Real Time Physics

y
www.matthiasmueller.info/realtimephysics

Matthias Müller Doug James Jos Stam Nils Thuerey

Cornell University
Schedule
• 8:30 Introduction, Matthias Müller
• 8:45 Deformable Objects, Matthias Müller
• 9:30 Multimodal Physics and User Interaction
Doug James

• 10:15 Break
• 10:30 Fluids, Nils Thuerey
• 11:15 Unified Solver, Jos Stam
• 12:00 Q & A
Real Time Demos
Before Real Time Physics
y

• Myy Ph.d. thesis:

– Find 3d shape of dense polymer systems

!BIOSYM molecular_data 4 !BIOSYM archive

@molecule POLYCARB_B0 PBC=OFF

CARB_1:C C cp C1/1.5 C5/1.5 HC C 3.313660622 -2.504962206 -11.698267937

CARB_1:HC H hc C HC 2.429594755 -2.005253792 -12.065854073
CARB_1:C1
_ C cp C/1.5 C2/1.5 H1 C1 4.420852184 -1.754677892 -11.284504890
CARB_1:H1 H hc C1 H1 4.390906811 -0.676178515 -11.332902908
CARB_1:C2 C cp C1/1.5 C3/1.5 C7 C2 5.566863537 -2.402448654 -10.808005333
CARB_1:C3 C cp C2/1.5 C4/1.5 H3 C3 5.605681896 -3.800502777 -10.745265961
CARB_1:H3 H hc C3 H3 6.489747524 -4.300211430 -10.377680779
CARB_1:C4 C cp C3/1.5 C5/1.5 H4 C4 4.498489857 -4.550786972 -11.159029007
CARB_1:H4 H hc C4 H4 4.528435707 -5.629286766 -11.110631943
... ...
... ...
Meeting
g Real Time Physics
y
• Post doc at MIT (1999-2001)
– Plan: Parallelization of packing algorithms
– Prof had left MIT before I arrived!

• Change of research focus

– Computer graphics lab on same floor
– Real
Real-time
time physics needed for a
virtual sculptor

B C tler et al.
B.Cutler al
1999
• Among my literature search:
– D. James et al., ArtDefo, Accurate Real Time
Deformable Objects, Siggraph 1999
– J.Stam, Stable Fluids, Siggraph 1999
• They brought physics brought to life!
• My assignment: make this real-time:
– J. O‘Brien at al., Graphical Modeling and Animation of
Brittle Fracture, Siggraph 1999
ArtDefo
• Boundary element method
• Haptic interaction
Doug
g James
• CV
– 2001: PhD in applied mathematics, University of British Columbia
– 2002: Assistant prof, Carnegie Mellon University
– 2006: Associate prof, Cornell Universtiy
– National Science Foundation CAREER award

• Research interests
– Physically based animation
– Haptic force feedback rendering
– Reduced-order modeling
Stable Fluids
• Semi-Lagrangian advection
• Equation splitting
Jos Stam
• CV
– PhD in computer science, University of Toronto
– Postdoc in Paris and Helsinki
– Senior research scientist at Alias|Wavefront, now Autodesk
– SIGGRAPH Technical Achievement Award

• Research interests
– Natural phenomena
– Physics based simulation
– Rendering and surface modeling
Animation of Brittle Fracture
• Finite elements, separation tensor
• Great results but 5-10 min/frame

J. O‘Brien et al.
Real-Time Fracture of Stiff Materials
• Hybrid rigid body – static FEM
• Not quite as realistic but 30 fps

M.Müller et al. Eurographics CAS 2001

Deformables and Water
• Post doc with ETH computer graphics lab

2004 2003 Video by D.Charypar

2003, D Charypar

FEM base deformables SPH fluids

NovodeX - AGEIA
• 2003 NovodeX as ETH spin-off
• 2004 Acquisition by AGEIA
• 2007 Nils Thuereyy AGEIA p
post doc
Nils Thuerey
y
• CV
– 2007: PhD in computer science from University Erlangen
– 2007: Post doc with AGEIA
– 2008: Post doc with ETH

• Research interests
– Lattice-Boltzmann based fluid simulation
– Real-time height
g field fluid simulation
– Fluid Control
Offline Physics
y
• Applications
– Special effects in movies and commercials

• Typical setup
– Millions of particles / triangles / tetrahedra / grid cells
– Expensive photorealistic rendering
– Impressive high quality results
– Seconds up to hours per frame

• Characteristics
– Predictable,
Predictable re-run
re run possible,
possible no interaction
Real Time Physics
y
• Applications
– Interactive
I t ti systems
t
– Virtual surgery simulators („respectable“, „scientific“)
– Games
G (not
( t so respectable
t bl but
b t true
t i 99%)
in

• Requirements
– Fast, 40-60 fps of which physics only gets a small fraction
– Stable in any possible, non-predictable situation

• Challenge:
– Approach offline results while meeting all requirements!
From Offline to Real Time
• Resolution reduction
– Blobby and coarse look
– Details disappear

• Use specialized real time techniques!

– Physics
y low-res,, appearance
pp hi-res ((shader effects))
– Reduction of dimension from 3d to 2d (height field fluids, BEM)
– Level of detail (LOD)
( )
– No equation solving, procedural animation for specific effects
Deformable Objects
Examples
p of Deformable Objects
j

• 1d: Ropes,
Ropes hair

• 2d: Cloth, clothing

• 3d: Fat,
Fat tires,
tires organs
Dimensionality
y

• Every real object is 3d

• Approximated object with lower dimentional models if
possible
• Dimension reduction substantially saves simulation time
Mass
ass Spring
Sp g Syste
Systems
s
Mass Spring
p g Meshes

• Rope: chain
– Additional springs for bending
and torsional resistance needed

• Cloth: triangle mesh

– Additional springs for bending
restistance needed

• Soft
S ft body:
b d tetrahedral
t t h d l mesh
h
Mass Spring
p g Physics
y
• Mass point: mass m, position x, velocity v
v
x m
f -f
• Springs: xi xj

l0
x j − xi ⎡ x j − xi ⎤
f= ⎢ k s ( x j − x i − lo ) + k d ( v j − v i ) ⋅ ⎥
x j − x i ⎢⎣ x j − xi ⎥⎦

• Scalars ks, kd, stretching, damping coefficients

Time Integration
g
• Newton: v& = f / m
x& = v

• Explicit Euler: vi
t +1
= v i + Δt
t 1
∑ f (x i
t t t t
, vi , x j , v j )
mi j
t +1 t +1
= x i + Δt v i
t
xi

• Assumes velocity and force constant within Δt

• Correct would be:
t + Δt
x(t + Δt ) = x(t ) + ∫ v(t )dt
t
Explicit
p Euler Issues
• Accuracy
– Better with higher order schemes e.g. Runge Kutta
– Not critical in real time environments

• Stability Δt2f/m
– Overshooting
– Big issue in
real time systems!
Implicit
p Integration
g
• Use values of next time step on the right
1
∑ f (x i , v i , x j , v j )
t +1 t +1 t +1 t +1 t +1
= v i + Δt
t
vi
mi j
t +1 t +1
= x i + Δt v i
t
xi

• Intuitively
– Don‘t extrapolate blindly
– Arrive at a physical configuration
Implicit
p Integration
g Issues
• Unconditionally stable (for any Δt)!
• Have to solve system of equations for velocities
– n mass p
points,, 3n unknowns
– Non linear when the forces are non-linear
in the positions as with springs
– Linearize forces at each time step (Newton-Raphson)

• Slow → Take large

g time steps
p
• Temproal details disappear, numerical damping
Position
os t o Based
ased Dynamics
y a cs
Force Based Update
p

penetration causes forces change velocities change

forces velocities positions

• Reaction lag
• Small ks → squashy system
• Large ks → stiff system, overshooting
Position Based Update
p

penetration move objects so that

update velocities!
detection only they do not penetrate

• Controlled position change

• Only as much as needed → no overshooting
• Velocity update needed to get 2nd order system!
Position Based Integration
g
Init all xi0, vi0
Loop
pi = xit + Δt·vit // prediction
xit+1 = modify pi // position correction
ui = [xit+1 – xit] /Δt // velocity update
vit+1 = modifyy ui // velocityy correction
End loop

• Explicit,
Explicit Verlet related
• If correction done by a solver → semi implicit
Position Correction
• Move vertices out of other objects
• Move vertices such that constraints are satisfied
• Example:
p Particle on circle
prediction

new velocity
correction
Velocity
y Correction
• External forces: vt = ut + Δt·g/m
• Internal damping
• Friction
• Restitution
corrected
velocity
l it
prediction
friction
restitution
collision correction
Internal Distance Constraint
x1 − x 2
Δx2 Δx1 = −
w1
( x1 − x 2 − l0 )
Δx1 l0 w1 + w2 x1 − x 2
m2
x1 − x 2
m1 Δx 2 = +
w2
( x1 − x 2 − l0 )
w1 + w2 x1 − x 2
wi = 1 / mi
• Conservation of momentum
• Stiffness: scale corrections by k ∈ [0..1]
• Easy to tune but effect dependent on time step!
General Internal Constraint
• Define constraint via scalar function:
Cstretch (x1 , x 2 ) = x1 − x 2 − l0
Cvolume (x1 , x 2 , x 3 , x 4 ) = [(x 2 − x1 ) × (x 3 − x1 )]⋅ (x 4 − x1 ) − 6 v0

C=0 ∇C

Rigid body modes

General Position Correction

• Correction along gradient

Δx i = − s wi ∇ xi C (x1 ,..., x n )

• Scalar s tells us how far to go

C (x1 ,..., x n )
s=
∑ w j ∇ xC (x1 ,..., x n )
2
j
Shape
p Matching
g Idea
• Optimally match
undeformed
d f d with
ith Δxi
deformed shape
pi
• Only
O l allow
ll t
translation
l ti
and rotation
• Global
Gl b l correction,
ti no
propagation needed
• No mesh needed!
Shape
p Matching
g
• Let xi be the undeformed vertex positions
• The optimal translation is
t = p cm − x cm where p cm = ∑ mi p i / ∑ mi and x cm = ∑ mi x i / ∑ mi
i i i i

• The optimal linear transformation is

−1
⎛ T ⎞⎛ T ⎞
A = ⎜ ∑ mi (p i − p cm )(x i − x cm ) ⎟⎜ ∑ mi (x i − x cm )(x i − x cm ) ⎟
⎝ i ⎠⎝ i ⎠
• The optimal rotation R is the rotational part of A
(use polar decomposition)
2d Shape
p Matching
g Demo
Working
g with Points and Edges
g
• No notion of volume or area
– Spring stiffness (N/m) not related to 3d stiffness (N/m2)

• Volumetric behavior dependent on

– Tesselation of volume
– Hand tune spring stiffnesses

• Often OK in real time environments

– Evenly tesselated physics meshes
– Fixed time step
Co-Rotated
Co otated Finite
te Elements
e e ts
Continuum Mechanics on one Slide

• Body as continuous set of points p(x)

• Deformation continuous function p(x)

• Elasticity
Elasticit theor
theory yields
ields felast(x)
( ) from p(x)
( )
• PDE of motion (Newton): x

ρ p tt (x, t ) = f elast (x, t ) + f ext (x, t )

• Solve for p(x,t)

p(x t)
• Analytical solution only for very simple problems
Finite Element Method on one Slide
• Represent body by set of finite elements (tetrahedra)
• Represent continuous p(x) by vectors pi on vertices
x3
p3
p1

x2 x1 p0
x0 p2

• pi induce simple continuous p(x) within each element

• Continuous elasticity theory yields forces at vertices
Hyper
yp Spring
p g
• Vertex forces depend on displacements of all 4 vertices
[f 0 , f1 , f 2 , f 3 ] = Ftetra (p 0 , p1 , p 2 , p 3 , x 0 , x1 , x 2 , x 3 )

• Tetrahedron acts like a hyper spring

• Compare to: [f0, f1] = Fspring(p0,p1,l0)
• Given Ftetra () -blackbox, simulate as mass spring system
• Ftetra () is non linear, expensive
Linearization
• Linearization Ftetra () of yields

⎡f 0 ⎤ ⎡p 0 − x 0 ⎤
⎢f ⎥ ⎢p − x ⎥
⎢ 1 ⎥ = K ⎢ 1 1 ⎥, K ∈ R12×12
⎢f 2 ⎥ ⎢p 2 − x 2 ⎥
⎢ ⎥ ⎢ ⎥
⎣f 3 ⎦ ⎣p 3 − x 3 ⎦

• K depends on x0, x1, x2, x3 and can be pre-computed

(see class notes for how to compute)

• Much faster to evaluate

Linearization Artifact
• Linearization only valid close to
the point off linearization

x p

linearized non-linear
Corotational Formulation
• Only rotations problematic, translations OK
• Extract rotation: RK(RTp-x)

f R
p-x

RT

RTp-x
p K(RTp-x)
Rotational Part
• Modified force computation
⎡f 0 ⎤ ⎡R 0 0 0 ⎤ ⎛⎜ ⎡R T p 0 ⎤ ⎡x 0 ⎤ ⎞
⎢f ⎥ ⎢ 0 R 0 0 ⎥ ⎜ ⎢ T ⎥ ⎢ ⎥⎟
⎢ 1⎥ = ⎢ ⎥K⎜ ⎢ R p1 ⎥ − ⎢ x1 ⎥ ⎟
⎢f 2 ⎥ ⎢ 0 0 R 0 ⎥ ⎜ ⎢R T p 2 ⎥ ⎢x 2 ⎥ ⎟
⎢ ⎥ ⎢ ⎥ ⎜ ⎢ T ⎥ ⎢ ⎥ ⎟⎟
⎣ 3 ⎦ ⎣ 0 0 0 R⎦ ⎝
f ⎣⎢ R p 3 ⎥⎦ ⎣ x 3 ⎦ ⎠

• Transformation matrix
A = [p1 − p 0 , p 2 − p 0 , p 3 − p 0 ][x1 − x 0 , x 2 − x 0 , x 3 − x 0 ]−1

• Rotation via polar decomposition of A

Advantages
g
• Matrix K can still be precomputed
• Artifacts removed
• Faster force computation in explicit
formulation
• Implicit time integration yields linear system
→ no Newton-Raphson
p solver needed
FEM Demo
Conclusions
• Trade-off speed, accuracy, stability
• Choose method accordingly
• Stability most important in real time systems
– Non predictable situations
– No time step adaptions
– No roll backs
• Remaining choice: accuracy vs. speed
Cloth in Games
Mesh Generation
• Input
– Graphical triangle surface mesh
– Extreme case: Triangle
g soup
p
• Output
– Input independent tesselation
– User specify resolution (LOD)
– Equally sized elements (stability, spatial hashing)
Surface Creation

• Input triangle mesh

• Each triangle adds density

to a regular
egu a ggrid
d
• Extract iso surface using
marching
g cubes
• Optional: Keep largest
connected mesh onlyy
• Quadric simplification
Tetrahedra Creation
• Delaunay
Tetrahedralization on
vertices of surface mesh
• Triangles of surface mesh
are used for clipping
tetrahedra (if necessary)
• Graphical mesh is moved
along with tetra mesh
using barycentric coords