This is a rough but fast implementation of GJK collision detection algorithm in plain C. Just one C file, less than 200 lines, no dependencies. It is in 2D for now, full 3D version is upcoming... This 2D-version uses Minkowski sums and builds a triangle-simplex in Minkowski space to tell if two arbitrary convex polygons are colliding. 3D-version will be roughly the same, but will build a tetrahedron-simplex inside a 3-dimensional Minkowski space. It currently only tells if there is a collision or not. C-code for distance and contact points coming soon.

Fuck licenses and copyright. I made it for learning purposes, this is public knowledge and it's absolutely free for any usage.

This is an illustration of the example case from dyn4j.

The two tested polygons are defined as arrays of plain C vector struct type. This implementation of GJK doesn't really care about the order of the vertices in the arrays, as it treats all sets of points as *convex polygons*.

```
struct _vec2 { float x; float y; };
typedef struct _vec2 vec2;
...
int main(int argc, const char * argv[]) {
// test case from dyn4j
vec2 vertices1[] = {
{ 4, 11 },
{ 4, 5 },
{ 9, 9 },
};
vec2 vertices2[] = {
{ 5, 7 },
{ 7, 3 },
{ 10, 2 },
{ 12, 7 },
};
size_t count1 = sizeof (vertices1) / sizeof (vec2); // == 3
size_t count2 = sizeof (vertices2) / sizeof (vec2); // == 4
int collisionDetected = gjk (vertices1, count1, vertices2, count2);
printf (collisionDetected ? "Bodies collide!\n" : "No collision\n");
return 0;
}
```

The goal of this explanation is to help people visualize the logic of GJK. To explain it we have to oversimplify certain things. There's no complicated math except basic arithmetic and a little bit of vector math, nothing more, so that a 3rd-grader could understand it. It is actually not very difficult to have GJK algorithm explained in a proper understandable way from ground up.

At the very top level GJK tells if two arbitrary shapes are intersecting (colliding) or not. The algorithm is used to calculate the depth of intersection (collision distance). A collision occurs when two shapes try to occupy the same points in space at the same time. The space can be of any nature. It might be your in-game world simulation, or a calculation on a table of statistical data, or a built-in navigation system for a robot or any other application you can imagine. You can use it for calculating collisions of solid bodies and numeric intersections of any kind. You can have as many dimensions as you want, the amount of dimensions does not really matter, the logic is the same for 1D, 2D, 3D, etc... With GJK you can even calculate collisions in 4D if you're able to comprehend this.

The algorithm itself does not require any hard math whatsoever, it's actually very intuitive and easy. It takes an arithmetic difference of two shapes by subtracting all points of one shape from all points of another shape. If two shapes share a common point, subtracting that point from itself results in a difference of zero. So, if zero is found in the result then there's a collision. In fact, the algorithm does not have to calculate all differences for all possible pairs of points, only a small subset of significant points is examined, therefore it works very fast while being very accurate.

In order to understand GJK one has to build an imaginary visualization of what is going on under the hood. Once you see the picture in your head, you can implement it and even tweak it for your needs.

Let's start with a naive example of computing a difference of two shapes in one dimension. A segment of a number line is an example of 1D-shape. Imagine we have two segments on the number line: segment `[1,3]`

and segment `[2,4]`

:

Zero is our point of reference on the number line, so we call it *the Origin*. Easy enough.
It is obvious that our segments occupy some common region of our 1D-space, so they must be intersecting or colliding, you can tell that just by looking at the representation of both segments in same 1D-space (on the same line). Let's confirm arithmetically that these segments indeed intersect. We do that by subtracting all points of segment `[2,4]`

from all points of segment `[1,3]`

to see what we get on a number line.

```
1 - 2 = -1
1 - 3 = -2
1 - 4 = -3
2 - 2 = 0
2 - 3 = -1
2 - 4 = -2
3 - 2 = 1
3 - 3 = 0
3 - 4 = -1
```

We got a lot of numbers, many of them more than once. The resulting set of points (numbers) is larger than each of our initial sets of points of two shapes. Let's plot these resulting points in our 1D-space and look at the shape of resulting segment on the number line:

So after all we have resulting segment `[-3,1]`

that covers points `-3`

, `-2`

, `-1`

, `0`

and `1`

. Because two initial shapes had some points in common the resulting segment contains a zero. This comes from a simple fact, that when you subtract a point (which is a number) from itself you inevitably end up with a zero. Note, that our initial segments had points 2 and 3 in common. When we subtracted 2 from 2 we got 0. When we subtracted 3 from 3 we also got a zero. This is quite obvious. So, if two shapes have at least one common point, because you subtract that point from itself, the resulting set must contain zero (the Origin) at least once. This is the key of GJK which says: if the Origin is contained inside the resulting set then initial shapes must have collided or kissed at least. Once you get it, you can then apply it to any number of dimensions.

Now let's take a look at a counter-example, say we have two segments `[-2,-1]`

and `[1,3]`

:

We can visually ensure that segment `[-2,-1]`

occupies a different region of number line than that of segment `[1,3]`

. There's a gap between our initial shapes, so these two shapes do not intersect in our 1D-space, therefore there's no collision. Let's prove that arithmetically by subtracting all points of any of the two segments from all points of the other.

```
1 - (-1) = 1 + 1 = 2
1 - (-2) = 1 + 2 = 3
2 - (-1) = 2 + 1 = 3
2 - (-2) = 2 + 2 = 4
3 - (-1) = 3 + 1 = 4
3 - (-2) = 3 + 2 = 5
```

And we again draw the resulting segment on a number line in our imaginary 1D-space:

We got another bigger segment `[2,5]`

that represents a difference of all the points of two initial segments but this time it does not contain the Origin. That is, the resulting set of points does not include zero, because initial segments did not have any points in common so they indeed occupy different regions of our number line and don't intersect.

Now, if our initial shapes were too big (long initial segments) we would have to calculate too many differences from too many pairs of points. But it's actually easy to see, that we only need to calculate the intersection between the endpoints of two segments, ignoring all the inside points of both segments.

Consider two segments `[10,20]`

and `[5,40]`

:

Now subtract their four endpoints from each other:

```
10 - 5 = 5
10 - 40 = -30
20 - 5 = 15
20 - 40 = -20
```

The resulting segment `[-30,15]`

would look like this:

Notice that when we subtract the *opposite points* of initial shapes from one another, the resulting difference number lands either to the left or to the right of the Origin (which is zero) on the number line. Say, we take the leftmost point of segment `[5,40]`

(number `5`

) and subtract it from the rightmost point of segment `[10,20]`

(number `20`

) the resulting number `20 - 5 = 15`

is positive, so it lands to the right of zero on the number line. Then we take the rightmost point of segment `[5,40]`

(number `40`

this time) and subtract it from the leftmost point of segment `[10,20]`

(number `10`

) the resulting number `10 - 40 = -30`

is negative, so it lands to the left of zero on the number line that is exactly opposite to the first positive difference point `15`

. We say these points are *opposite* directionwise. Therefore the Origin is between two resulting points *enclosing* it.

We ignored all insignificant internal points and only took the endpoints of initial segments into account thus reducing our calculation to four basic arithmetic operations (subtractions). We did that by switching to a simpler representation of a segment (only two endpoints instead of all points contained inside an initial segment). A simpler representation of the difference of two shapes is called a *simplex*. It literally means 'the simplest possible'. The simplest possible *thing* on a one-dimensional number-line is a number (a single point in 1D-space is called a *0-simplex*). A segment is the simplest possible shape sufficient to contain multiple points of a number line (a segment of two points in 1D space is called a *1-simplex*). Even if one segment covers the other segment in its entirety (a segment fully contains another segment) – you can still detect an intersection of them in space. And it does not matter which one you're subtracting from, the resulting set will still contain the Origin at zero.

GJK also works if two initial shapes don't intersect but just barely touch. Say, we have two shapes, segment `[1,2]`

and segment `[2,3]`

:

These two segments only have one point in common. Subtracting their endpoints from each other gives:

```
1 - 2 = -1
1 - 3 = -2
2 - 2 = 0
2 - 3 = -1
```

And the resulting difference looks like:

Notice, that in case of only one common point the Origin is not inside resulting segment, but is actually one of its endpoints (the rightmost in this example). When your resulting segment has the Origin at one of its endpoints that means that your initial shapes do not collide, but merely touch at a single point (two segments have only one common point of intersection).

What GJK really says is: if you're able to build a simplex that contains (includes) the Origin then your shapes have at least one or more points of intersection (occupy same points in space).

WORK IN PROGRESS, A live demo of GJK in a 1D-space and a video of GJK in action coming up soon )

Now let's take a look at the picture in 2D. Our 2D-space is now an xy-plane (which is represented by two orthogonal number lines instead of a single number line). Every point in our 2D-space now has two xy-coordinates instead of one number, that is, each point is now a 2D-vector. Suppose we have two basic 2D-shapes – a rectangle `ABCD`

intersecting a triangle `EFG`

on a plane.

These shapes are represented by the following sets of points (2D-vectors, which are pairs of xy-coordinates):

```
A (1, 3)
B (5, 3)
C (5, 1)
D (1, 1)
E (2, 4)
F (4, 4)
G (3, 2)
```

Now, because we have much more numbers here, the arithmetic becomes a little more involved, but it's still very easy – we literally subtract all points of one shape from all points of another shape one by one.

```
A - E = (1 - 2, 3 - 4) = (-1, -1)
A - F = (1 - 4, 3 - 4) = (-2, -1)
A - G = (1 - 3, 3 - 2) = (-2, 1)
B - E = (5 - 2, 3 - 4) = ( 3, -1)
B - F = (5 - 4, 3 - 4) = ( 1, -1)
B - G = (5 - 3, 3 - 2) = ( 2, 1)
C - E = (5 - 2, 1 - 4) = ( 3, -3)
C - F = (5 - 4, 1 - 4) = ( 1, -3)
C - G = (5 - 3, 1 - 2) = ( 2, -1)
D - E = (1 - 2, 1 - 4) = (-1, -3)
D - F = (1 - 4, 1 - 4) = (-3, -3)
D - G = (1 - 3, 1 - 2) = (-2, -1)
```

After plotting all of 12 resulting points in our 2D-space and connecting the *outermost* points with lines we get the following difference shape:

GJK says that if we're able to enclose the Origin within the resulting shape then two initial shapes must have collided.
We immediately see that this shape actually contains the Origin. Therefore we can visually confirm that our initial rectangle `ABCD`

indeed intersects our initial triangle `EFG`

.

Also note, that we subtracted all points from all other points (calculated all combinations of pairs) which yields some points inside our resulting shape. But as with 1D, we will optimise the algorithm to skip all internal points entirely.

Now, remember, in 1D to determine whether the resulting segment contains the Origin we check if a primitive inequality `leftEndpoint < Origin < rightEndpoint`

holds. We can think of it as if the segment *surrounds* the Origin from all sides of 1D space (from both left and right, as there are only 2 relative sides in one dimension on our sketch). In 1D for an object to be able to surround any point (the Origin is the zero point on a number line), that object must itself consist of at least two of its own points (in other words, it must be a segment defined by its two points on a number line). Therefore the 1D-version is trying to build a simplex of two endpoints (a segment on a number line). And then it checks whether the Origin is contained within the resulting segment.

In 2D the process is similar. We can immediately tell if the Origin is inside the resulting polygon shape just by a brief visual inspection. This is done by the natural wiring of human brains. But a machine cannot distinguish *inside* from *outside* unless you teach it how to do that. It's a lot easier for a machine to see if a point is contained inside a simple shape like a triangle or a circle, rather than a more complex shape like a general polygon with N vertices or some arbitrary irregular shape. Moreover, the algorithm has to do that in a fast but accurate way in order to be suitable for general-purpose collision detection.

And this is where the true power of simplicity of GJK comes into play. Think this way: you need at least two points in 1D to surround the Origin. So a 1D segment of two points is a 1-simplex. But in 2D two points are not enough to surround anything on a plane. Two points define a single straight line on a plane, but a straight line cannot enclose anything, because it's a line and it is straight. Therefore in 2D you need at least three points connected by segments, that is a triangle, to be able to enclose at least some area of the plane. The simplest possible shape that can *enclose* something in 2D is a triangle also known as *2-simplex*.

Now our task of arithmetically enclosing the Origin within our resulting shape simplifies a little bit. We have to find such three points from our resulting set of points, that make a triangle that encloses the Origin. In other words, we need to build a 2-simplex that would satisfy our criteria. GJK can build a triangle and test if a certain point lies within that triangle, so we just made this task solvable by a machine.

After we subtracted our initial shapes one from another we got a resulting set of all of the points of a new shape that represents the difference of initial shapes. The most straightforward way to build such an Origin-enclosing triangle from a given set of points is to start taking triples of points (combinations of three points) to see if they form a triangle with the Origin inside it. If a triple of points makes such a triangle, then we can conclude that the difference of two shapes contains the Origin, so initial shapes must have collided or intersected. If not, we try some other triple, and that is done in a loop until we run out of points. If none of the triples did enclose the Origin, then no such triangle was found and there was no collision at all.

So, if we randomly select any three of our points and connect them with line segments, we will probably end up with a triangle similar to one of the following:

All of these triangles satisfy our criteria (all of them contain the Origin), so if we randomly select one of these and find the Origin inside, then we can immediately tell that there was a collision. But if we selected a triple of points that does not contain the Origin we might end up with something like this:

As long as the choice of points for our 2-simplex is random we might have to try all possible triples in worst case. Random walk is not the best strategy for finding a triangle that satisfies our criteria, there's a better algorithm for that ;) The goal of the algorithm is to find a combination of the best three points that enclose the Origin out of all possible triples of points, if such a combination exists at all. If such a triple does not exist then there is some distance between our initial shapes. The search for a simplex is the core of GJK.

In order to build the simplex efficiently the algorithm introduces a special routine that calculates the difference of two points of initial shapes. It is called a *support function* and it is the workhorse of the search.

Remember that in a 1D-space to obtain the resulting 1-simplex you subtract points from one another. The algorithm can skip 'internal' points and only compute the difference of outermost opposite points. The support function calculates the difference of opposite points in a more general way. As a bonus it also allows *flat vs curved* collisions! With it you can detect intersections of ellipses, circles, curves and splines in 2D and rounded shapes and more complex objects in 3D (which is cool).

Let's think of opposite points in two dimensions. Take a look at these examples of opposite points of a 2D-shape:

That is trivial. Now imagine you take not one but two arbitrary shapes and pick a random point of the first shape then pick a second point on the opposite side of the other shape. You might end up with something similar to this:

A difference of two points yields another point of resulting shape. That point is literally the distance vector. So, if you take a point of a shape and then choose a point of the other shape and then subtract the two points from one another, you get exact distance and direction between your shapes (at specific points).

Note, that when you subtract *opposite* points of two shapes your resulting point-vector will always land somewhere on the contour (an outermost edge) of your resulting shape. Below is an illustration of how a difference of two opposite points (on the left) finally lands on the contour of resulting shape (on the right).

The left side of the picture shows our simulated world space. On the right is our 2D *Minkowski space*. In GJK you can think of Minkowski space as if it was an imaginary world of shape differences. By subtracting two points in real world we therefore obtain a new point in another surreal world. It is also often called a *mapping* or a *projection* of a real-world intersection into Minkowski space (into the world of differences). The result in Minkowski space looks like a weird inside-out union as if one shape is *"swept"* along the contour of the other. But this is how all intersections really look like in 2D.

It's easy to show arithmetically that the resulting point is obtained by taking the difference of the two opposing points:

```
A(x1, y1) - B(x2, y2) = C(x1 - x2, y1 - y2)
A(3, 1) - B(1, -1) = C(3 - 1, 1 - (-1)) = C(2, 2)
```

Resulting point `(2, 2)`

ends up exactly on the contour of our difference shape. If initial points are opposite their difference will always be one of the outermost (*"external"*) points of the resulting shape. Also notice, that resulting distance vector `(2, 2)`

is in one-to-one correspondence to the distance between the two initial opposite points. They are literally the same.

The support function of GJK maps a difference of two real-world points into Minkowski space. It seeks for two opposite points which are furthest apart along a given direction and returns their difference. In seek of opposite points along a direction the support function does the following:

- Start at the Origin and choose any direction you like (denoted as
`D`

on the image above). A direction is itself a vector, pointing somewhere away from the Origin. It can be random, you choose whatever you want for a start, later you'll see why initial direction doesn't really matter. The direction vector always starts at the Origin and is sometimes written as`OD = D(x,y) - O(0,0) = D(x,y)`

(that is a direction from`O`

towards`D`

). - Take the first of two shapes. It does not matter which one of the two is first. From the Origin seek for the furthest point of first shape in direction
`D`

. To find the furthest point in direction`D`

calculate the dot product of all the point-vectors of the first shape with direction`D`

. This is the same as taking magnitudes (or distances) from the Origin to each point of the first shape in direction`D`

. It is also often called*projecting a point-vector onto a direction-vector*. The distance to point`A`

along direction`D`

is the projection of vector`A`

onto vector`D`

. The most distant point`A`

along`D`

will have the greatest dot product of`A`

and`D`

. This way our first most distant point from the Origin along direction`D`

is found. Note that a dot product can result in a negative projected vector. - Next an opposite point of the
*other*shape must be found. To do that the support function switches the direction`D`

to the opposite, literally flipping it`D = -D`

. The process from previous step is repeated, but this time with the other shape in a reverse direction`-D`

. It looks for a point of the second shape that is most distant from the Origin along direction`-D`

. To find the most distant point it calculates all dot products of all the point-vectors of the second shape and direction-vector`-D`

. Then it takes the point which has the greatest dot product with`-D`

. This way the second point`B`

is found that is most distant from the Origin and is also opposite to the first point. - Once two opposite points
`A`

and`B`

have been found, subtract one of them from the other and done. It doesn't matter which one you are subtracting from. The resulting vector is a mapping of their difference into 2D Minkowski space.

In other words, the support function takes an arbitrary line and two opposite points furthest from the Origin along that line, one point from each of the initial shapes. This is similar to subtracting segments in 1D. The support function finds the endpoints of two shapes along some direction-line and returns their difference that is another point in Minkowski space. On a one-dimensional number line the resulting point is a 1D-number. One a two-dimensional coordinate plane the resulting point is a 2D-vector (it has two coordinates).

Now its easy to show that it does not matter which initial direction you choose to start with. The support function does not care about given initial direction at all. For example, if we choose a different arbitrary `D`

, we might end up with something like this:

It doesn't matter which direction `D`

you choose to seek for opposite points `A`

and `B`

. By taking their difference you get a point `C`

somewhere on a contour in Minkowski space. And it's easy to verify arithmetically that the intersection of opposite points `A`

and `B`

along any direction `D`

still yields one of many points on the contour of our resulting shape:

```
A(2, -2) - B(-1, 2) = C(2 - (-1), -2 - 2) = C(3, -4)
```

So, in general the support function can work in any given direction and in any given space. You give it two shapes and a direction, and then it finds two opposite points in your space and returns their intersection in Minkowski space. In mathematics the intersection is usually denoted with symbol `∩`

.

```
C = support (shape1, shape2, D); // return a point of (shape1 ∩ shape2) along arbitrary direction D
```

The intersection of two points always yields a third point. In 1D a point is a number. Thus, an intersection of two numbers or two points in 1D gives a third number or another 1D-point. In 2D an intersection of two vectors or two points gives a third vector or another 2D-point.

To calculate the difference of two numbers we subtract one of them from the other like this: `A - B = C`

. In 1D which number of the two is A and which one is B does not matter, the algorithm works either way. This is because in 1D you have only one possible direction that is the actual number line itself. But in general when dealing with 2D or 3D coordinate vectors (which is usually the case in many applications) the order of subtraction actually does matter. A more accurate way of representing the difference of two vectors in Minkowski space is to take one vector and sum it with a negated version of the other vector, so that `A + (-B) = C`

. Because of this fact the Minkowski support function is often defined as a sum of the first point-vector with the negated version of the second point-vector. After negating the second vector you simply add it to the first one to get their arithmetic total result. This is why the support function is called *Minkowski addition* or *Minkowski sum* and you will probably never hear of *Minkowski subtraction* nor *Minkowski difference*.

Here's how an implementation of the Minkowski sum support function for any space of arbitrary amount of dimensions might look like in pseudocode:

```
//-----------------------------------------------------------------------------
// Return an arithmetic sum of two point-vectors
// A 1D-vector has only one component (one coordinate on a number line)
// In general a vector has one or more components
vec sum (vec a, vec b) {
return a + b; // [ a.x + b.x, a.y + b.y, a.z + b.z, ... ]
}
//-----------------------------------------------------------------------------
// Dot product is the sum of all corresponding components of both vectors multiplied
float dotProduct (vec a, vec b) {
return a * b; // a.x * b.x + a.y * b.y + a.z * b.z + ...
}
//-----------------------------------------------------------------------------
// Get furthest vertex along a certain direction d
// This is the same as finding max dot product with d
// In 1D direction is always 1 or -1 (to the left or to the right of the Origin)
size_t indexOfFurthestPoint (const vec * vertices, size_t count, vec d) {
size_t index = 0;
float maxProduct = dotProduct (d, vertices[index]);
for (size_t i = 1; i < count; i++) {
float product = dotProduct (d, vertices[i]); // may be negative
if (product > maxProduct) {
maxProduct = product;
index = i;
}
}
return index;
}
//-----------------------------------------------------------------------------
// Minkowski sum support function for GJK
vec support (const vec * vertices1, size_t count1, // first shape
const vec * vertices2, size_t count2, // second shape
vec d) { // direction
// get furthest point of first body along an arbitrary direction
size_t i = indexOfFurthestPoint (vertices1, count1, d);
// get furthest point of second body along the opposite direction
// note that this time direction is negated
size_t j = indexOfFurthestPoint (vertices2, count2, -d);
// return the Minkowski sum of two points to see if bodies 'overlap'
// note that the second point-vector is negated, a + (-b) = c
return sum (vertices1[i], -vertices2[j]);
}
```

It does not really care how many dimensions are there in space. So given a direction and two shapes the support function always returns another point regardless of how many dimensions you have. It works the same for 1D, 2D, 3D, etc... The point returned by the support function is always on the contour of intersection. This is because the initial points involved in the intersection are opposite.

Remember, the whole purpose of having a support function was to help us quickly build the simplex for GJK. The support function is used in the search for a 2-simplex that encloses the Origin in 2D. Now that we have such a function, we can move on and build the rest of the algorithm on top of it.

This is where the actual logic of GJK kicks in. The general plan of GJK is:

- Find best points for a simplex with the help of the support function.
- Check if a simplex of those points encloses the Origin from all sides.
- If it does, there is a collision – hooray and thanks for the support, support function )
- If it doesn't, well... try other points, why not?
- If a simplex cannot be built at all no matter how many times you try – it might be a
*degenerate case*(explained later). - If it's not a degenerate case, then there's no collision.

All of that is done in a loop. The algorithm tries different points and checks if they satisfy the condition. The whole process is called *evolution*. If it succeeds to enclose the Origin in another branch of evolution (upon another loop iteration), then the intersection is detected. If not, it tries again until finally it either succeeds or fails.

Some people might be reasonably worried of the possibility that the evolution of GJK runs out of control and continues on and on without ever stopping. It might even become intelligent some day and who knows what could happen... So they add a limit of iterations into their implementations which is a countdown that cuts power off when the game is over, forcing the algorithm to stop. It's like an emergency halt or a safety button in case something goes wrong with float number precision or whatever. But that is actually not necessary in general.

In search for a 2-simplex the algorithm has to obtain 3 points one by one to make a triangle. The resulting set of points is initially empty. The process of finding three points is done step-by-step. First, the algorithm finds the first point and adds it to the resulting set. Then it finds and adds the second point, and then it finds and adds the third one. So, while being built, the 2-simplex kinda *evolves* actually passing through all stages of evolution from a 0-simplex (one point, the first one in the resulting set), through a 1-simplex (segment of two points, with the second point added to the set), to a 2-simplex (three points, with the third point added to the resulting set).

A 2-simplex (a triangle) contains or consists of 1-simplices (segments). A 1-simplex (a segment) consists of even simpler 0-simplices (points). It's clear that each new dimension adds one more point to the simplex. So, in general:

```
0-simplex = nothing + 1 point = 1 point
1-simplex = 0-simplex + 1 point = 2 points
2-simplex = 1-simplex + 1 point = 3 points
3-simplex = 2-simplex + 1 point = 4 points
... = ... + 1 point = ...
N-simplex = (N-1)-simplex + 1 point = N+1 points
```

GJK evolves the simplex from the very beginning each time, restarting the evolution when no further progress is possible in current branch. So a simplex passes through all stages of its evolution upon each iteration of the main loop.

Let us do a single iteration of the main evolution loop by an example, step-by-step. Below is an illustration of a pair of slightly different intersecting shapes (for example), and their intersection Minkowski sum on the surreal side to the right. The Minkowski sum is unknown to our algorithm at the moment, as we haven't calculated all the differences of points yet and, in fact, we won't need the full resulting shape, we only need best three points to surround the Origin and that's it. So the Minkowski sum is currently unknown and we're drawing it in full on the right just for visual explicacy:

To build a 2-simplex in 2D we need three points from our resulting Minkowski set that would enclose the Origin within a triangle and we have our nice support function for finding best points for that. Before we begin the search for three points we reserve a place for each one of them and label or tag each place with capital letters `C`

, `B`

and `A`

in that order. Their naming might be a little confusing, as we've used `A`

and `B`

to denote opposite points previously in this text, but it's an ancient tradition, sorry, can't do much about that. These will be placeholders for the points we might find in the search process. Initially our resulting set of points is empty (all placeholders are empty).

The first point `C`

is the easiest to obtain. You just choose a random direction `D`

and call the support function which calculates that point for you as we did earlier. Initial direction `D`

is chosen randomly and passed along with the shapes to the support function. The support function then finds opposite points of two shapes along the given direction (labelled below as `a`

and `b`

in lowercase for less confusion) and returns their Minkowski sum which is point `C`

on the right, the first point of our simplex (which is a 0-simplex for now).

Once you get the first resulting point, you place it in placeholder `C`

. From now on we will be referring to that point by that label or tag `C`

. Don't forget, that `C`

is a point on the difference contour line in Minkowski space. Now we have a 0-simplex (a point) in our resulting set and it is the first stage of current iteration of evolution. Note, that the support function looks in both `D`

and its opposite `-D`

in search for points `a`

and `b`

. This is what we get after executing the first step and obtaining point `C`

in Minkowski space:

Once we got the first difference of most distant opposite points along a certain direction-line we need to choose a different direction to find the next point of our simplex in Minkowski resulting set. Remember, that the magic of GJK says that we have to surround the Origin from all sides to detect a collision. As with 1D number line where a segment can surround the Origin with two of its endpoints, in 2D we can surround the Origin from all sides with three *endpoints* or vertices of a triangle. So our goal is to find such a 2-simplex that would enclose the Origin within itself.

The logic behind this can be described in the following way: imagine we *stand at* point `C`

and look towards the Origin from that perspective. We want to find next point that would be *beyond* the Origin as seen by us. If there's no such point then we are not reaching far enough and we will not be able to surround the Origin with a triangle using current point `C`

.

If we cannot reach for a point beyond the Origin from current standpoint `C`

, then there's probably some distance between initial shapes (at least those two shapes don't intersect along direction `D`

). So, in that case no collision is detected. The halting criteria of GJK algorithm says that current iteration stops if our simplex fails to include (or surround) the Origin, handling these cases will be explained in more detail later in this text.

Standing at point `C`

we need to check if there's a point that is further away from us than the Origin is, in the direction from us towards the Origin. The direction from point `C`

towards the Origin `O`

is the direction `CO`

which is the reversed opposite of direction from Origin to point `C`

, so that `CO == -OC`

. Therefore we should be looking in direction `CO`

next as we hope to find some point beyond the Origin there.

The idea of *looking beyond Origin from your current standpoint* is the key principle of the search for simplex in GJK. It helps to quickly find the biggest simplex by taking the most distant opposite points of Minkowski sum. The bigger the simplex the higher the chances that it will surround the Origin upon some early iteration.

The second point `B`

is quite easy to get. From now on we set new direction `D = -OC`

which is equal to `CO`

, and we will refer to it as our new direction `D`

without the negative sign. The support function finds the other two most distant opposite points `a`

and `b`

and it looks in both directions `D`

and `-D`

(yes, again) in the process. Note, that new direction vector `D`

is different from the first initial direction vector `D`

as it now points in the opposite way.

You call the support function with your new opposite direction `D`

and get the second point which is labelled or tagged as capital `B`

. And now there's a 1-simplex (a segment of two points, `C`

and `B`

) in the resulting set, and it is the second stage of that same iteration of evolution. Here's what we get after the second step:

By calling the support function the second time with the opposite direction you get a second point `B`

which is another point on the same contour in Minkowski space. That second point `B`

is exactly opposite to the first point `C`

, because initial directions passed into the support function were opposite, right?

Having point `B`

we need to verify that it is indeed beyond Origin as seen from point `C`

. In other words, we have to check if point `B`

is further away from point `C`

than the Origin is (in direction from `C`

towards the Origin). If a dot product of `CO ⋅ OB`

is positive, then point `B`

is really beyond the Origin as seen from point `C`

and the evolution continues, otherwise there's no collision and the evolution stops or restarts in a different direction. In geometry the sign of dot product of two vectors basically tells if their orientation is roughly the same.

`if ((CO ⋅ OB) > 0) // test if point B is beyond Origin as seen from point C`

This test for *a point beyond Origin* can also be explained as follows: we know that point `C`

is located on the contour of Minkowski sum (this is a feature of our support function) and point `B`

is on the opposite side of that same contour, so the Origin must be in between `C`

and `B`

, otherwise we're not reaching far enough to surround it, and if we cannot surround it, then there's probably some distance between the two initial shapes, therefore no intersection.

Remember that the algorithm doesn't know anything about the whole set of Minkowski points, it currently only knows about points `C`

and `B`

that were found during first two stages of the evolution. Now let's take a look at the Minkowski space as seen by GJK algorithm at this point. The left part of the image below shows the 1-simplex of two points (segment `CB`

) in Minkowski space, and on the right we see the same segment with the full set of Minkowski sum points added for visual reference and for explanation purposes (the algorithm only sees the left side of this picture, not the right side):

The algorithm's current view into surreal Minkowski world is on the left side of the image above. There's 1-simplex segment `CB`

of two points. The simplex has evolved from being a single point `C`

0-simplex into a 1-simplex or a segment of two points `CB`

, in other words, our simplex has become a little more complex and obtained an extra dimension )

Before we proceed to the third point `A`

we might need to gain even more intuition about what is going on here. Seeing that segment `CB`

on the image above and seeing the whole set of Minkowski sum points on the right side, humans can quickly visually locate the third point to enclose the Origin. Easy. But for a machine to know where to look for the third point, we have to make a critical decision in which direction it should be looking next.

The logic of GJK goes this way: imagine we *stand anywhere on the segment* `CB`

and look towards the Origin from that perspective. We want to find such a third point that would be at least *beyond* the Origin (as seen by us relatively to the segment `CB`

that we're *standing on*). So, again, if we cannot find a third point beyond the Origin then we will not be able to surround it using current segment `CB`

and there's no collision at least on this iteration.

The image above shows the direction to look for the last point of triangle simplex. To find the exact direction vector, you just take one of perpendiculars to the segment `CB`

. If you tilt your head to the right a little while looking at the segment `CB`

, you will immediately notice the following simple fact: any segment kinda cuts the coordinate plane into two halves, to the left and to the right of the segment. And the Origin always ends up either on one side of the segment `CB`

or on the other side. All we have to do to find the next direction vector is just take a perpendicular (often called *a normal*) to `CB`

that points towards the Origin.

There are two possible directions perpendicular to segment `CB`

, the one pointing towards the Origin and the one pointing the opposite way (away from the Origin). In order to get the one pointing towards the Origin, we can do a simple vector math trick, called *vector triple product*. It works like this: take a cross product of the segment `CB`

with a segment `CO`

(from endpoint to the Origin) and then take a cross product of the result with the segment `CB`

again, basically two cross products done sequentially, involving the same segment `CB`

twice:

`CB ⨯ CO ⨯ CB`

There's also a very fast formula for this (an expanded equation) aka *triple product expansion* with only a few multiplications and subtractions for calculating the resulting perpendicular:

`a ⨯ (b ⨯ c) = b(a ⋅ c) - c(a ⋅ b)`

You have to be precise with the orientation of vectors when using the expansion formula, but if done carefully and correctly, this always gives a perpendicular to segment `CB`

pointing towards the Origin.

We set our final direction `D`

to that perpendicular to `CB`

pointing towards Origin. That will be a direction to search for point A of our triangle simplex. This is also the last stage of the evolution of the 2-simplex. We call our support function and pass new direction `D`

along with the two shapes. The image below shows the result after calling the support function for the 3rd time:

We obtained point A and now we have a complete triangle of three points in total. Our simplex has evolved from a 0-simplex point through a 1-simplex segment into a 2-simplex triangle. Also notice, that we did not calculate all points of Minkowski sum, we only calculated three of all possible points. This is one of the reasons why this algorithm actually works very fast in narrow-phase collision detection. It just doesn't have to calculate all points, so it does not know anything about the full Minkowski set. Here's how it sees the triangle with point A added to the simplex in memory (remember, it only knows about the left side of the image below, but not about the right side which is there for explanation purposes):

Now that we have all three points for a complete triangle (a full 2-simplex), the only task that's left is to test if the Origin is really contained inside that triangle. In other words, if our resulting simplex surrounds the Origin from all sides, then the whole Minkowski sum also does surround the Origin as well, and it follows that there was a collision between two initial shapes. The most straightforward way to test if a point is inside a triangle (the Origin is a point at zero) is to do a bunch of dot product operations.

For each edge (each side) of a triangle you need two things: a normal vector (a perpendicular) to that side towards the Origin and a vector from the Origin to opposite vertex of the triangle. Then you check if dot product of the two vectors is positive (greater than zero)...

WORK IN PROGRESS, to be continued soon. A live demo of GJK in a 2D-space and a video of GJK in action coming up )

...

WORK IN PROGRESS, to be continued soon... )

...

A collision is when two bodies occupy the same points in space. In two dimensions a collision is either an intersection of two shapes (when shapes kinda "overlap") or they might not intersect but instead one shape could just touch the other, and that is also considered to be a collision. There's even more details to the nature of collisions, because there are different types of touch...

A full-on collision when a shape overlaps or penetrates another shape usually looks like this:

When 2D-shapes overlap there is usually at least one way to build a 2-simplex that encloses the Origin. But there can be other types of non-penetrating collisions without overlapping, when shapes touch edge-to-edge or meet at one single point. In GJK these collisions are usually called *degenerate case*. They are not collisions but contacts in common sense. By design GJK handles all degenerate cases absolutely fine.

Here are some examples of what a degenerate case collision (a touch) in GJK is:

It may seem like a lot of special cases to handle, but in fact, GJK already does that intrinsically.

A non-overlapping collision will yield a Minkowski sum that has the Origin on its *edge* or *contour*. Remember, in 1D, when two segments have only one common point, the Origin lands on the endpoint of resulting segment. In 2D, when there is only one common point, the Origin in Minkowski space will be located on the contour of your resulting 2D-intersection shape. If two shapes share a common face, then the Origin will be on one of the sides of triange simplex.

So, the Origin can either be inside the Minkowski sum, or it can be on the edge of the sum. And the trick is to check if the Origin is actually one of the points of your simplex. The Origin can also end up exactly on one of the sides of your simplex, between two points of the simplex.

Basically, to each stage of evolution you add a check, to see if the Origin is already included in your not-fully-evolved simplex or not. That is enough to detect all degenerate cases. If the Origin is already included in your half-built simplex upon some early stage of evolution, then you don't have to evolve it further. Then you can stop and signal a collision (or no collision, at your discretion). You can differentiate between a penetrating collision or not, by counting degenerate collisions or not counting them as being collisions.

WORK IN PROGRESS, to be continued soon... )

What's very remarkable of this algorithm, is that the support function can be also defined for a huge variety of shapes. If every point of a shape can be derived from a simple formula (aka *parametric equation*) then your support function can use that equation to calculate differences of points. When geometry is defined in a parametric way there is no need to keep all points in memory. And you can also scale everything up and down as you like, do boolean combinations of shapes and much more... That's what vector graphics is about, but it's a whole other topic in itself, we won't dive deep into it here.

The key point is: a parametric equation can be an equation of a circle `((x - h)/r)² + ((y - k)/r)² = 1`

or an equation of an ellipse `(x - h)²/a² + (y - k)²/b² = 1`

, or some *spline* (!), or even a combination of polygons and conic sections. So, you can detect collisions and intersections of round shapes very accurately up to an exact point of contact. GJK is not restricted to polygons, you can do perfect circle collisions, and as we'll see later in the 3D section, spherical collisions are also possible as well.

The only restriction that still holds is that your shapes should be convex, not concave, so, any line crossing your shape should not intersect its contour more than twice. As long as your shapes are fat and bulgy, you'll be fine. But there's a hint: you can always make a concave shape from multiple convex shapes.

WORK IN PROGRESS, to be continued soon... )

A careful reader might have already noticed a pattern in the algorithm...

For a 1D number line we need a 1-simplex of two points to enclose the Origin. If we can find such two points then our shapes do intersect indeed. If we cannot find such two points then our initial shapes must have some distance (non-zero difference) between them. For 2D coordinate plane a 2-simplex of three points (a triangle) is necessary to enclose the Origin. In 3D space we need a 3-simplex of four points (a tetrahedron). These are all examples of simplest possible shape primitives in given dimensions.

WORK IN PROGRESS, A live demo of GJK in a 3D-space and a video of GJK in action coming up soon )

The version of GJK that gives a boolean answer to a yes/no collision test is somewhat simplified. The algorithm is able to not only detect the fact of intersection, but is also capable of giving back the exact depth of penetration and information about points of contact, so that the collision could be handled properly.

The simplified yes/no test is often called a *bastardized* version of GJK algorithm in comparison to its original purpose of calculating detailed collisions. But the simplified version described in the text above is easier for understanding. Having understood the simple yes/no GJK test it is much easier to grasp the Gilbert-Johnson-Keerthi algorithm in its entirety along with a few useful concepts. We will proceed to cover the rest of GJK functionality below.

WORK IN PROGRESS, to be continued soon... )

WORK IN PROGRESS, to be continued soon... )

Most of the info (along with reference implementation) was taken from dyn4j. It has a very thorough explanation of GJK and it is definitely a must visit for those who need to understand the intricacies of the algorithm.

- http://www.dyn4j.org/2010/04/gjk-gilbert-johnson-keerthi/ GJK description (+ a lot of other useful articles)
- http://mollyrocket.com/849 Awesome old-school GJK / Minkowski space video by Casey Muratori
- https://awesomeopensource.com/project/wnbittle/dyn4j Quality source code for reference
- https://www.youtube.com/watch?v=-GZOGJ26yxE – A video of a seminar by Elmer Gilbert Professor Emeritus Aerospace Engineering, University of Michigan

Alternatives To Gjk.cSelect To Compare

Related Awesome Lists

Top Programming Languages

Get A Weekly Email With Trending Projects For These Topics

No Spam. Unsubscribe easily at any time.

C (193,862)

Shape (5,981)

Physics (5,354)

Collision (2,665)

Simplex (541)

Collision Detection (499)

Gjk (16)