3D Video Games 2020-04-23

05: Game Physics - part3

Course Plan

lec. 1: Introduction 
lec. 2: Mathematics for 3D Games 
lec. 3: Scene Graph 
lec. 4: Game 3D Physics  + 
lec. 5: Game Particle Systems 
lec. 6: Game 3D Models 
lec. 7: Game Textures 
lec. 8: Game 3D Animations 
lec. 9: Game 3D Audio 
lec. 10: Networking for 3D Games 
lec. 11: Artificial Intelligence for 3D Games 
lec. 12: Game 3D Rendering Techniques 
90

Verlet integration method

 Idea: remove velocity from state

 Current velocity is implicit
 It’s defined from: 𝐩
 current pos 𝐩
𝐩
 last pos 𝐩
which we need to record

𝐩 =𝐩 + 𝑣⃗ · 𝑑𝑡 Euler & friends

𝑣⃗ = (𝐩 −𝐩 )/𝑑𝑡 Verlet

91

Marco Tarini
Università degli studi di Milano 1
3D Video Games 2020-04-23
05: Game Physics - part3

Verlet integration method

init 𝐩 =...
state 𝐩 = ...

𝑓⃗ = 𝑓𝑢𝑛(𝐩 ,…)
one
𝑎⃗ = 𝑓⃗/𝑚 expanding
this…
step 𝑣⃗ = (𝐩 − 𝐩 )/𝑑𝑡
𝑣⃗ = 𝑣⃗ + 𝑎⃗ ⋅ 𝑑𝑡
𝐩 =𝐩 + 𝑣⃗ ⋅ 𝑑𝑡

92

Verlet method

init 𝐩 ⟵ ...
state 𝐩 ⟵ ...

𝑓⃗ ⟵ 𝑓𝑢𝑛(𝐩 )
𝐩 ⟵𝐩
one 𝑎⃗ ⟵ 𝑓⃗/𝑚
step
𝐩 ⟵𝐩
𝐩 ⟵ 2𝐩 −𝐩 + 𝑎⃗ ⋅ 𝑑𝑡

93

Marco Tarini
Università degli studi di Milano 2
3D Video Games 2020-04-23
05: Game Physics - part3

Verlet: characteristics

 Velocity is kept implicit

 but that doesn’t save RAM:
we need so store previous position instead
 Good efficiency / accuracy ratio
 per step error: linear with dt
 accumulated error: order of dt2 (second order method)
 Extra bonus: reversibility
 it’s possible to go backward in t and
reach the initial state
 that, in theory… careful with implementation details

94

Verlet: caveats

 it assumes a constant dt (time-step duration)

 if it varies: corrections are needed! (which ones?)
 Q: how to act on velocity (which is now implicit)?
 e.g. to apply impulses
 A: change 𝐩 instead
 Q: how to act of positions w/o impacting velocity?
 e.g. to apply teleports / kinematic motions
 A: displace both 𝐩 and 𝐩 by the same amount
 Q: how to apply velocity damps?
 A: act on 𝐩 or 𝐩 (see below)
95

Marco Tarini
Università degli studi di Milano 3
3D Video Games 2020-04-23
05: Game Physics - part3

dt updates in Verlet
(if they are not constant)

Problem:
if 𝑑𝑡 now changes to a new 𝑑𝑡′
then, all pold must be updated to some 𝐩

current velocity 𝑣⃗
Find p'old : 𝑣⃗ = (𝐩 −𝐩 )/𝑑𝑡
and position 𝐩
𝑣⃗ = (𝐩 −𝐩 )/𝑑𝑡′ must not change

𝐩 =𝐩 ⋅ (𝑑𝑡 − 𝑑𝑡′)/𝑑𝑡 + 𝐩 ⋅ 𝑑𝑡′/𝑑𝑡

96

Velocity damping in Verlet

implicit

 Velocity at next frame: 𝑣⃗ = (𝐩 −𝐩 )/𝑑𝑡

 We want to multiply 𝑣⃗ a factor 𝑐 e.g. 0.98

 before applying accelerations obtained as
(1-𝑑𝑡·𝑐 )
 We can do that using a more general formula for 𝐩

𝐩 = 2 ⋅𝐩 − 1 ⋅𝐩 + 𝑑𝑡 ⋅ 𝑎⃗

𝐩 = 1+𝑐 ⋅𝐩 −𝑐 ⋅𝐩 + 𝑑𝑡 ⋅ 𝑎⃗
98

Marco Tarini
Università degli studi di Milano 4
3D Video Games 2020-04-23
05: Game Physics - part3

Velocity damping in Verlet

(geometric interpretation)
a bit shorter

𝑣⃗ 0.98𝑣⃗
𝐩 𝐩
𝑣⃗ 𝑣⃗
𝐩 𝐩

𝐩 𝐩

𝐩 = 2 ⋅𝐩 − 1 ⋅𝐩 𝐩 = 1.98 ⋅ 𝐩 − 0.98 ⋅ 𝐩

That is , That is ,
𝐩 is an extrapolation 𝐩 is a different extrapolation
of 𝐩 , 𝐩 : of 𝐩 , 𝐩 :

𝐩 = 𝑚𝑖𝑥( 𝐩 , 𝐩 , 2) 𝐩 = 𝑚𝑖𝑥( 𝐩 , 𝐩 , 1.98)

99

Verlet with PBD

“Position Based Dynamics”
init 𝐩 ⟵ ...
state 𝐩 ⟵ ...

𝑓⃗ ⟵ 𝑓𝑢𝑛(𝐩 )
𝐩 =𝐩
𝑎⃗ ⟵ 𝑓⃗/𝑚 𝐩 =𝐩
one
step 𝐩 ⟵ 2𝐩 −𝐩 + 𝑎⃗ ⋅ 𝑑𝑡

Enforce constraints on (𝐩 ) 💡

100

Marco Tarini
Università degli studi di Milano 5
3D Video Games 2020-04-23
05: Game Physics - part3

Position Based Dynamics

a formula
 A positional constraint is with ‘=‘ ‘>’ ‘<‘ etc.

an equation/inequality
involving the positions of particles.
 Useful, for example, to model consistency conditions
 Like “solid objects don’t compenetrate each other”,
or “steel bars won’t bend”
 We will see specific examples soon
 We enforce (impose) positional constraint directly
💡
by displacing the positions of particles
 Thanks to Verlet: this displacement automatically cause
some appropriate update of the velocity!
 not necessarily the correct one, but a plausible one

101

Verlet + Position Based Dynamics.

Advantages
 flexibility: different constraints can be used to model
many different phenomena
 Useful constraints are straightforward to define
 They are easy to impose (they involve only few particles)
 They can be used to model many possible phenomena
 Esamples: see following slides
 robustness : plausibility is ensured by explicitly
enforced the conditions we want to see
 For exampe: a ball won’t ever be seen outside the box
containing it (at lest, not for long)
 Bypasses the need to using forces / impulses to enforce
the same consistency condition
 Which is much more difficult to enforce

102

Marco Tarini
Università degli studi di Milano 6
3D Video Games 2020-04-23
05: Game Physics - part3

Example of positional constraint

«Particles must stay

within [0 – 100] x [0 – 100] »
100 Imposing constraint: simple clamp !
a
ex:
for(int i=0; i<NUM_PARTICLES; i++)
b
{
p[i].x = clamp( p[i].x, 0, 100 );
p[i].y = clamp( p[i].y, 0, 100 );
}
0 100

Imposing constraints like this is a first part of collision response.

For re-bounces, impulses must still be added (see collisions).
103

Example of positional constraint:

equidistance constraint
«Particles a and b must stay at distance d »

104

Marco Tarini
Università degli studi di Milano 7
3D Video Games 2020-04-23
05: Game Physics - part3

Enforce equidistance constraints

if 𝐩 − 𝐩 >𝑑

if 𝐩 − 𝐩 <𝑑

105

Enforce equidistance constraints:

pseudo code
Vector3 pa, pb; // curr positions of a,b
float d; // distance (to enforce)

Vector3 d = pa – pb;
float currDist = v.length;

d /= currDist; // normalization of d

float delta = currDist – d ;

pa += ( 0.5 * delta) * d;
pb -= ( 0.5 * delta) * d;
assuming equal mass, each particle moves half the way
(see later for the general case)

106

Marco Tarini
Università degli studi di Milano 8
3D Video Games 2020-04-23
05: Game Physics - part3

Enforcing sets of constraints

 Many constraints to impose:

when you solve one  you break another one!
 Simultaneous enforcement: computationally expensive

 Practical solution: enforce them in cascade

(Gauss-Seidel fashon):

Constr. Constr. ... Constr.

1 2 N

Repeat until convergence (= max error below threshold)

…but at most for N times! even 1 (remember: soft real time)

107

Enforcing sets of constraints

 Note:
 The whole loop for imposing the constraints happen in
the constraint enforcement phase on one physics step!
 Convergence:
 if constraints are not contradictory
 if convergence not reached (or solution doesn’t exist):
never mind, next frames will fix it (it’s fairly robust)
 needed iterations (typically): 1 ~ 10 (efficient!).
 Optimization (to decrease number of needed iterations):
solve the most unsatisfied constraints first
 Problem: it’s a sequential approach! 
 parallelized versions (similar to Jacobi) are possible
 they have a worse convergence in practice
(they require more iterations)
108

Marco Tarini
Università degli studi di Milano 9
3D Video Games 2020-04-23
05: Game Physics - part3

Equidistance constraints
VS springs
 They are similar
 they both mean:
these 2 particles “want to be” at this distance (not more, not less)
 Differences:
 equidistance constraint:  spring:
 applied during  applied during
constraint enforcement force evaluation step
 directly affecting  affecting forces,
positions therefore accelerations
 models a rigid rod  models a deformable spring
between the two particles between the two particles
 of a given length  of a given length
 sometimes called  sometimes called
an “HARD” constraint a “SOFT” constraint
 A physic engine can combine them in one object!
109

We can combine equidistance

constraints to obtain…
 Rigid bodies

 Articulated bodies

 Ragdolls

 Cloth

 Non-elastic ropes

 And more
110

Marco Tarini
Università degli studi di Milano 10
3D Video Games 2020-04-23
05: Game Physics - part3

Compounds of particles
disguised as rigid bodies

111

Combining equidistance constraints

we obtain rigid objects
 Rigid body dynamics
as emerging behavior
 without explicitly updating
their orientation, angular vel,
angular acc., etc.
A box?
(rigid object)
A configuration of:
• 4 particles
• 6 equidistance constraints

112

Marco Tarini
Università degli studi di Milano 11
3D Video Games 2020-04-23
05: Game Physics - part3

Example

NO NO NO

STEP 0 STEP 1 STEP 1

before constraints after 1st constraint

113

Example

In total: the “box”,

under gravity + impact
• had rotated
• gained angular velocity
(will keep rotating by
inertia)
even the system does not
(explicitly) handle rotations
or
NO NO angular velocities

(works in 3D as well!)

STEP 1 STEP 1
after all constraints (implicit) velocities
multiple times
114

Marco Tarini
Università degli studi di Milano 12
3D Video Games 2020-04-23
05: Game Physics - part3

More examples of
positional constraints
 Preserve volume of some object: «Volume is 𝑣 »
 How to impose it:
1. Estimate current total volume 𝑣
2. uniform scaling of the entire object of 𝑣 /𝑣
 Fixed positions: «particle 𝑎 stays in 𝐩 »
 particles «pinned in position»
 trivial to impose, but useful!
𝐩
 Angle constraints, e.g. 𝛂 < 𝛂 𝛂
 e.g. on joints: «elbows cannot bend backward» 𝐩
 Coplanarity / collinearity 𝐩

 Non interpenetration
 this is part of collision handling – see collisions later

115

Enforcing a positional constraint:

the general case.
 Check: does the equation/inequality hold?
 If so, nothing to do!
 Else:
 All positions must be displaced a bit so that it does
 Infinite ways to achieve this. Which one to pick?
 Answer:
minimize the sum of squared displacements
(with respect to current position)
weighted by particle masses
 Find minimizer by analytically solving simple problems
(in closed form, “analytically” = “on paper”)
116

Marco Tarini
Università degli studi di Milano 13
3D Video Games 2020-04-23
05: Game Physics - part3

Enforcing a positional constraint

the general case: formally problem
 We want to enforce a constraint 𝒞 on particles a , b , c,…
with have mass ma, mb, mc …
 𝒞 defined as an equation/inequality of their positions pa , pb , pc , …

 We must apply the displacements 𝑑 , 𝑑 , 𝑑

which minimize:

argmin ma 𝑑 + mb 𝑑 + mc 𝑑 +⋯
, , ,…

such that 𝒞 pa + 𝑑 , pb + 𝑑 , pc + 𝑑 ,…

among all the choices that satisfy this,

we want the one which minimizes this

117

Enforcing positional constraint

Example: equidistance constraint
 To enforce the constraint
“particles a and b must stay at distance k ”
 input: current positions pa, pb
 input: masses ma, mb
 We need to the the displacements 𝑑 , 𝑑
found by minimizing:
argmin ma 𝑑 + mb 𝑑
,
such that pa + 𝑑 − pb + 𝑑 =𝑘
 And the solution (in closed form) is…
118

Marco Tarini
Università degli studi di Milano 14
3D Video Games 2020-04-23
05: Game Physics - part3

Equidistance constraints: solution for

non-equal masses
Vector3 pa, pb; // curr positions of a,b
float ma, mb; // masses of a,b
float d; // distance (to enforce)

Vector3 v = pa – pb;
float currDist = v.length;

v /= currDist; // normalization of v

float delta = currDist – d ;

/* solutions of the minimization: */

pa += ( mb/(ma+mb) * delta) * v;
pb -= ( ma/(ma+mb) * delta) * v;
119

Enforcing positional constraint

Example: “don’t sink into a plane”
 We want to enforce the constraint
“particle a must be above a constant plane ”
 Given: position of the particle pa and its mass ma
 Point on a plane pq and its normal (unit vec) 𝑛
 We need to apply the displacement 𝑑
found by minimizing:
argmin ma 𝑑
,
such that pa − pq 𝑛 >0
 And the solution (in closed form) is, trivially…
120

Marco Tarini
Università degli studi di Milano 15
3D Video Games 2020-04-23
05: Game Physics - part3

In pseudocode

Vector3 pa; // curr positions of a

float ma; // mass (no effect here)
Vector3 pq; // point on the plane
Vector3 nq; // normal of the plane (unit)

Vector3 v = pa – pq;
float currDist = Vector3.dot( v , n );

if (currDist < 0.0)

pa -= currDist * n; // just project!
else {} // constrain holds, do nothing

121

Rigid objects as compounds of

constrained particles: advantages
 Interesting/rich/useful set of “emerging behaviors”
(i.e. effects with “just automatically happens”) :
 rigid, deformable, jointed objects
 made of particles + hard constraints consequence
 their angular velocities you don’t of
 rotation around proper axis need to constraints
compute disallowing
 their barycenter
or store compene-
 their momentum of inertia these tration
 angular velocity is maintained

 somewhat believable bounces on “impacts”

 for more control: impact impulses can be added (see collisions)
122

Marco Tarini
Università degli studi di Milano 16
3D Video Games 2020-04-23
05: Game Physics - part3

Particles + constraint,
or rigid bodies?
 Rigid-body based systems:
 explicitly compute dynamics for rigid bodies
 updating their rotation, angular speed,…
 Particles-based systems:
 only compute dynamics for particles
 rigid (or deformable, or jointed) bodies
as an emerging behavior
 Mixed systems:
 use both
 may even dynamically swap between the
two representations for rigid bodies

124

Rigid body as particles + constraints:

Challenges
 Approximations are introduced
 e.g.: mass is concentrated in a few locations
 Scalability issues
 many constraints to enforce, many particles to track
 Some of the info which is kept implicit
is needed by the rest of the game engine
 and must therefore be extracted 
 example: the transform (position + orientation) of the
“rigid body” is needed to render the associated mesh
 similarly: angular speed, barycenter pos, velocity…

125

Marco Tarini
Università degli studi di Milano 17