~1~ Computer Graphics, Volume 22, Number 4, August 1988

Collision Detection and Response for Computer Animation

Matthew Moore and Jane Wilhelms

Computer Graphics & Imaging Laboratory

Computer & Information Sciences Board
University of California at Santa Cruz
Santa Cruz, California 95064

Abstract tional problem of detecting and controlling object interactions

When several objects are moved about by computer ani- is encountered. When no special attention is paid to object in-
teractions, the objects will sail majestically through each oth-
marion, there is the chance that they will interpenetrate. This
is often an undesired state, particularly if the animation is er, which is usually not physically reasonable and produces a
seeking to model a realistic world. Two issues are involved: disconcerting visual effect. Whenever two objects attempt to
detecting that a collision has occurred, and responding to it. interpenetrate each other, we call it a collision.
The former is fundamentally a kinematic problem, involving The most general requirement that arises from this is an
the positional relationship of objects in the world. The latter ability to detect collisions. Most animation systems at present
is a dynamic problem, in that it involves predicting behavior do not provide even minimal collision detection, but require
according to physical laws. This paper discusses collision the animator to visually inspect the scene for object interac-
detection and response in general, presents two collision tion and respond accordingly. This is time-consuming and
detection algorithms, describes modeling collisions of arbi- difficult even for keyframe or parameter systems where the
trary bodies using springs, and presents an analytical collision user explicitly defines the motion; it is even worse for pro-
response algorithm for articulated rigid bodies that conserves cedural or dynamical animation systems where the motion is
linear and angular momentum. generated by subroutines and laws defining their behavior.
Though automatic collision detection is somewhat expensive
C R Categories and Subject Descriptors: 1.3.5: [Computer both to code and to run, it is a considerable convenience for
Graphics]: Computational Geometry and Object Modeling - animators, particularly when more automated methods of mo-
Geometric algorithms; 1.3.7: [Computer Graphics]: Three Di- tion control, such as dynamics or behavioral control, are
mensional Graphics and Realism - Animation. used.[24, 31] This paper describes two collision detection al-
gorithms. One algorithm deals with triangulated surface
Key W o r d s and Phrases: computer animation, collision
representations of objects, and is appropriate for flexible or ri-
detection, collision response, analytical solution, dynamical
gid surfaces. The other algorithm applies to objects modeled
simulation.
as rigid polyhedra. Both algorithms are simple, robust, and not
dreadfully expensive.
1. O V E R V I E W
The related issue is response to collisions once they are
Computer animation provides a number of methods for
detected. Even keyframe systems could benefit from automat-
controlling object motion.[28] The object's positions and ic suggestions about the motion of objects immediately fol-
orientations as functions of time may be interpolated from
lowing a collision; animation systems using dynamical simu-
keyframes or parameter specification,[27] or may be the out-
lation inherently must respond to collisions automatically and
put of special computer programs written by the user,[23] or realistically. Linear and angular momentum must be
may be produced by physical simulation of the effect of inter-
preserved, and surface friction and elasticity must be reason-
nal, model-derived, and user-specified forces and
able. This article presents two methods that satisfy these cri-
torques.J1, 12, 16,29, 32] In any such scheme, the main ques-
teria. One is the obvious method, based on temporary springs
tions when animating a single object are how to achieve real- introduced at collision points. The other method is an analyti-
istic motion and how to economize on the human animator's cal linear system solution. The former method is more gen-
time. When several objects are animated at once, the addi- eral, working equally well for flexible, rigid, and articulated
Permission to copy without fee all or part of this material is granted bodies. The latter, limited to rigid and articulated objects, is
provided that the copies are not made or distributed for direct typically faster. Furthermore, while the spring solution as-
commercial advantage, the ACM copyright notice and the title of the sumes the ability to use the dynamics equations of motion to
publication and its date appear, and notice is given that copying is by predict the motion immediately after impact, the analytical
permission of the Association for Computing Machinery. To copy solution could be used within a kinematic animation system.
otherwise, or to republish, requires a fee and/or specific permission.

@1988 ACM-0-89791-275-6/88/008/0289 $00.75

289
 SIGGRAPH '88, Atlanta, August 1-5, 1988

2. C O L L I S I O N D E T E C T I O N the time step.

Collision detection involves determining when one ob- The hard case is solved by setting up the parametric vec-
ject penetrates another. It is clearly an expensive proposition, tor equation
particularly when large numbers of objects are involved and P +V t =Po+Vo t +((PI-Po)+(VI-Vo) t )u
the objects have complex shapes. Collision detection has been
+ ((P2-Po) + (V2-V0) t ) v
extensively pursued in the fields of CAD/CAM and robot-
ies,[2, 3, 6,7, 11, 30] and it is with some diffidence that we where P is the point (with velocity V per time step), the Pi s
offer any more algorithms. Some published algorithms[2, 3] define the triangle vertices (with velocity Vi per time step),
solve the problem in more generality (and at higher cost) than and t,u,v are the parametric variables. Rearranging, we can
we have found to be necessary for computer animation. Oth- write this as three linear equations in three unknowns.
ers[6] do not easily produce the collision points and normal a u +b v +c t = d
directions necessary if collision response is to be calculated. eu+fv+gt=h
VoxeI-based methods have also been used,J30] but would not iu+jv+kt=l
be appropriate for all applications. Finally, many collision where
detection algorithms are quite intricate and must deal with a =(P1,-Pax)+t(Vlx-Vax)
many special cases, which we wished to avoid for software
b = (Pz~- Pax) + t(Vz~- Vax)
engineering reasons. Two collision algorithms are discussed
here: the first is designed to test the interpenetration o f sur- C ~--V x
faces modeling flexible objects; the latter is designed to test d =Px -(Pax + tVox)
for interpenetration o f convex polyhedra. e =(Ply-Poy)+t(Vty-Voy)

2.1. Collision Detection for Flexible Surfaces f = (P2y- Coy) + t(V2y- roy)
Surfaces are modeled as a grid of points connected to g =-Vy
form triangles.[19] Collisions between surfaces axe detected h =ey - ( e o y +tVoy)
by testing for penetration of each vertex point through the
i = (P l , - P a x ) + t ( V l , - Vo,)
planes of any triangle not including that vertex (thus, self-
intersection of surfaces is detected). The surfaces are assumed J = ( P ~ - Po,) + t ( V ~ - Vo,)
to be initially separate. For each time step of animation, the k =-V z
positions of points at the beginning and the end of the time t = P , -(P0= +tVoD
step must be compared to see if any point went through a tri-
angle during that time step. If so, a collision has occurred. The Pw s and Vws are the position and velocity components of
The algorithm is O (nm) for n triangles and m points. the point, and the Piw S and Viw s are the position and velocity
components of the triangle vertices. The velocities are per
A correct test must consider edges and triangles, as po-
lyhedral objects can collide edge-on without any vertices be- time step.
ing directly involved. However, in many cases merely testing The linear system above can be solved for t and expand-
points versus triangles produces acceptable results. This algo- edto
rithm only tests points versus triangles. It is worth noting that 0 = a U a - / b ) ( h a - c a ) - a (da-ib)(ga --cc ) t
the mathematics for testing intersection of a moving point + (fa -eb )(ka-ic ) t - (la-id)ffa--eb)
with a fixed triangle is the same as for testing a fixed edge
Substitution of the actual expressions for a through l gives a
versus a fixed triangle. Thus the fully general edge versus tri-
5'th order polynomial in t. If further substitutions were made,
angle tests could be done at fixed instants in time, with the
the equations could be written in the form
same advantages and disadvantages that will be discussed for
the second collision detection algorithm. C 5t5 .l- c 4t4 -1-c 3t3 -I-c2t2-I-Cl tl -I- Co= O
The question of whether a moving point has intersected a Polynomials of order 5 and above cannot be solved
surface can be divided into two cases. The easy ease requires analytically,[10] so a binary search technique is used to find
the surface to be fixed in space, whereas the hard case allows approximate values for t.[5] Binary search is used because it
the surface to be moving also. When the surface triangle is is guaranteed to converge, and because, using economizing
fixed, the parametric vector equation techniques described below, this algorithm is not used often
enough to warrant large efforts at optimization. The interval
P + (P" - P ) t = Po + (Pi-Po) u + (P2---Po) v
from t=0 to t=l is subdivided into a number of sub-intervals,
where P and U a.re the beginning and ending positions of the and the left-hand side is evaluated at each dividing point. If
point and the Pi's define the triangle, is set up and solved for the sign of that value is different for the two endpoints of
the variables u,v,t, u and v are parametric variables for the some subinterval, then some t for which the equation is true
plane defined by the triangle, whereas t is a time variable must lie within that interval. A binary search of values of t
which is 0 at the beginning of the simulation step in question, within that interval brings the brackets around that value of t
and 1 at the end. The left hand side is the parametric equation closer together, until a limit is reached (after 10 iterations, in
for the path of the point, and the right hand side is the our system) and an approximate value of t is found. Each
parametric equation for any point o n the plane. This vector value of t thus arrived at is used to get values for u and v by
equation represents three scalar equations in three unknowns back substitution, and then the standard 0<_t<_l and u~0 and
and is solved by matrix inversion. If 0,St_<l and u:20 and v:20 v~0 and u + v ~ l test is used to determine whether a collision
and u+v_<l, then the point has intersected the triangle during has occurred.

290
 ~)' Computer Graphics, Volume 22, Number 4, August 1988

To minimize the cost of executing the above calcula- Figure 1 illustrates a two-dimensional version o f this pro-
tions, a preliminary step is used. Every point is compared to cedure. The points A through 1 were inserted into an initially
every triangle. The perpendicular distance of the point from empty quadtree in alphabetical order, so that A is the root ele-
the plane defined by the triangle is first derived, by substitut- ment of the tree. The tree is to be searched for all points inside
ing into the plane equation,[25] for the beginning and the end the dotted box. A is inside, so all of it's subtrees must be
of the time step. If the sign of the perpendicular distance has searched. B is above and to the right of the dotted box, so only
not changed intersection is assumed not to have occurred. If its lower left subtree must be searched. This finds C, which is
the sign has changed, then the more expensive tests outlined inside the box. If C had subtrees, they would all have to be
above must be done, but in practice this test eliminates most searched. The next subtree of A starts with F. F is above the
point-triangle pairs. bounding box, so both of its lower subtrees must be searched.
A special kind of bounding box can also be used to One is empty, and the other contains only I, which is also out-
minimize computation. This bounding box includes the be- side the box. A ' s third subtree contains only E, which is below
ginning and ending position o f the triangle. This box is then the box. If E had subtrees, only the upper ones would need to
grown by the distance between the beginning and ending posi- be searched. A ' s fourth and last subtree contains only G,
tions of the point being tested. (This is necessary to avoid the which is outside the box. If G had subtrees, only the two fight
point passing unnoticed completely through the box during the hand ones would be searched. A large, bushy quadtree would
time step. A similar growth technique is used in the Lozano- be very fast to search (if the dotted box were small relative to
Perez path planning algorithm.)[15] the area covered by the tree) because the unsearched subtrees
would often contain large numbers o f points.
The basic algorithm is O (nm), for n triangles and m
points. Use of an octree[19] and bounding boxes can reduce
2.2. Collision Detection for Convex P o l y h e d r a
the time to O (mlogm) to construct the octree, and O (n logm) to
search it (assuming that the tree is almost balanced and that The detection of collisions between solids (or closed sur-
the bounding boxes are small compared to the space covered faces) can be treated somewhat differently, for the objects
by the tree). have a distinguishable inside and outside. The problem is
somewhat more complex than might be initially thought.
The search finds all point - bounding box pairs that must Edges as well as vertex points may be involved in collisions.
be examined more closely for possible intersection. All of the This method for detecting collisions is based on the
points in the model are inserted into an octree, which is creat- Cyrus - Beck clipping algorithm.[25] Collisions of articulated
ed anew for each round of collision detection. This octree is objects can be detected by applying this algorithm to all pairs
based on the points themselves, with each point P having up of the polyhedra making up the two objects. The two polyhe-
to 8 subtrees containing points in each of the octants of space dra are assumed to be convex; concave polyhedra can be
defined by the P's position. This is an obvious generalization decomposed into collections of convex ones. The basic algo-
of the well known binary search tree.[13, 14] A pseudo- rithm is O(n2m 2) for n polyhedra and m vertices per po-
random number generator is used to scramble the order of lyhedron. Methods for reducing these exponents are discussed
insertion; in this way, Knuth assures us, J13] the tree will be al- below.
most balanced, i.e. the height of the octree will be O (logm) al-
most always. The two-dimensional Cyrus - Beck algorithm[25] tells
whether a point is inside a convex polygon. It takes the dot
Each triangle's bounding box is grown by the distance product of each side's outward normal vector (n) with a vector
between the starting and ending positions of the fastest point from some point (v) on the side to the point in question (p). If
being tested. Each bounding box is then recursively compared
that dot product is negative for all edges of the polygon, then
against the octree to find the points inside it. If a point is in- the point is inside; if not, it is outside (see Figure 2).
side the box, all of its subtrees must be searched recursively. If
a point is outside the box, at least half of its subtrees do not
need to be searched. If a point is found to be inside a box, then dot pro.duct
the algorithm above must be run to determine if the point in- negative,
tersected the associated triangle during the time step.
~ p

r. . . . . . . . . . . . . . . . . D n l
'

Figure 2
Cyrus - Beck Clipping
Figure 1
Searching a Quadtree

291
¢ SIGGRAPH '88, Atlanta, August 1-5, 1988

The collision detection algorithm is developed as a This will result in a collection of intersection points P ly-
three-dimensional analogy to Cyrus - Beck clipping. The al- ing along the edge. Intersection points with t < 0 or t > 1 do
gorithm works by testing whether representative points of one not lie on the actual edge and are discarded. The remaining in-
polyhedron are inside the other polyhedron. First points from tersections are sorted into order according to their t values,
polyhedron B are tested against polyhedron A, and then the forming a sequence of points from one vertex to the other
process is reversed and points from A are tested for inclusion along the edge. Each adjacent pair o f points in this sequence,
in B. These two steps combine to cover all special cases and including those made by the vertices and the first and last sub-
give a reliable answer. The algorithm given below terminates division points, defines a sub-segment of the edge. The mid-
when a single point o f interpenetration is found, which is point of each resulting line segment is checked for being in-
sufficient for collision detection. If collision response is also side A by the same method that was used for vertices, above.
required, the algorithm below should be modified to find all Again, if any of these midpoints is inside A the algorithm ter-
points of interpenetration. The rest of this section describes minates with a detected collision.
the test of points from B against A. The third step tests for the infrequent case where two
Let A consist of a set of planar polygonal faces (Pi). Each identical polyhedra are moving through each other with faces
polygon contains a set o f vertices (uii) and an outward point- perfectly aligned. Here, the centroid point of each face of B is
ing normal vector ni. Let B consist of a set of vertices (vD, a tested against A by the method used for vertices, above. If
set of edges (el), and a set of planar polygonal faces ffi). All any of these centroids is inside A the algorithm terminates
coordinates of B have been transformed into the reference with a detected collision.
frame of A. If the algorithm survives the above three steps without
The first step tests for the presence of vertices of B inside detecting a collision, and also does not detect one when rev-
of A. Each vertex of B is compared to every face of A ; if any ersing and comparing A against B, then the two polyhedra do
vertex is on the inward side of all such faces, it is inside A and not interpenetrate.
the algorithm terminates having detected a collision. For each The above algorithm can be speeded up by a variety of
vertex i of B and for each face j of A, form the dot product tricks. A bounding box or bounding sphere test can be applied
(vi-ujl) • nj. If this dot product is negative the vertex is on the to every pair of polyhedraa, yielding an immediate "no colli-
inward side of the face. sion" result in most cases. Many of these bounding box tests
The second step tests for penetration of the edges of B can even be eliminated by octree or voxel methods. [4,9]
through the faces of A. Each edge of B is divided into a When a point is to be tested against a polyhedron, it chn first
number o f smaller line segments by intersecting it with the be compared to the polyhedron's bounding box, which will
infinite planes corresponding to every face o f A . See Figure 3. probably eliminate the need to compare it against all of the
This subdivision is done as follows. Let some edge of B con- faces. The bounding box can be aligned with the coordinate
nect the vertices vi and vy, and let us compare it against some axes of the polyhedron's local frame to make this point elimi-
face of A that has an outward pointing normal n, and a vertex nation test particularly fast.
point u~l. First the perpendicular distance of each vertex from It should be noted that this algorithm, or indeed any algo-
the plane defining the face is calculated, by substitution into rithm which point samples the positions o f objects over time,
the plane equation.[25] If the perpendicular distances differ in could fail if one object moved entirely through another during
sign, then the edge intersects the plane, and the intersection a single time step. This is a rather unusual occurence in pro-
point P can be calculated. cedural or dynamic animation because simulation time steps
are normally small relative to the velocities of the objects. The
di = (vi - U k l ) " nk
correct solution to this problem is to generalize to four dimen-
a j = (vj - a k ~ ) ' n , sions;J3] the starting and ending positions of the polyhedra
td~t define 4-D hyper-polyhedra which are checked for interpene-
t-- tration by higher-dimensional analogues to the algorithm
mail + IdyE
given above. The more practical approach is either to ignore
P =v~ + t (vy -v~) the problem (as we do) or to restrict the animation step size so
that the change in any object's position in any step is small re-
lative to the object's size.

V2 3. C O L L I S I O N R E S P O N S E
In keyframed and procedural animation systems, colli-
sion detection is the main requirement; collision response usu-
ally consists of informing the animator or the motion control
program that a collision has occurred, and trusting them to
handle it. In animation systems using dynamics to generate
planes seen edge-on motion, the system itself must respond to a collision by deter-
mining new linear and angular velocities for the colliding ob-
jects. These new velocities must conserve linear and angular
momentum, or else the resulting "funny bounce" will be very
Figure 3 obvious to viewers of the animation. The elasticity of the sur-
faces must also be taken into account, as this determines how
Edge Subdivision
much kinetic energy is lost in the collision; no-one will be-

292
 @ * ComputerGraphics,Volume22, Number4, August1988

lieve that a bean bag should bounce off of a hard surface as if This section develops the solution in stages. First, an
it were a golf ball. analytical solution for the collision of two rigid bodies is
presented; this result is due to MacMillan.[17] MacMillan's
3.1. Collision Response Using Springs solution is extended to tree-like articulated rigid objects with
revolute joints. Then, the restriction to wee-like objects is re-
The most intuitive way to handle collisions is with
moved, and finally the method is extended to encompass joints
springs. Dynamic simulation systems must already have a
with one or two sliding degrees of freedom.
method for applying external forces to objects. Thus, when a
collision is detected, a very stiff spring is temporarily inserted
between the points of closest approach (or deepest interpene-
3.2.1. Single Rigid Bodies
tration) of the two objects. The spring law is usually K / d , or MacMillan gives a general solution for the collision of
some other functional form that goes to infinity as the separa- two arbitrary rigid objects. Each object has a linear velocity
tion d of the two objects approaches 0 (or the interpenetration vector vi, an angular velocity vector toi, a mass m i, a center of
depth approaches some small value). K is a spring constant mass vector cl, and an inertial tensor matrix Ii which is rela-
controlling the stiffness o f the spring. The spring force is ap- tive to the center of mass. All of these quantities, for both ob-
plied equally and in opposite directions to the two colliding jects, are expressed in the same inertial reference frame. In
objects. The direction of the force is such as to push the two addition, each object has a vector Pi which points from its
objects apart (or to reduce their depth of interpenetration). center of mass to the collision point~ The solution also re-
quires three orthogonal unit vectors i ,j,k that define the "colli-
Our particular implementation handles variable elasticity sion frame", k will be perpendicular to the plane of collision
by making a distinction between collisions where the objects and i and j will be in that plane. The definition of the plane of
are approaching each other and collisions where the objects collision is somewhat arbitrary; for convenience we will
are receding from each other. For e = 1, i.e. perfectly elastic define it as follows. If a vertex of one object is colliding with
(hard) collisions, the spring constant K will be the same a face of the other, then that face defines the plane of collision.
whether the objects are approaching or receding. For e = 0, i.e. If an edge of one object is colliding with an edge of the other,
totally inelastic (soft) collisions, the spring will act as noted these two'edges define the plane of collision. If two vertices
above as long as the objects are approaching each other, but as are colliding, k is directed along the line joining them. See
soon as they start to move apart the spring force will decrease Figure 4.
to 0. For elasticities between 0 and 1, the two spring constants
will be related by K,e,,d~ = e Kot,p r ~ h .
The spring method is easy to understand and easy to pro-
gram. It applies equally well to rigid bodies (articulated or
not) and to flexible bodies, whether modeled as point masses
connected by springs, or by energy of deformation tech-
niques.[29] The main problem with this method is that it is
computationally expensive; stiffer springs mean stiffer equa-
tions, which require smaller time steps for accurate numerical
integration. [8] The numerical effort required goes up with the
"q ot
violence of the collision; as the springs are compressed more
and more, the equations become stiffer and stiffer, and smaller
and smaller time steps are needed. This was the motivation for
seeking a better method o f collision response.

3.2. Collision Response Using an Analytical Solution

An analytical solution for the collision of two arbitrary C2
articulated rigid objects is available. The analytical solution
depends upon the conservation of momentum during a colli- Figure 4.
sion, and results in a new angular and linear velocity for each Collision Problem - Two Rigid Objects
body. Thus, the solution bypasses the question of collision
forces and can be used independent of dynamic simulation, as- It is desirable to assume that there is only one collision
surning information concerning the bodies mass and mass dis- point in any given collision; this restriction is not totally
tribution can be provided. necessary, but it simplifies the formulations given below. It is
Some combination o f spring and analytical collision reasonable to say that whenever two objects collide in the real
response may be desirable for a dynamic animation system. world, there is one point at which they collide first (other col-
Analytical solutions are typically faster for strong collisions, lisions may follow within microseconds). Thus, the collision
because the solution need only be found once. However, for detection algorithm must furnish a single collision point
gentle collisions, such as a body resting quietly atop another between two objects. Because of the time-stepped nature of
body, springs may be desirable.[26] In such a case, gravity dynamics simulations, this will only be an approximate colli-
may eonsistently cause the two objects to interpenetrate and, sion point; a good heuristic is to take the point of greatest in-
thus, the analytical solution would have to be applied time and terpenetration o f any two objects in the simulation, provided
time again. A simple spring that counteracts gravity will be that the relative velocities of the two objects at that common
faster and more stable in this case. point are such that the interpenetration depth is increasing. If
adaptive step size control is available, this heuristic can be

293
 SIGGRAPH '88, Atlanta, August 1-5, 1988
i i

refined by stating that interpenetration to greater than some point. This corresponds to the vector equation:
threshold depth iz unacceptable, and causes backtracking and
V2+~xp2-~l-~ lxpl=0
reduction of the step size. This allows the simulation to close
in on a collision point very close to the surfaces of the objects In between perfectly smooth and perfectly rough colli-
by a process similar to binary search. Multiple collision sions lies the great middle ground of partially rough friction.
points can be handled by a straightforward extension to the al- Modeling partial (i.e. realistic) friction can become quite com-
gorithms given below, by inventing multiple collision im- plex; the simple treatment given here is from MacMillan[17]
pulses and incorporating them into the matrix. and McLean[18] and is sufficient to produce visually reason-
able results.
The solution involves solving a set of 15 linear equations
in 15 unknowns. The fifteen unknowns are: the new linear The coefficient of friction, 7, is the maximum allowed ra-
velocity vector for each object (~1, V2); the new angular velo- tio of force parallel to the collision plane versus force perpen-
city vector for each object (c01, o2); and the impulse vector R. dicular to that plane. Although properly speaking, y is a pro-
An impulse has units of momentum and can be thought of as a perty of pairs of surfaces, we assign a y value to each surface,
huge force applied for a tiny time. Because the collision is as- and then use the larger of the 77values of the colliding objects.
sumed to occur in a negligible time (approximately instan- When the two objects have finite 77and e = 0, the collision can
taneous), only the collision impulse itself matters; any other be solved in two steps. First the collision is solved as if it were
forces being applied to the objects will be too small to have an infinitely rough. Then the resulting collision impulse, R, is
effect. By convention, the impulse is directed from object 2 to examined. If the allowed ratio, 77, of the components of R
object 1. parallel and perpendicular to the collision plane is not exceed-
Twelve linear equations can be written down immediate- ed (i.e. if77R, k ~ I R - k ( R - k ) I), all is well and the solution
ly, expressing the change in linear and angular momentum stands, because the objects should stick.
that each object experiences as a result of the collision im- Otherwise, the objects should slide. The system of equa-
pulse R. tions must be set up and solved again with different collision
conditions. These new conditions will give a smaller restrain-
rnlv- 1 = r n l v 1 + R
ing parallel force, because only a limited amount of friction
ra2V2 = rnxvz - R earl act against sliding motion. Two constants ct and 13are cal-
culated, such that the collision impulse will exactly fulfill
I101 =ll001 + Pl ×R
77R - k = I R - k (R • k) I, or in other words such that the ratio
12~2 = 12002- Pa × R of the parallel and perpendicular components of R is exactly y,
and the direction of the parallel component of R is the same as
The last three linear equations come from some assump-
before. This gives the maximum parallel force allowed by the
tions about the collision conditions; the assumptions that we necessary perpendicular force and the coefficient of friction.
will use are that the elasticity, e, is zero (so that the two collid- The collision conditions axe then:
ing objects come to rest relative to each other, at least at the
collision point) and that the surfaces are frictionless (so that R -i =txR -k
the impulse must be perpendicular to the collision plane). Oth- g -j = ~ R -k
er assumptions are possible and are discussed below. Our as-
sumptions require the dot products of R with the collision (V2 + ~2 x p2 - Vl - ~ l x pl) " k = O
frame unit vectors i and j to be zero, and the difference in the ct and ~ are calculated as follows, with Q the component
velocity of the collision point, as seen from each of the two of R perpendicular to the collision plane, and P the unit direc-
objects, to be zero in the k direction. We can write: tion vector of the component of R parallel to the collision
R.i=O plane:
R "j=0 Q=k(R .k)

(V2 + 02 x p 2 - V l - ~ l x pl)"k =0 R-Q

P
tR-Qt
These equations can be solved by standard Gauss-Jordan ct = 77(P . i )
elimination with maximal pivoting,[5] LU-decomposition,[22]
or by more advanced sparse matrix methods.[20, 21] It is pos- ~=77 (P - j )
sible at this point in the algorithm to find the solution for an To reiterate, the full algorithm for solving a general colli-
elastic collision. The actual elasticity of the collision can be sion of two rigid objects is to transform the required quantities
taken as the lower of the elasticities of the two colliding sur- (incoming velocities, tensor matrices, Pi, etc) from the ob-
faces. A new collision impulse Rae~l Can then be calculated jects' local coordinate frames to a common inertial frame,
as Ractual = ( l + e a a u a t ) R . This new collision impulse is then define the collision frame's orthogonal unit vectors i , j , and k,
plugged back into the defining equations above, to solve for choose appropriate collision conditions, set up and solve the
the ~ and ~i vectors that are required. The ~ vectors come system of equations as outlined above, and transform the new
out easily; the oi vectors require inverting the I i inertial tensor linear and angular velocities back to the objects' local frames.
matrices. This may seem like a drastic amount of work when compared
Next consider including friction. If the objects axe with inserting a simple spring between the two objects, and it
infinitely rough and e = 0, the collision condition requires that does require more lines of computer code, but this method is
the objects come to rest (relative to each other) at the collision usually applied only once for each collision, whereas springs

294
 ~ ComputerGraphics, Volume 22, Number 4, August 1988

generally must be applied over a large number of very small colliding. Thus, the total size of the linear system to be solved
time steps. This analytical method is cheaper computationally is approximately 9n for n rigid objects involved in the coUi-
unless the collision is very gentle indeed, and the cost of this sion. The sparsity of the matrix increases as n increases, so
collision solution does not depend upon the violence of the that if sparse matrix methods are used the solution should be
collision, certainly a desirable property. around O (n).[20, 21]
Once again, the unknowns to be solved for are g. and ~i
3.2.2. A r t i c u l a t e d R i g i d B o d i e s - T r e e - S t r u c t u r e d , R e v o -
for i = 1...n, R, and Rq for all pairs of objects i and j connect-
lute Joints
ed by a joint. The equations for objects 1 and 2, and for the
Now we extend MacMillan's solution to tree-like articu- collision impulse, still look familiar. The extra summation
lated rigid objects with revolute joints. The various rigid ob- terms reflect the change in linear and angular momentum
jects that make up the tree-like linkages will be numbered resulting from any attachment impulses felt by those objects.
from 1 to n. Objects 1 and 2 will be the colliding objects, and
n
the rest will be linked to one or both of them, either directly or
m l ~ t = m l v l +R + ~ R l i
through some number o f intermediaries. Note that this solu- iffil
tion allows the links of an articulated object to collide with n
other links of the same object or with another object entirely. m2V2 =m2v 2 - R + ~ R ~
i=1
Each rigid object will again have a linear velocity vector v i , an
n
angular velocity vector toi, an inertial tensor mawix li, a mass 11~1 =I1091 +pl xR + ~Pli ×Rli
mi, and a center of mass ci, all expressed in a common inertial i=1
reference frame. n
12~2 =12(.022-- P2 × R + ~ 92i ~R2i
The revolute joints connecting the various rigid objects i=1
will be assumed to be ideal, that is, perfectly elastic and with
no mechanical tolerance. The single joint that connects object The conditions on the collision impulse R are still the
i to object j will be described by the vector Pij that points same.
from c i to the joint, and the vector Pji that points from cj to R "i = 0
the joint. As well as the collision impulse R, this solution will
calculate an attachment impulse Rij for each joint. Unlike the R 'j=0
collision impulse, the attachment impulses Rij are uncon- (V2+~xp2-VI-~I x Pl)" k = 0
strained as to direction. By convention, the attachment im-
pulse Rij points from object j to object i, and Rij =-Rji. Rij For the objects that are not directly colliding (for object
will be (0,0,0) if objects i and j are not connected by a joint. i = 3..n ), the momentum conservation equations are
n
See Figure 5.
rrti~. =miv i + ~Rij
j=l
n
lif-Oi =liOOi + ~Pij ×Rij
j=t
Each joint requires three more linear equations to make
the system of equations complete and solvable. These are
derived from the basic requirement of a revolute joint: the
velocity of the joining point, when seen as part of either of the
rigid objects which it connects, must be the same. Otherwise,
the joint would tend to pull apart. For each joint connecting
objects i and j , three more equations can be written.
. . . . collision
~',,,,. point ~ +~ xpu =~ +~j xp~
Once again, the algorithm requires that the necessary in-
formation about all of the objects be transformed from their
local reference frames to a common inertial reference frame.
The collision frame orthogonal unit vectors i, j , and k must
be determined. The (potentially rather large) linear system is
set up and solved for the variables V/, 0~i, R, and Rij, by stan-
Figure 5. dard methods.[5, 22, 20, 21] The actual elasticity of the colli-
Articulated Collision Problem sion is determined as above, and actual impulses are deter-
mined by multiplying R and the R q ' s by (1 + I~aa.al). The ac-
tual impulses are then put back into the equations above to get
For a collision involving n rigid objects there are 6n the final solution for linear and angular velocities. The last
unknowns corresponding to the resulting linear and angular step is to transform the solution back to the object's local
velocities of the objects, 3 unknowns for the collision impulse, frames.
and either 3 ( n - l ) unknowns corresponding to the attachment
impulses if the objects are all part of one articulated linkage,
or 3(n-2) unknowns if two different articulated objects are

295
 SIGGRAPH '88, Atlanta, August 1-5, 1988

3.2.3. Articulated Rigid Bodies - Revolute Joints RU "di = 0

The above solution for tree-like articulated rigid objects (vj +~j XPji -v'i - ~ j Xpij)'dj =0
can be extended by removing the requirement for tree-like
linkage. Since the two articulated objects that are colliding are (~ +~j xpj~-~ - ~ j xpv)-ak =o
defined to be connected objects, some subset of the attachment The argument for a planar sliding joint is similar. In this
points will define tree-like linkages. The first step is to set up case the attachment point on object j is allowed to move free-
the problem as above for the objects and for those joints. Then ly in the d~ and dj directions, constraining the collision im-
the extra joints are added; each contributes another attachment pulse to exactly the dk direction, but leaving its magnitude
impulse Rij, and thus adds three variables to the problem. unknown. The attachment point velocities must still match in
Each joint also allows three more linear equations to be writ- the dk direction, but are allowed to vary in the other two direc-
ten down, the familiar velocity matching condition. tions. The equations are as follows.
n~i .d~ =o
This larger system once again contains as many variables eli .dj =o
as equations, and so can be solved by standard techniques.
The actual collision and attachment impulses are calculated (~j + ~ x p j i - ~ - ~ j × P 0 ) ' dk = 0
and applied in the same way as above, and the solution values
The sliding joints described above allow the two objects
are transformed back to the objects' local coordinate systems.
that they connect either 4 (linear) or 5 (planar) degrees of free-
It is even possible to have more than one joint connecting two
dom of movement relative to each other. Sliding joints that
objects i and j , although in that case the notation used above
provide fewer degrees of freedom can be constructed by ad-
would have to be expanded slightly. If two joints connect ob-
ding one or more extra joints of the above types. For instance,
jects i and j , the objects will have only one degree of freedom
suppose that a planar sliding joint is desired such that one ob-
of motion relative to each other; they will be able to twist
ject can slide relative to the other, but the objects cannot rotate
around the line connecting the two joints. If a third joint is ad-
relative to each other. This can be accomplished by defining
ded connecting i and j , which is not colinear with the other
three planar sliding joints (all using the same plane) with the
two, the two objects will be locked in position relative to each
sliding points not colinear. A piston type joint (one degree of
other, and will form in effect a single rigid object. This is
translational freedom and one degree of rotational freedom)
probably not a desirable state of affairs, but the solution algo-
can be described by two linear sliding joints of the above type,
rithm does permit it.
with the connection points constrained to slide along the same
3.2.4. Articulated Bodies - Sliding Joints line.
A sliding joint is one in which the joining points on the
3.3. Collisions of Dynamic Objects with Non-Dynamic
two objects are allowed to move freely with respect to each
Objects
other in one or two dimensions, but are at the same time con-
strained to a fixed relationship in the other dimensions. A A complication seems to arise when dynamically con-
linear sliding joint allows sliding motion in one degree of free- trolled objects collide with objects that are controlled in other
dom, while controlling two others; a planar sliding joint con- ways (such as keyframe interpolation). In these cases the velo-
trois one degree of freedom, allowing sliding motion in the cities of the non-dynamic objects involved in the collision
other two. For now, assume that these joints allow three revo- cannot change. Thus, ~ =vi andtoi =o)i for the objects that
lute degrees of freedom as well. are not under dynamic control, and V/ and o01 are not variables
in the linear system formulations above. The systems can be
For a linear sliding joint, assume that objects i and j are
connected, and that a joint coordinate system is defined by reformulated with fewer rows and columns, and the solution
three orthogonal unit vectors di, dj, and dk. These vectors are proceeds just as before. The result is collisions that do not
stated in the local coordinate system of object i and rotate conserve linear or angular momentum. The keyframe or pro-
cedurally controlled objects move along their assigned paths
with it. Further assume that the attachment point on object j is
with lordly disdain, brushing aside the dynamically controlled
allowed to move freely in the d; direction, but must maintain
objects as if they had negligible mass.
some particular value for its position along dj and dk as seen
from object i. Note that the "attachment point" on object i is More complex responses from the non-dynamic objects
in fact a line; this requires us to calculate P0, the actual attach- are possible. Programs that control objects could be written to
ment point on object i, separately for each collision. Note also take the results of dynamic collisions into account. In effect,
that the joint coordinate system orthogonal unit vectors must the procedural object could become dynamic for the duration
be rotated into the common inertial frame. of the collision, and its velocity could change. The program
Notice that the attachment impulse R V is required to lie would have to be alert to this possibility. This would be fairly
in the dj, dk plane. As it is possible to write down three linear simple to implement using the analytic solution, which does
constraining equations about this joint, such a joint can be not require setting up and solving the complete dynamics
treated within the linear systems described above. One equa- equations of motion. Altematively, collisions between
tion expresses the constraint that Rij must lie in a plane; the dynamic and keyframed objects could be defined as excep-
other two equations constrain the attachment point velocities tional events that require that the human animator be notified.
to match in two of the three joint coordinate system directions.

296
 ~ ComputerGraphics, Volume 22, Number4, August 1988

4. CONCLUSIONS 11. J.E. Hopcroft, J.T. Schwartz, and M. Sharir, "Efficient

Collision detection is important for any animation sys- Detection of Intersections among Spheres," The Interna-
tem. The coding requirements are not excessive, and, while a tional Journal of Robotics Research, vol. 2:4, pp. 77-80,
naive approach to collision detection can consume large Winter 1983.
amounts of computer time, several tricks are available to keep 12. Paul M. Isaacs and Michael F. Cohen, "Controlling
the cost reasonable. Dynamie Simulation with Kinematic Constraints," Com-
Dynamical simulation systems must resolve collisions puter Graphics, vol. 21, no. 4. Proceedings of SIG-
after detecting them. The obvious method of inserting tem- GRAPH'87 (Anaheim, CA, July 27-31, 1987)
porary springs is general and easy to program, but exacts a 13. Donald Knuth, Fundamental Algorithms, Addison-
severe execution time penalty, particularly for violent colli- Wesley Publishing Co., Reading, MA, 1975.
sions. This makes an analytical collision resolution algorithm 14. Donald Knuth, Searching and Sorting, Addison-Wesley
attractive. On the other hand, for objects resting against each Publishing Co., Reading, MA, 1975.
other but encouraged by forces to interpenetrate, the spring
solution is more appropriate. A dynamical simulation system 15. Tomas Lozano-Perez and Michael A. Wesley, "An
Algorithm for Planning CoUision-Free Paths Among
should have a combination of both methods available.
Polyhedral Obstacles," Communications of ACM, vol.
ACKNOWLEDGEMENTS 22, no. 10, pp. 560-570, October, 1979.
This work was supported by National Science Founda- 16. Richard V. Lundin, "Motion Simulation," Proceedings
tion grant number CCR-8606519. We wish to thank Robert of Nicograph 1984, pp. 2-10, November, 1984.
Skinner, David Forsey, and Peter Valtin for contributing to the 17. William D. MacMillan, Dynamics of Rigid Bodies,
dynamical animation software that we used to implement Dover Publications, Inc, New York, 1936.
these algorithms. We would also like to thank our reviewers
18. W . G . McLean and E. W. Nelson, Engineering Mechan-
for their incisive and helpful comments and references.
ics: Statics and Dynamics, Shaum's Outline Series,
McGraw-Hill Book Co., New York, 1978.
References
19. Matthew Moore, " A Flexible Object Animation Sys-
1. William W. Armstrong and Mark W. Green, "The tem," Masters Thesis, University of California, Santa
Dynamics of Articulated Rigid Bodies for Purposes of Cruz, Computer & Information Sciences, Santa Cruz,
Animation," Proceedings of Graphics Interface '85, pp. California, March, 1988.
407-415, Canadian Information Processing Society,
20. Ole Osterby and Zahari Zlatev, Direct Methods for
Toronto, Ontario, Canada, May 1985.
Sparse Matrices, Springer-Verlag, Berlin, 1983.
2. John W. Boyse, "Interference Detection Among Solids
21. Sergio Pissanetsky, Sparse Matrix Technology,
and Surfaces," Communications of the ACM, vol. 22:1, Academic Press, London, 1984.
pp. 3-9, January, 1979.
22. William H. Press, Brian P. Flannery, Saul A. Teukolsky,
3. John Canny, "Collision Detection for Moving Polyhe- and William T. Vetterling, Numerical Recipes, Cam-
dra," MIT Aar. Lab Memo 806, October, 1984. bridge University Press, Cambridge, England, 1986.
4. Ingrid Carlbom, "An Algorithm for Geometric Set 23. Craig W. Reynolds, "Computer Animation with Scripts
Operations Using Cellular Subdivision Techniques," and Actors," Computer Graphics, vol. 16, no. 4, pp.
IEEE Computer Graphics and Applications, vol. 7, pp. 289-296, Association for Computing Machinery, July,
44-55, Computer Society of the IEEE, Los Alamitos, 1982. Proceedings of SIGGRAPH'82
CA, May 1987. 24. Craig W. Reynolds, "Flocks, Herds, and Schools: A Dis-
5. Brice Carnahan and James O. Wilkes, Digital Computing tributed Behavioral Model," Computer Graphics, vol.
and Numerical Methods, John Wiley and Sons, Inc., New 21, no. 4, pp. 25-34, Association for Computing
York, 1973. Machinery. Proceedings of SIGGRAPH'87 (Anaheim,
6. Scott E. Fahlman, " A Planning System for Robot Con- CA, July 27-31, 1987)
struetion Tasks," Artificial Intelligence, vol. 5, pp. 1-49, 25. David F. Rogers, Procedural Elements for Computer
1974. Graphics, McGraw-HiU Book Company, New York,
7. Wm. Randolph Franklin, "Efficient Polyhedron Intersec- 1985.
tion and Union," Proceedings of Graphics Interface 26. Robert Skinner, U Cal. Santa Cruz, CIS Dept. personal
1982, pp. 73-80, 1982. communication.
8. C. William Gear, Numerical Initial Value Problems in 27. Scott N. Steketee and Norman I. Badler, "Parametric
Ordinary Differential Equations, Prentice-Hall, Engle- Keyframe Interpolation Incorporating Kinetic Adjust-
wood Cliffs, NJ, 1971. ment and Phrasing Control," Proceedings of SIGGRAPH
9. Jeffrey Goldsmith and John Salmon, "Automatic Crea- '85, vol. 19, no. 4, pp. 255-262, July, 1985.
tion of Object Hierarchies for Ray Tracing," IEEE Com- 28. David Sturman, A Discussion on the Development of
puter Graphics and Applications, vol. 7, pp. 14-20, Com- Motion Control Systems, Association for Computing
puter Society of the IEEE, Los Alamitos, CA, May 1987. Machinery, July 1987. SIGgraph '87 Course 10 Notes:
10. I.N. Herstein, Topics in Algebra, Xerox College Publish- Computer Animation: 3-D Motion Specification and
ing, Lexington, MA, 1964. Control.

297
 SIGGRAPH '88, Atlanta, August 1-5, 1988

29. Demetri Terzopoulous, John Platt, Alan H. Barr, and

Kurt Fleischer, "Elastically Deformable Models," Com-
puter Graphics, vol. 21, no. 4. Proceedings of SIG-
GRAPH'87 (Anaheim, CA, July 27-31, 1987)
30. Tetsuya Uchild, Toshiaki Ohashi, and Mario Tokoro,
"Collision Detection in Motion Simulation," Computers
& Graphics, vol. 7:3-4, pp. 285-293, 1983.
31. Jane Wilhelms, "'Towards Automatic Motion Control,"
IEEE Computer Graphics and Animation April, 1987,
vol. 7, no. 4, pp. 11-22, April, 1987.
32. Jane Wilhelms, "Using Dynamic Analysis for Anima-
tion of Articulated Bodies," IEEE Computer Graphics
and Applications, vol. 7, no. 6, June, 1987.

Example 1: early stage rockpile

Example 2: man sitting Examples 3-5: foiling domino movie

positioned with collision detection

298