allowing us to solve constraints at the position level, and without
additional stabilization terms. In the following section we develop a scheme using regularized
constraints that have a direct correspondence to well-defined energy
A number of works use a global approach in solving constraints. potentials, and show how to solve these in a time step and iteration
Goldenthal et al. [2007] proposes a projection method for enforcing count independent manner.
inextensible constraints similar to PBD, but where the constrained
system is globally linearized and solved with a direct method at Algorithm 1 XPBD simulation loop
each iteration. Our method may be seen as a compliant version of
their Fast Projection algorithm, combined with an iterative Gauss- 1: predict position x̃ ⇐ xn + ∆tvn + ∆t2 M−1 fext (xn )
Seidel solver. Tournier et al. [2015] extend a compliant constraint 2:
formulation with second order derivative information to avoid insta- 3: initialize solve x0 ⇐ x̃
bilities due to linearization. In contrast to their work, our method 4: initialize multipliers λ0 ⇐ 0
uses only first order constraint derivatives, and avoids stability 5: while i < solverIterations do
problems related to global linearization through repeated local lin- 6: for all constraints do
earizations per-time step. 7: compute ∆λ using Eq (18)
8: compute ∆x using Eq (17)
Liu et al. [2013] propose a local–global solver for mass-spring sys- 9: update λi+1 ⇐ λi + ∆λ
tems, where a local solve is used to handle non-linear terms relat- 10: update xi+1 ⇐ xi + ∆x
ing to direction, while the global solve, which can be pre-factored, 11: end for
is used for handling stretching. The stiffness is largely indepen- 12: i⇐i+1
dent of the iteration count and the solution approaches that of New- 13: end while
ton’s method as the number of iterations increases. The idea is 14:
generalized to other constraints in Projective Dynamics by Boaziz 15: update positions xn+1 ⇐ xi
update velocities vn+1 ⇐ ∆t
1
xn+1 − xn


et al. [2014]. Wang [2015] proposes a Chebyshev type approach 16:
that combines results from previous iterations to obtain better con-
vergence. Narain et al. [2016] applies the alternating direction
method of multipliers (ADMM) for implicit time integration. Their 4 Our Method
method allows handling nonlinear constitutive models, hard con-
straints, and they show that projective dynamics is a special case of We derive our extended position-based dynamics algorithm
the ADMM. (XPBD) by starting with Newton’s equations of motion subject to
forces derived from an energy potential U (x):
These global approaches typically rely on the fact the global solve
can be done efficiently, which often means pre-factoring the global Mẍ = −∇U T (x). (3)
matrix. However, if the constraint topology changes during run-
time such as due to tearing or fracturing, the global matrix needs to Here x = [x1 , x2 , · · · , xn ]T is the system state. In PBD this is
be re-factored, which can be computationally intensive. They also often simply particle positions, however it may represent any gen-
tend to be much more involved to implement compared to PBD. eralized coordinate model, e.g.: rigid body transforms. Note that

50
 ∂g
we use the convention that the gradient operator ∇ is a row vector where K = ∂x . This system may be solved for ∆x and ∆λ and
of partial derivatives. the positions and multipliers updated accordingly:
We perform an implicit position-level time discretization of our λi+1 = λi + ∆λ (13)
equations of motion (3), where the superscript n denotes the time xi+1 = xi + ∆x. (14)
step index:
This is a fixed-point iteration that satisfies our implicit equations
xn+1 − 2xn + xn−1
 
T n+1 of motion (8, 9) for any sequence such that |xi+1 − xi | → 0 and
M = −∇U (x ). (4) |λi+1 − λi | → 0. Generally this method works well, however
∆t2
it may require a line search strategy to be robust, and computing
the system matrix can be expensive. In particular, we would like
The energy potential U (x) may be further specified in terms of a
to avoid computing K, which requires the evaluation of constraint
vector of constraint functions C = [C1 (x), C2 (x), · · · , Cm (x)]T Hessians.
as
We now introduce two approximations that simplify the implemen-
tation of our method and allow us to make a connection back to
1
U (x) = C(x)T α−1 C(x), (5) PBD. First, we use the approximation that K ≈ M. This omits the
2 geometric stiffness and constraint Hessian terms, and introduces a
local error on the order of O(∆t2 ). We note that this approxima-
where α is a block diagonal compliance matrix corresponding to tion may change the rate of convergence, but it does not change the
inverse stiffness. The force from an elastic potential is then given global error or solution to the fixed-point iteration. Specifically, it
by the negative gradient of U with respect to x, can be seen as a quasi-Newton method.
Next, we assume that g(xi , λi ) = 0. This assumption is justi-
felastic = −∇x U T = −∇CT α−1 C. (6) fied by noting that it is trivially true for the first Newton iteration
when initialized with x0 = x̃ and λ0 = 0. In addition, if the con-
We convert this to a compliant constraint formulation by following straint gradients change slowly then it will remain small, and will
Servin et al. [2006] and decomposing the force into its direction and go to zero when they are constant. Furthermore, the modified linear
scalar components by introducing the Lagrange multiplier, system now corresponds to the optimality conditions for a mass-
weighted minimization to the constraint manifold starting from the
current iterate xi . This variational point of view was discussed in
λelastic = −α̃−1 C(x). (7) detail by Goldenthal et al. [2007], and our method can be seen as a
compliant version of their Fast Projection algorithm.
Here λelastic = [λ1 , λ2 , · · · λm ]T is a vector of constraint multi- Including these approximations, our updated linear subproblem is
pliers. Henceforth we will drop the subscript unless necessary for given by
clarity. Note that we have included the time step from the left-hand
side of Equation (4) by folding it into our compliance matrix and
α
defining α̃ = ∆t 2.
    
M −∇CT (xi ) ∆x 0
=− , (15)
Substituting in our expression for λ we have the discrete con- ∇C(xi ) α̃ ∆λ h(xi , λi )
strained equations of motion:
we may then take the Schur complement with respect to M to ob-
tain the following reduced system in terms of the unknown ∆λ:
M(xn+1 − x̃) − ∇C(xn+1 )T λn+1 = 0 (8)
n+1 n+1 h i
C(x ) + α̃λ = 0, (9) ∇C(xi )M−1 ∇C(xi )T + α̃ ∆λ = −C(xi ) − α̃λi . (16)

where x̃ = 2xn − xn−1 = xn + ∆tvn is called the predicted, The position update is then given directly by evaluating
or inertial, position. To solve this non-linear system we design a
fixed point iteration based on Newton’s method. In the following,
we omit the time step superscript (n + 1) to emphasize the per- ∆x = M−1 ∇C(xi )T ∆λ. (17)
iteration unknown, indicated by the subscript (i + 1).
Although the solution returned by our method can no longer be said
We label equations (8, 9) as g and h respectively. Our goal is to to solve the implicit equations of motion exactly, in practice the
find an x and λ that satisfies: error is small. We investigate the accuracy of our method further in
Section 6.
g(x, λ) = 0 (10)
h(x, λ) = 0. (11) 4.1 A Gauss-Seidel Update

We now make a connection back to PBD by considering a Gauss-

Linearizing equations (10, 11) we obtain the following linear New- Seidel solution for our linear system of equations (16). If we take
ton subproblem: a single constraint equation with index j, we can directly compute
its Lagrange multiplier change by evaluating
    
K −∇CT (xi ) ∆x g(xi , λi ) −Cj (xi ) − α̃j λij
=− (12) ∆λj = .
∇C(xi ) α̃ ∆λ h(xi , λi ) (18)
∇Cj M−1 ∇CjT + α̃j

51
This equation forms the core of our method. During constraint solv- Substituting in our discrete approximation of velocity v =
ing, we first calculate ∆λj for a single constraint, then we update 1
∆t
(xn+1 − xn ), and linearizing with respect to λ, we have the
the system positions and multipliers using Equations (13, 14). Al- following updated Newton step:
gorithm 1 summarizes our method. It is identical to the original
PBD algorithm with the addition of lines 4, 7, and 9.
 
If we examine the equation for ∆λj more closely, we can see that α̃β̃ −1 T
(I + )∇C(xi )M ∇C(xi ) + α̃ ∆λ = −h(xi , λi ).
in the case of αj = 0 it corresponds exactly to the scaling factor ∆t
sj in the original PBD algorithm (2). From our new definition, we (25)
understand sj as being the incremental Lagrange multiplier change
for an infinitely stiff constraint. In the case of compliance, we have
αj ≥ 0, and an additional term in both the numerator and denom- In terms of a single constraint equation, our Gauss-Seidel update is
inator appears. These terms act to regularize the constraint in such now given by
a way that the constraint force is limited, and corresponds to the
elastic potential given by Equation (5). −Cj (xi ) − α̃j λij − γj ∇Cj (xi − xn )
∆λj = . (26)
(1 + γj )∇Cj M−1 ∇CjT + α̃j
The numerator in Equation (18) includes reference to λij . This is
the total Lagrange multiplier for constraint j at the current iteration
i. We must store and update this variable in addition to the system A common case is that α̃ and β̃ are simple diagonal (as opposed to
positions. The additional storage of one scalar per-constraint is a block-diagonal), then we have γj = ∆t j j α̃ β̃
, i.e.: the time step scaled
modest overhead, and provides useful information on the total con- product of compliance and damping parameters for the constraint.
straint force, and can be used to drive force dependent effects (e.g.: We stress that all the additional terms here are easily and efficiently
breakable joints), or haptic devices. computable.

5 Damping 6 Results
Our method is derived from an implicit time stepping scheme that We test our method on a variety of deformable models and con-
naturally includes some dissipation of mechanical energy. Never- straint types. For our 2D results we have used a CPU implemen-
theless, it can be useful to model additional constraint damping. To tation with Gauss-Seidel style iteration which we validate against
do so, we define a Rayleigh dissipation potential a traditional non-linear Newton solver applied directly to the dis-
crete equations of motion (8, 9). The Newton solver uses the robust
1
D(x, v) = Ċ(x)T β Ċ(x) (19) Cholesky decomposition from the Eigen math library [2010] as its
2 linear solver.
1
= vT ∇CT β∇Cv, (20) Our 3D results were captured from a GPU based solver using a Ja-
2
cobi style iteration on a NVIDIA GTX1070. Collisions are handled
exactly as in the original PBD method. We assume zero compli-
where β is a block diagonal matrix corresponding to the constraint ance in contact, meaning it is not necessary to store the Lagrange
damping coefficients. Note that β is not an inverse parameter like multiplier for the contact constraint.
compliance, and should be set as a damping stiffness. According to
Lagrangian dynamics the force from a dissipation potential comes 1.5
Exact
from the negative gradient of D with respect to velocity 1.4
XPBD
PBD 1
PBD 5
1.3 PBD 10

T T
fdamp = −∇v D = −∇C β∇Cv. (21) 1.2

1.1
Offset

As in the elastic case, we separate the scalar component of this force 1

vector into a constraint multiplier 0.9

0.8

λdamp = −β̃ Ċ(x) = −β̃∇Cv, (22) 0.7

0.6

0.5
where we have included the time step in our damping parameter by 0 10 20 30 40 50
Frame
60 70 80 90 100

defining β̃ = ∆t2 β.
Figure 2: A simple harmonic oscillator modeled as a distance con-
It is possible to solve for λdamp by itself, and in so doing obtain a straint with compliance α = 0.001. XPBD matches the analytic
constraint force multiplier for the damping and elastic components solution closely, while PBD’s solution is heavily dependent on the
individually. However, in most cases we only care to know the total number of iterations.
constraint force, and because the damping force must act along the
same direction as the elastic force, we can combine our multipliers
into a single equation 6.1 Spring

λ = λelastic + λdamp = −α̃−1 C(x) − β̃∇Cv. (23) We test our method on the case of a simple harmonic oscillator
modeled as a distance constraint, with rest position of of x = 1,
initial position x0 = 1.5, α = 0.001, and mass m = 1. As we
Re-arranging (23) into our constraint form we have can see from Figure 2, PBD cannot reproduce the correct period of
oscillation. We plot the PBD simulation with 1,5, and 10 iterations
h(x, λ) = C(x) + α̃λ + α̃β̃∇Cv = 0. (24) to illustrate how PBD’s oscillation period and damping are heavily

52
 600 0.1
Newton Newton
XPBD 50 XPBD (Gauss−Seidel)
0.09
XPBD 100 XPBD (Jacobi)
XPBD 1000
550 0.08

0.07

500 0.06

|h ( x, λ ) | 2
Force (N)

0.05

450 0.04

0.03

400 0.02 Figure 5: A cantilever beam modeled with triangular FEM ele-
0.01 ments and a linear, isotropic constitutive model, simulated for 50
350
75 80 85
Frame
90 95 100
0
0 5 10 15 20 25
Iteration
30 35 40 45 50 frames. Left: Reference Newton solver. Right: XPBD, using 20
iterations our method is visually indistinguishable from the refer-
(a) (b)
ence.
Figure 3: (a) The constraint force magnitude at the fixed support
over a series of frames in the Chain example. (b) Residual error
in h(x, λ) for a single frame taken from the Beam example, both The compliance matrix is then given by the inverse stiffness matrix,
Gauss-Seidel and Jacobi converge linearly as expected for an itera- which, in terms of the Lamé parameters is
tive method.
 −1
λ + 2µ λ 0
−1
αtri = K = λ λ + 2µ 0 . (28)
iteration count dependent. In this simple case XPBD closely repro- 0 0 2µ
duces the analytic result regardless of time step and iteration count.
We note that in this example the damping in the reference solution The advantage of this formulation over Strain-Based Dynamics
is a side-effect of our implicit discretization of the equations of mo- [Müller et al. 2014] is that, not only does it correspond to traditional
tion. material parameters, it correctly couples the individual strains to
model Poisson’s effect. This allows XPBD to accurately simulate
6.2 Chain models with material properties obtained from experimental data.
We evaluate our algorithm’s accuracy by simulating a cantilever
beam modeled using a St Venant-Kirchhoff triangular FEM dis-
cretization (Figure 5). We use a linear, isotropic constitutive model
with Young’s modulus of E = 105 , Poisson’s ratio µ = 0.3, and
∆t = 0.008. We run the simulation for 50 frames and find that 20
iterations of XPBD are sufficient to be visually indistinguishable
from our reference Newton solver. The residual error over iteration
is plotted in Figure 3b.
Figure 4: A time-lapse view of a hanging chain of 20 particles
falling under gravity. Left: Reference Newton solver. Right: XPBD 6.4 Cloth
with 50 iterations.

One of the key benefits of our method is its ability to return con-
straint force magnitudes. An important question then, is how ac-
curate are the returned constraint forces? To quantify the error in-
troduced by the PBD update we simulate a weakly extensible chain
of particles as illustrated in Figure 4. The chain consists of 20 par-
ticles, each with m = 1.0, and connected by distance constraints
with α = 10−8 . We simulate the scenario over 100 frames and
measure the resultant constraint force magnitude at the fixed par-
ticle at the top of the chain. Figure 3a compares the result of our
reference Newton solver’s solution to XPBD’s. The maximum rel-
ative error for our method over the course of the simulation is 6%,
2%, and 0.5% for 50, 100, and 1000 iterations respectively. We
believe this is acceptable for most graphics applications.
Figure 6: Hanging cloth with 20, 40, 80, and 160 iterations re-
spectively (left to right). Top row: PBD, Bottom row: XPBD. PBD
6.3 Cantilever Beam stiffness is dependent on iteration count in a non-linear way, while
XPBD’s behavior is qualitatively unchanged.
Traditional finite element methods (FEM) can be reformulated in
the compliant constraint framework [Servin et al. 2006]. In the case
of a linear isotropic constitutive model, the entries of the strain ten- We test the iteration count independent nature of our algorithm by
sor  for each element are treated as a vector of constraint functions, simulating a hanging cloth subject to gravity. The cloth model con-
here in Voigt notation for a triangular element: sists of a 64x64 grid of particles connected by a graph of 24k dis-
tance constraints. In Figure 6 we show the effect of fixing the con-
 straint stiffness and varying the iteration count. Cloth simulated
x with PBD becomes progressively stiffer and more heavily damped
Ctri (x) = tri =  y  . (27) as iteration count increases, while in XPBD behavior is consistent
xy regardless of iteration count.

53
To clearly show how the PBD behavior changes with varying it- only trivial modifications to an existing PBD solver we expect it to
eration counts we have used an artificially low constraint stiffness be easily adopted by industry and practitioners.
of k = 0.01. We note that XPBD does not converge faster than
PBD, simply that behavior remains consistent at different iteration References
counts. Indeed, in the limit of zero compliance XPBD is equivalent
a constraint stiffness of k = 1 in PBD. B ENDER , J., KOSCHIER , D., C HARRIER , P., AND W EBER , D.
We measure the performance impact of XPBD on our cloth example 2014. Position-based simulation of continuous materials. Com-
and find that it adds a small per-iteration cost that is typically less puters & Graphics 44, 1–10.
than 2% of the total simulation time. Please refer to Table 1 for B ENDER , J., M” ULLER , M., OTADUY, M. A., T ESCHNER , M.,
more detail. AND M ACKLIN , M. 2014. A survey on position-based simula-
tion methods in computer graphics. Computer Graphics Forum,
Table 1: Per-step simulation time (ms) for the hanging cloth exam-
1–25.
ple at varying iteration counts.
B OUAZIZ , S., M ARTIN , S., L IU , T., K AVAN , L., AND PAULY, M.
Iterations 20 40 80 160 2014. Projective dynamics: fusing constraint projections for fast
simulation. ACM Transactions on Graphics (TOG) 33, 4, 154.
PBD 0.95 1.75 3.25 5.61
XPBD 0.97 1.78 3.34 5.65 G OLDENTHAL , R., H ARMON , D., FATTAL , R., B ERCOVIER , M.,
AND G RINSPUN , E. 2007. Efficient simulation of inextensible
cloth. In ACM Transactions on Graphics (TOG), vol. 26, ACM,
49.
6.5 Inflatable Balloon
G UENNEBAUD , G., JACOB , B., ET AL ., 2010. Eigen v3.
We apply our method to an inflatable balloon where the interior air http://eigen.tuxfamily.org.
pressure is modeled using a global volume constraint, and the outer
surface as a cloth mesh with stretch, shear, and bending constraints L IU , T., BARGTEIL , A. W., O’B RIEN , J. F., AND K AVAN , L.
(see Figure 1). In terms of additional storage, the interior volume 2013. Fast simulation of mass-spring systems. ACM Transac-
constraint requires only a single auxiliary multiplier, plus one mul- tions on Graphics (TOG) 32, 6, 214.
tiplier for each surface constraint. Our method makes it easy to M ÜLLER , M., H EIDELBERGER , B., T ESCHNER , M., AND
adjust the relative stiffness of constraints, for example, if necessary G ROSS , M. 2005. Meshless deformations based on shape
we may increase solver iterations to obtain stronger volume preser- matching. In ACM SIGGRAPH 2005 Papers, ACM, New York,
vation without affecting surface stiffness. NY, USA, SIGGRAPH ’05, 471–478.
M ÜLLER , M., H EIDELBERGER , B., H ENNIX , M., AND R AT-
7 Limitations and Future Work CLIFF , J. 2007. Position based dynamics. J. Vis. Comun. Image
Represent. 18, 2 (Apr.), 109–118.
In the limit of α = 0 our method is identical to PBD with k = 1,
and as such, requires the same number of iterations to obtain a stiff M ÜLLER , M., C HENTANEZ , N., K IM , T.-Y., AND M ACKLIN , M.
result. Similar to PBD, low iteration counts that terminate before 2014. Strain based dynamics. In Proceedings of the ACM SIG-
convergence will result in artificial compliance. Convergence speed GRAPH/Eurographics Symposium on Computer Animation, Eu-
may be improved by replacing Gauss-Seidel or Jacobi iteration with rographics Association, Aire-la-Ville, Switzerland, Switzerland,
a more sophisticated linear solver at each iteration. Recent work SCA ’14, 149–157.
that focused on accelerated iterative methods such as the Chebyshev
method presented by Wang et al. [2015] could be applicable. NARAIN , R., OVERBY, M., AND B ROWN , G. E. 2016. ADMM
⊇ projective dynamics: Fast simulation of general constitutive
One promising avenue that our method opens up is temporal coher- models. In Proceedings of the ACM SIGGRAPH/Eurographics
ence, because we now have a record of the total constraint force it Symposium on Computer Animation. To appear.
should be possible to warm-start the constraint solve based on the
previous frame’s Lagrange multipliers. N EALEN , A., M ÜLLER , M., K EISER , R., B OXERMAN , E., AND
C ARLSON , M. 2005. Physically Based Deformable Models
Unlike Projective Dynamics [Bouaziz et al. 2014] our method is in Computer Graphics. In Eurographics 2005 - State of the
only an approximation of an implicit Euler integrator. For real- Art Reports, The Eurographics Association, Y. Chrysanthou and
time graphical applications we believe the error is acceptable, and M. Magnor, Eds.
in many cases indistinguishable. However, traditional methods may
be more suitable for applications requiring greater accuracy guar- S ERVIN , M., L ACOURSIERE , C., AND M ELIN , N. 2006. Interac-
antees. tive simulation of elastic deformable materials. In Proceedings
of SIGRAD Conference, 22–32.
8 Conclusion S TAM , J. 2009. Nucleus: Towards a unified dynamics solver for
computer graphics. In Computer-Aided Design and Computer
We have presented a simple extension to position-based dynamics Graphics, 2009. CAD/Graphics’ 09. 11th IEEE International
that addresses its most well known limitation, namely, time step and Conference on, IEEE, 1–11.
iteration count stiffness dependence. Our method requires comput- T OURNIER , M., N ESME , M., G ILLES , B., AND FAURE , F. 2015.
ing and storing only a single additional scalar per-constraint, but al- Stable constrained dynamics. ACM Transactions on Graphics
lows PBD to simulate arbitrary elastic and dissipative energy poten- (TOG) 34, 4, 132.
tials. In addition, it can provide accurate constraint force estimates
for force dependent effects. We believe this widens the scope of WANG , H. 2015. A chebyshev semi-iterative approach for accel-
PBD to applications which require greater accuracy and correspon- erating projective and position-based dynamics. ACM Transac-
dence to traditional material models. Because our method requires tions on Graphics (TOG) 34, 6, 246.

54